` tag the server sends with the page. ## Step 2: Structure the Data, Kill the Stories AI does not care about your brand journey. It wants hard facts. If you are burying what you actually do inside five paragraphs of "synergistic workflows tailored to empower your unique enterprise journey," the LLM cannot extract it. **The audit.** Open your Services and Pricing pages. Are your deliverables buried in heavy paragraphs? Are your prices hidden in an accordion that requires a click? Are you describing what you do in adjectives when you could be describing it in nouns? **The fix.** Serve product details and pricing in structured formats an LLM can parse instantly. Lists, bullet points, clean tables. This site does this on every pricing and feature block, which is why an AI can answer "what does Practical AI offer and at what price" without hallucinating. | Human-Optimised (Bloated) | AI-Optimised (Structured) | | --- | --- | | "Since our founding, we have prided ourselves on delivering bespoke, synergistic workflows tailored to empower your unique enterprise journey." | Core Service: Automated WorkflowsDeliverable: API integration and AI routingTarget: Operations Directors | ## Step 3: Fix Your Semantic Tree H1, H2, and H3 tags are not font sizes. They are a database schema that tells the LLM exactly how to index your business. Most websites get this wrong. Multiple H1s on a single page, H2s used purely for visual decoration, an H3 that contains the company name and tagline in a single run-on. **The audit.** Run a free SEO checker (W3C validator, Lighthouse, or the ahrefs free audit) and look at the header tag tree. You are looking for exactly one H1 per page, H2s for major sections, H3s for items within those sections. **The fix.** Your H1 should explicitly state what the business does, in plain language. Your H2s should be your core services or sections. Your H3s should be the features, deliverables, or pricing tiers. Use them logically. While you are in there, check whether you have structured data (Schema.org JSON-LD) for your products, services, or organisation. It is the same idea at a deeper level, and most AI engines prefer it to unstructured prose. ## Step 4: Validate With an AI The closing test. Open Claude, ChatGPT, or Gemini and ask, out loud: "What does [your company] do, what do they offer, and at what price?" Do not look at your site while you ask. Then ask: "Which of their services would suit a [your target customer]?" If the answer is wrong, hedged, or vague (if it hallucinates a service you do not offer, or misses one you do), your prospects are getting the wrong information too. The audit fixes above directly address whatever the AI got wrong. Run the same validation in a month, after the fixes have shipped, and compare the answers. You can run all four steps in 15 minutes without writing a line of code. The fixes typically take a developer half a day if you find problems. Pilot the audit on your homepage first. If your homepage passes, the rest of the site usually follows. ## Beyond the Basics: /llms.txt for Builders If you have a developer on hand, there is one more thing worth doing. `/llms.txt` is a new standard (live at [llmstxt.org](https://llmstxt.org)) where you publish a clean Markdown file at the root of your site describing what your business does, what it offers, and how to contact you. AI scrapers are starting to look for it the way they look for `sitemap.xml` or `robots.txt`. This site has one. You can read it at [practical-ai.space/llms.txt](/llms.txt). It is around 30 lines of plain Markdown. A developer can write it in an hour, and it gives AI engines a clean, structured summary to cite from. Not a replacement for the audit above. A complement. Most businesses do not need this yet, but if you are serious about being recommended by AI, it is the next step. Optimising for LLMs is not a dark art. It is a basic data pipeline problem. Fix the text, structure the facts, label the schema, and validate the result. ## Where to Start Run Step 1 in the next 15 minutes. Open your homepage with JavaScript off. If the AI can read your value proposition, you are ahead of most of your competitors. If it cannot, you have a punch list of fixes that will pay back every time a prospect asks an AI about your category. Once the basics are solid, the free ["5 AI Wins You Can Action This Week"](/5_AI_Wins_You_Can_Action_This_Week.pdf) guide on this site shows how to put AI to work inside your business once your website is doing its job. ### This is what an AI strategy session actually covers Website readability is one piece. The bigger picture is which AI ecosystem, which frontier model, how to govern AI safely, and where to start. Our Executive AI Strategy sessions work through your specific business: your current site, your existing contracts, your team's actual workflows. You leave with a 30-day action plan that prioritises the changes that move the needle first. [Book a Free Discovery Call](https://calendly.com/practical-ai/30min) --- AI is being positioned to us as if there is one decision: pick an ecosystem, and you are done. But thats not really the case. Its more like we moving to make two independent decisions stacked on top of each other, and conflating them looks to be a common strategic mistake. The first decision is where your daily work lives (we probably spend significant amount of our day here): email, documents, meetings, calendar. The second decision is what your team reaches for when the work can potentially be optimised with AI tools or services: writing and reasoning, coding, data analysis, deck-building, agentic automation. Most businesses pick one suite from each layer and combine them. The CFO still gets email in Outlook. The analyst still builds the model in Excel. But when someone needs to reason through the quarter or draft the board deck, they reach for Claude. This guide walks through this 'dual-stack' model and how they fit together. No code - but rather a decision framework for calls that will shape your operating costs and your team's workflows for the next two to three years. ## The Two Layers Think of your AI setup as two stacked layers, not one flat choice. **Layer 1: Your productivity backbone.** Where your email, documents, meetings, and calendar already live. The honest answer here is mostly already decided by what you pay for today. If your team is on Microsoft 365, the M365 stack is the obvious place to start. If you are Google-native, the Workspace path is. The decision is real but slow-moving, and most of it is settled by your existing contracts. **Layer 2: Your frontier AI model.** The powerful general-purpose AI your team reaches for when the work is hard. Writing and reasoning. Coding and code review. Data analysis and visualisation. Building decks. Running agentic workflows. This is where Claude, ChatGPT, and Gemini compete head-to-head on capability. The decision is fast-moving, high-stakes, and where AI competence actually differentiates a business. The two decisions are independent. You can keep Microsoft 365 and layer Claude on top. You can keep Google Workspace and layer ChatGPT. You can commit to a single vendor on both layers if you want, but you don't have to. Most guides treat them as one. That is the mistake this guide is built to fix. ## Layer 1: Your Productivity Backbone Three serious contenders. Each has a clear shape, a clear cost, and a clear set of businesses it suits. The Most Likely Default ### Microsoft 365 Copilot **Cost:** Around $30 per user per month, on top of your existing M365 licence (verify against current vendor pricing) [View pricing →](https://www.microsoft.com/en-us/microsoft-365/copilot/business) · **Best for:** Businesses already on M365, especially with Teams and Outlook at the centre of daily work **What it actually does.** Copilot sits across every M365 app. For most businesses the most-used touch points are Teams (meeting recaps, action items, chat summaries, message drafting) and Outlook (email summaries, reply drafting, calendar triage). Beyond those, Copilot lives inside Word, Excel, PowerPoint, OneDrive, and SharePoint, with shared context across all of them. Recent connector updates let it reach into Salesforce, ServiceNow, and Jira. **Who it suits.** Any business already paying for M365 across the team, especially those above 20 employees where the productivity layer needs to be consistent across departments. Financial services, professional services, and regulated businesses benefit from Microsoft's mature enterprise governance and UK/EU data residency controls. **The honest downside.** A serious financial commitment. Across 30 people, $30 per seat per month is roughly £10,800 a year on top of what you already pay for M365 itself. Lock-in is real once your team is using Copilot daily in Teams, Outlook, and Word. Quality varies across apps: Teams and Outlook are mature; Excel and PowerPoint are still catching up to what dedicated AI tools can do. The Google-Native Choice ### Google Workspace with Gemini **Cost:** Bundled into Workspace Business and Enterprise plans (verify against current vendor pricing) [View pricing →](https://workspace.google.com/pricing) · **Best for:** Businesses already on Google Workspace, especially those with a younger team or strong Google-first habits **What it actually does.** Gemini is built directly into Gmail (Help me write, Suggested Replies, thread summaries), Google Meet (real-time translation, automated notes), Docs (AI side panel for drafting), Sheets (formula help, table generation), and Drive (search and summarisation across files). **Who it suits.** Businesses where Gmail and Google Meet are already the daily centre of work. Creative agencies, tech startups, modern SMBs that never moved to Microsoft. Also a strong choice if cost matters: Gemini is included in Workspace plans you are likely already paying for. **The honest downside.** Less mature enterprise governance than Microsoft. Fewer controls over how AI features are deployed across the team. If you are in a regulated industry or have strict UK/EU data residency requirements, Google's setup needs more careful review. Sheets and Slides are less mature for power users than Excel and PowerPoint with Copilot. The Meeting-Centric Overlay ### Zoom AI Productivity Suite **Cost:** Per-seat pricing on top of Zoom Workplace, plus AI Credits for metered actions (verify against current vendor pricing) [View pricing →](https://zoom.us/pricing) · **Best for:** Consultancies, agencies, advisory firms whose primary work product comes out of meetings **What it actually does.** Turns conversations into deliverables: Canvas for written work, Slides for presentations, Sheets for data, Paper for notes. ZoomMate is the agentic search layer that pulls context across your meetings, chats, and connected apps. A 60-minute call becomes a polished follow-up deck, a structured spreadsheet of action items, and a draft client email. **Who it suits.** Consultancies, agencies, coaching and advisory firms where the meeting is the deliverable. If your business lives in Zoom and the bottleneck is what happens after the call, Zoom AI is worth a serious look. **The honest downside.** You will still need a primary productivity stack underneath it (M365 or Google) for email, identity, finance, and document workflows. Zoom AI is a meeting-centric overlay, not a full backbone replacement. If your team is split between Zoom and Teams or Meet, it becomes another tool to maintain rather than a unifying layer. ## Layer 2: Your Frontier AI Model Three serious contenders. The decision here is capability-led, not contract-led. The Reasoning and Analysis Leader ### Claude **Cost:** Claude Pro from around £18 per user per month; API pricing varies by model tier (verify against current vendor pricing) [View pricing →](https://www.claude.com/pricing) · **Best for:** Teams whose hardest work is reasoning, analysis, writing at length, coding, or agentic workflows **Where Claude wins.** Long-document reasoning (200K context window, handles full board packs and contracts in a single prompt). Claude Code for developer workflows. Data analysis and visualisation with the artifacts interface. Agentic work and multi-step planning. Deck creation and structured outputs. Instruction following that holds up under complex prompts. Connectors into Microsoft 365, Gmail, Slack, and Notion mean Claude can also act as an intelligent front-end for the productivity backbone, not just a parallel tool beside it. **Who it suits.** Any business whose hardest work is thinking, analysing, writing at length, coding, or running multi-step agentic tasks. Particularly strong for finance, legal, consulting, and research-heavy roles. **The honest weakness.** Narrower consumer reach than ChatGPT. Smaller plugin and integration ecosystem. Weaker on real-time voice and browser-native experiences. Anthropic is more focused on capability than consumer surface. The Ecosystem and Surface-Area Leader ### ChatGPT **Cost:** ChatGPT Plus from around £20 per user per month; API pricing varies by model tier (verify against current vendor pricing) [View pricing →](https://openai.com/chatgpt/pricing/) · **Best for:** Consumer-facing, browser-driven, or plugin-heavy work **Where ChatGPT wins.** The Atlas browser turns ChatGPT into a browsing-native agent that can act on what it sees. Agent mode runs multi-step tasks with computer use. The largest consumer and ecosystem footprint. Mature voice and real-time features. The biggest plugin and integration store. **Who it suits.** Businesses whose hardest work is consumer-facing, browser-driven, or benefits from the largest plugin ecosystem. Sales, marketing, customer ops. Any team that wants the broadest surface area and the most third-party integrations. **The honest weakness.** Weaker long-context reasoning than Claude. Less refined instruction following for code-heavy work. Less mature for analyst-style data work where depth matters more than breadth. The Cost-Efficient Google-Native Option ### Gemini **Cost:** Bundled into Google Workspace plans; standalone Gemini Advanced from around £19 per user per month (verify against current vendor pricing) [View pricing →](https://gemini.google/pricing) · **Best for:** Google-native teams that want one vendor across both layers, or multimodal-heavy work **Where Gemini wins.** Native to the Google stack, so it doubles as both your backbone and your frontier model if you are Google-native. Strong multimodal and video understanding. Aggressive pricing, especially inside Workspace. **Who it suits.** Google-native businesses that want one vendor across both layers. Any team whose work is multimodal-heavy (video, image, audio). Any business that wants the lowest-cost path to a serious frontier model. **The honest weakness.** Less mature as a standalone tool outside the Google stack. Smaller third-party ecosystem than ChatGPT. Weaker than Claude on long-context reasoning for the hardest analytical work. ## The Headline Pattern: Backbone Plus Frontier, Layered The most common successful AI setup in mid-market right now is one suite from Layer 1 plus one model from Layer 2, layered. Not picked together. Picked separately, and combined deliberately. **Microsoft 365 plus Claude.** The canonical mid-market 2026 setup. Outlook, Teams, Excel, Word, PowerPoint stay where they are. The hard work (analysis, drafting, building, agentic workflows) runs through Claude. You get enterprise governance from Layer 1 and frontier capability from Layer 2, with no forced compromise on either. This is the configuration most businesses with serious AI ambitions land on. For executives especially, the connector pattern matters in practice: with Claude wired into Microsoft 365, Claude can read and triage the inbox, summarise a Teams thread, and pull context from a SharePoint document. The user spends less time inside Outlook and Teams directly, and more time working through Claude as the intelligent surface in front of the whole stack. **Google Workspace plus Gemini.** The equivalent if you are Google-native. The advantage: Gemini doubles as both layers, which keeps total cost lower. The trade-off: less best-in-class reasoning than Claude for the hardest analytical work. **Microsoft 365 plus ChatGPT.** The right pick if your hardest work is consumer-facing, browser-driven, or depends on the largest plugin ecosystem. Less strong for long-context reasoning and code-heavy workflows than Microsoft plus Claude. **Standalone Claude-only.** The right pick for very small teams running lean. Skip Layer 1 commitments entirely, pay for Claude Pro, and connect it to your existing email and documents where you can. The pattern to avoid: treating the two decisions as one and committing to a single vendor across both. It almost always means overpaying for one layer or underspec'ing the other. You can run a small pilot on each layer before committing across the business. Microsoft, Google, and Zoom all allow small-seat trials for Layer 1. Claude, ChatGPT, and Gemini all have free or low-cost entry points for Layer 2. Pick 3 to 5 power users per layer, give them 30 days, and measure honestly. Two pilots running in parallel will tell you more than any vendor demo. Practitioner's Note, Not Consultant Advice In my own practice I run Claude daily as the frontier layer on top of Microsoft 365. Outlook, Teams, Excel, Word: all Microsoft. The hard work (analysis, drafting, building decks, agentic workflows) runs through Claude, and a meaningful slice of inbox and Teams interaction happens through Claude as the interface rather than inside Outlook or Teams directly. That is the configuration this guide is implicitly recommending, and I want to be transparent about it. If your team is Google-native and cost matters, Gemini-on-Workspace is a credible second choice. If your hardest work is browser-driven or consumer-facing, ChatGPT on top of Microsoft is the better fit. Both honest answers, both covered above. For the connector mechanics that make this practical, see the [Connect AI to Your Email Client](/resources/connect-ai-to-email) guide. ## Decision Checklist Run each layer through its checklist. Then look at the combination. Layer 1: Productivity Backbone - We are already paying for Microsoft 365 (or Google Workspace, or Zoom Workplace) across the team - Our team's daily work happens in Outlook and Teams, or Gmail and Meet, or Zoom - We have specific UK or EU data residency requirements - We have 20 or more employees who would actively use AI features - Most customer-facing work happens in Word, Excel, PowerPoint (or Google Docs, Sheets, Slides) **Three or more ticked:** Microsoft 365 Copilot, Google Workspace with Gemini, or Zoom AI is your most likely Layer 1 answer. Plan a 30-day pilot before committing across the team. Layer 2: Frontier AI Model - Our hardest work is long-document reasoning, analysis, or writing at length - We have developers, analysts, or research-heavy roles - We need multi-step agentic workflows - Instruction-following quality matters more to us than the broadest plugin ecosystem - We can pay for frontier model subscriptions without needing them bundled into existing contracts **Three or more ticked:** Claude is your most likely Layer 2 answer. If your hardest work is browser-driven or consumer-facing and you need the largest plugin store, ChatGPT. If you are Google-native and want to minimise cost, Gemini. **Combining the layers.** M365 plus Claude is the canonical mid-market 2026 setup. Google plus Gemini is the cost-efficient equivalent. M365 plus ChatGPT suits browser-driven and consumer-facing work. Standalone Claude-only suits very small teams running lean. ## Where to Start The next step is not to sign a contract today. Map where your team's AI-relevant work happens, who would use each layer most, and which ecosystem your current contracts nudge you toward. Then run a 30-day pilot on each layer in parallel, 3 to 5 power users per pilot. Measure two things: did output get better, and did people keep using the AI features after the novelty wore off? If the answer to either is no, you have your answer before committing. If this guide convinced you that your team is not ready for either commitment yet, that is also a useful answer. Run standalone Claude or ChatGPT subscriptions for your power users, layer them onto existing tools where the integrations allow, and revisit in six months. ### This is what an AI strategy session actually covers Both decisions, picked separately and combined deliberately. Build vs buy for each layer. Data governance across both. Procurement timing. Our Executive AI Strategy sessions are built for leaders making exactly these calls. We work through your specific business, your existing software contracts, and the governance framework that fits. You leave with a 30-day action plan for each layer, not a slide deck. [Book a Free Discovery Call](https://calendly.com/practical-ai/30min) --- Most businesses pick one AI provider and run every AI task through it. The simple ones (e.g. what is the weather in London tomorrow), the complex ones (e.g. multi-step agentic workflows) - some of these a much smaller model could easily handle in a quarter of the time, and at a much more cost efficient rate. It feels straightforward though to just use one - less admin and overhead. However you are paying for convenience, and it is also more fragile (anyone seen the uptime for Anthropic API recently?), and potentially limiting you to a single supplier's roadmap and even their worldview and ethics. This is a decision guide for leaders thinking strategically about how their organisation should access AI. No code, no setup. Just a framework you can reference to reason about AI usage. ## One-Model Thinking Is Costing You If your business uses AI in any meaningful way, you have probably picked a favourite. ChatGPT. Claude. Gemini. Copilot. You bought the premium plan, bit of FOMO made you spend a bit too much (maybe) and now you are off to the races. You wired it into the workflows, you did the team education sessions and all-hands discussions. You moved on to the next business objective. This choice was reasonable a year ago. Today, it is the single biggest reason businesses are overpaying for AI. It is the same logic as sending every letter using priority next-day courier. Fine for the urgent contract. Wasteful if it is just a postcard. AI works the same way: different tasks need different models (even providers?), and you should not be paying frontier prices for jobs a far cheaper model handles just as well. ### The price gap is bigger than most realise: Here is what the market actually looks like today (June 2026), per million tokens of input and output: | Tier | Example models | Input | Output | | --- | --- | --- | --- | | Frontier | Claude Opus, GPT-5 | $5.00 | $25.00 | | Mid-tier | Claude Sonnet, GPT-5 mini | $3.00 | $15.00 | | Fast / cheap | Claude Haiku, Gemini Flash | $1.00 | $5.00 | | Open-weight value | MiniMax M2.7, DeepSeek | $0.30 | $1.20 | | Floor | Compact open models | $0.09 | $0.29 | That is the same task, with a 20 to 80 times difference in price between the top and the bottom. Most categorising, summarising, drafting, and tagging work does not need frontier intelligence. It needs adequate intelligence at a reasonable cost. Rule of thumb: only the highest-stakes work needs the most expensive model. Triage, classification, summarisation, and routine drafting can usually run on something 10 to 20 times cheaper with no quality difference your customers would notice. More recently running this locally on consumer hardware is also becoming more realistic and feasible for non-tech teams and business leaders. ## What an AI Router Actually Is An AI router (sometimes called an AI gateway) is a layer between your business and the AI providers. The simplicity is you have one account, one API key, one monthly bill, but access to two hundred or more models from every major provider through a single connection. The most well-known is OpenRouter. There are others (Portkey, LiteLLM, TokenMix). The category matters more than the brand: a router decouples your business from any single AI vendor and gives you control over which model handles which job. It is the difference between having a relationship with one supplier and having a procurement function. For a small business, that shift used to require enterprise budget. It no longer does. ## Three Strategic Benefits ### 1. Cost routing You can tell a router: "for this workflow, use the cheapest available model that meets my quality bar." OpenRouter calls this floor pricing. You name a small group of candidate models and let the router pick the lowest-priced one available right now. When prices shift (and they shift constantly), your bill quietly drops without anyone touching the code. For most small businesses running AI in their workflows, this single feature pays for the switch within a month. ### 2. Resilience If your chosen provider is down, rate-limiting you, or returns an error, the router automatically tries the next model on your list. You are only billed for the successful run. For a business that has built any kind of workflow on top of an AI API, this is the difference between a quiet Friday afternoon and an evening spent firefighting. Last year saw three separate multi-hour outages across the major AI providers. Every business running through a single endpoint went down with them. Every business routing through a gateway with fallback enabled stayed up. We expect this on our other enterprise applications - so why not AI? ### 3. No lock-in You can compare Claude against GPT against Gemini against MiniMax on your own work without rebuilding anything. Switch when prices change. Run a quiet A/B between two models on important outputs and keep both bills under fifty pounds a month while you decide. This is the part that should matter most. The AI market is moving fast. Pricing changes. Models get deprecated. Companies pivot. Building your workflows around a single provider is taking a bet on that provider's roadmap. Routing through a gateway means you can re-bet at any time. ## The Honest Caveats This is not a free lunch though, and there are three things you should know before deciding. ### There is a platform fee Routers do not run for free. OpenRouter charges roughly a 5.5% platform fee on pay-as-you-go usage. For most small businesses, the saving from picking the right model dwarfs the fee. For very high-volume workloads it is worth running the math both ways. Either way, it is rare to find a setup where the fee outweighs the saving from cost routing. ### Your data still flows through the provider A router does not change which providers see your data. It just adds a layer in front of them. If you are sending sensitive customer information, you still need to know exactly which providers can see what, and you still need to vet their data policies. Good routers let you allow-list specific providers and block the rest. Bad routers do not. Always check. ### It is a developer-grade tool You do not set this up by clicking a button. Either someone on your team wires it into your code, or you use a platform (Make, n8n, most modern AI workflow builders) that accepts a custom API endpoint. If your entire use of AI today is a ChatGPT Plus or Claude.ai subscription, a router will not help you yet. This is the next step up, not the first one. ## Do You Actually Need This? Whether or not you adopt a router this year, these four ideas should shape how you think about AI procurement from this point on. ### 1. Do not bet the business on one provider Every major AI provider has had outages, changed pricing without much notice, deprecated models that businesses had built on, and revised terms of service. A router is the cheapest insurance policy you can buy against any of that. Even if you only ever use one model in practice, knowing you could switch in a day changes the conversation. ### 2. Models are commodities. Prompts and skills are becoming your IP - treat them as such! The underlying model is becoming interchangeable. What is hard to replicate is the prompts your team has refined, the workflows you have built, and the institutional knowledge of which model to use for which task. Design for portability. The model you use this year is unlikely to be the model you use in 2028. ### 3. Start small, measure honestly Pick one workflow. Run it through a router for a month. Compare the bill, the quality, and the reliability against your current setup. Either you have saved money, or you have proven your current setup is the right one. Both outcomes are valuable. Both are cheap to learn. ### 4. Budget per model, not just per month A single monthly AI budget tells you nothing. Set spending caps per model and watch where the money actually goes. You will find the surprises quickly: the workflow that quietly costs ten times what it should, the experiment that nobody turned off, the team that defaulted to the most expensive model for tasks the cheapest one would have handled. Adopting and managing AI technologies well means treating it like any other line item: measure it, optimise it, and keep your options open. ## Where to Start If you read this and thought "we should be doing this," the next step is not to sign up to OpenRouter today. It is to map the AI work your business is actually doing, identify which workflows are running on the wrong model, and put a small pilot in place. That mapping is the strategic work. The technical setup is the easy part. If you read this and thought "we are not there yet," that is also a useful answer. It tells you the next twelve months should be about building up AI use across the business until the cost decisions start to matter. The guides on email, meeting notes, and workflow automation on this site are designed for exactly that stage. ### This is what an AI strategy session actually covers Cost routing. Build vs buy. Data governance. Which models for which tasks. Our Executive AI Strategy sessions are built for leaders making decisions like this one. We work through your specific business, your actual AI use, and the procurement and governance framework that fits. You leave with a 30-day action plan, not a slide deck. [Book a Free Discovery Call](https://calendly.com/practical-ai/30min) --- You likely spend over two hours a day reading, sorting, and replying to emails that follow the exact same patterns: pricing questions, meeting requests, and support queries. An AI inbox gatekeeper fixes this. It reads each email as it arrives, classifies the urgency, drafts a reply in your tone, and drops it into your Drafts folder. You review for 10 seconds, hit send, and move on. No code. No developers. Just your existing inbox and a workflow you can set up in 15 minutes. ## Two Approaches: Pick the One That Fits Before you start, it's worth knowing there are two ways to get AI help with email. They solve different problems. ### Approach 1: Built-in AI connectors (manual, per-email) Claude, Gemini, and some third-party plugins already connect directly to Gmail. You can ask them to read an email, draft a reply, or search your inbox. No workflow tool needed. **Claude** connects to Gmail through its Google Workspace connector (available on Pro, Max, Team, and Enterprise plans). You can ask things like "Draft a response to the vendor's pricing proposal" or "Are there any unanswered emails I should follow up on?" Claude reads the email, drafts a reply, and saves it to your Drafts. It won't send anything without your approval. **Gemini** is built directly into Gmail. "Help me write" generates drafts from a short prompt, and Suggested Replies offer one-click responses that match your writing style. If you have a Google Workspace or Google AI plan, Gemini can also summarise long email threads and pull details from your other emails and Drive files to personalise responses. **ChatGPT** doesn't have a native Gmail connector, but third-party Chrome extensions like "GPT for Gmail" add AI drafting directly inside your inbox. These are great for handling individual emails faster. But they're manual. You still have to open each email and ask the AI to help. ### Approach 2: Automated workflow (hands-off gatekeeping) This is what the rest of this guide covers. You connect a workflow tool to your inbox so that every incoming email is automatically read, classified, and drafted without you lifting a finger. The AI works in the background. You just review the drafts when you're ready. If you want AI help on individual emails as you go, the built-in connectors above are the simplest option. If you want emails classified and drafted automatically while you focus on other things, keep reading. ## What You'll Need Before you start, make sure you have these ready: | You'll need | Free option | Notes | | --- | --- | --- | | An AI account | Claude.ai, ChatGPT, or Gemini (all have free tiers) | Any of these will work. Pick whichever you prefer. | | A workflow tool | See options below | This connects your email to the AI. Think of it as the wiring. | | A Gmail or Outlook account | Your existing inbox | The workflow will read new emails and save draft replies. | | 15 minutes | Right now | Seriously, that's all this takes. | ### Choosing a workflow tool You only need one. Here are the options worth considering. **Zapier** is the simplest choice for most people. It has the largest library of app connections (8,000+) and the lowest learning curve. The free plan is generous: 100 tasks per month with no time limit, which is enough to get your inbox gatekeeper running and test it properly before deciding whether to upgrade. Paid plans start at $19.99/month. This is what we recommend if you've never built an automation before. **Make** (formerly Integromat) is a strong alternative if you want better value for money. The free plan gives you 1,000 operations per month, which is significantly more generous than Zapier. The visual builder shows your workflow as connected circles, which some people find easier to follow. Paid plans start at $9/month. **Relay.app** is the most AI-native option. You can describe what you want in plain language and it builds the workflow for you. It also has built-in "human-in-the-loop" approval steps, which is ideal for email: the AI drafts, you approve. Good if you find Zapier or Make intimidating. For the step-by-step instructions below, we'll cover Zapier. The process is similar in Make and Relay.app. ## Step 1: Write Your Classification Prompt This is the instruction you'll give the AI. It tells the AI how to read each email, sort it into a category, and draft a reply. You don't need to get this perfect on the first try. Start with something close and refine it over the first week as you review the drafts. Here's a prompt you can copy and adapt. ### For Claude Copy Classification & drafting prompt You are my email triage assistant. You've worked with me for years and understand my communication style: professional, direct, and friendly without being overly formal. When I give you an email, do two things: 1. Classify it into one of these categories: - URGENT: Needs my personal response today - STANDARD: Needs a reply but not time-sensitive - LOW: Newsletter, notification, or FYI only - SALES: Unsolicited pitch or cold outreach 2. Draft a reply (skip this for LOW and SALES emails). Rules for drafting: - Match my tone: concise, warm, no corporate waffle - Keep replies under 100 words unless the topic genuinely needs more - If I need to check something before replying, say so in the draft with [CHECK: what I need to verify] so I can spot it instantly - If a meeting is being requested, suggest a specific time rather than "let me know when works" - Never commit me to a deadline, cost, or deliverable. Flag these with [CONFIRM BEFORE SENDING] Respond in this format: CATEGORY: [category] DRAFT: [your draft reply] FLAGS: [anything I should check before hitting send, or "None"] ### For ChatGPT Use the same prompt above. ChatGPT handles this format well. One small addition: if you're using the free tier, add this line at the top of the prompt: Keep your response concise. Do not explain your reasoning or add commentary. Just give me the category, draft, and flags. This stops ChatGPT from adding preamble like "Sure! Here's my analysis of this email..." which clutters the output. ### For Gemini Same prompt works. If you're using Gemini through Google Workspace, you already have "Help me write" and Suggested Replies built into Gmail. Those handle individual email drafting well (see Approach 1 above), but they don't classify emails or flag decisions for you. The workflow approach below gives you both classification and drafting, running automatically in the background. ## Step 2: Connect It to Your Inbox This is where the workflow tool comes in. You're building a simple three-step chain: New email arrives → AI reads and classifies it → Draft reply saved ### Using Zapier 1. Go to zapier.com and sign in (or create a free account). 2. Click **Create a Zap**. 3. Set up the trigger: - App: Gmail (or Outlook) - Event: **New Email** - Choose the mailbox or label you want to monitor. - Tip: start with a specific label like "To Process" rather than your entire inbox. Move emails there manually for the first week while you test. 4. When Zapier asks for Gmail permissions, only tick what the workflow actually needs. The screen looks like this: Tick only "Read, compose and send emails". Leave contact access off — the gatekeeper only needs to read incoming mail and create drafts. Limiting permissions is the simplest way to keep risk low while you're testing. 1. Add an AI step: - App: **AI by Zapier** (built in, no extra account needed) or **Claude** or **ChatGPT** (requires connecting your account). - If using AI by Zapier: paste your classification prompt in the Instructions field. Map the email subject and body into the input. - If using Claude or ChatGPT directly: select **Send Message** as the action. Paste your prompt, then add the email subject and body as variables. 2. Add the final action: - App: Gmail - Event: **Create Draft Reply** - Map the AI's output into the draft body. - The draft will appear in the correct email thread, ready for your review. That's it. Three steps. Test it by sending yourself an email and watching the draft appear. ## Step 3: Add a Filter (Optional but Recommended) You probably don't want the AI drafting replies to every single email. Add a filter between the email trigger and the AI step to skip: - Emails from yourself (CC/BCC loops) - Newsletters and automated notifications - Emails already in specific folders (Spam, Promotions) In Zapier, use a **Filter** step. In Make, click the line between modules and add a filter condition. A simple starting filter: only process emails that land in your Primary inbox (Gmail) or Inbox folder (Outlook), and skip anything where you're in CC rather than To. ## Step 4: Refine Over the First Week Your first few drafts will be decent but not perfect. That's expected. Here's how to improve them quickly. **Days 1 to 3: Review every draft before sending.** Notice patterns in what the AI gets wrong. Is it too formal? Too wordy? Missing context about your business? **Day 4: Update the prompt.** Add specific instructions for the mistakes you keep seeing. For example: - "When someone asks about pricing, always direct them to practical-ai.space/resources rather than quoting a number." - "Keep replies to clients under 3 short paragraphs." - "If the email is from [specific client], always copy in [colleague name]." **Day 7: Check your categories.** Are emails landing in the right buckets? If "STANDARD" is catching things that should be "URGENT," add clearer definitions. For example: "URGENT means: the sender is a current client, the email mentions a deadline within 48 hours, or the subject contains words like 'issue,' 'broken,' or 'ASAP.'" The prompt is a living document. The more specific you make it, the better the drafts get. ## What to Watch Out For **Never auto-send.** Always review before sending. The AI will occasionally misread tone, miss context from a previous conversation, or draft something that's technically correct but not what you'd actually say. The "draft and review" approach gives you a 10-second check rather than a 10-minute writing session. **Be careful with sensitive information.** Don't pipe emails containing financial data, passwords, personal health information, or anything confidential through a third-party AI. If your inbox handles sensitive data, set up filters to exclude those emails from the workflow. **Start small.** Don't connect your entire inbox on day one. Start with one email category (support emails, or a specific label) and expand once you trust the system. ## What You've Just Built Congratulations. You've built your first AI agent. Not the sci-fi kind. The practical kind: an AI connected to your existing tools, given clear instructions, and set up to take action on your behalf. It reads, classifies, drafts, and waits for your approval. That's what an AI agent actually is. And you just built one without writing a single line of code. ## Three Ways to Extend This Once your inbox gatekeeper is running, you can expand it. 1. **Add a priority Slack notification.** When the AI classifies an email as URGENT, have the workflow send you a Slack message (or text) with a one-line summary. You'll catch critical emails even when you're not checking your inbox. 2. **Log to a spreadsheet.** Add a step that logs every classified email to a Google Sheet: date, sender, category, subject. After a month, you'll see exactly where your email time goes and which categories you could automate further. 3. **Build a second agent.** Now that you know the pattern (trigger → AI → action), apply it to meeting notes, weekly reports, or proposal drafting. The "Build Your First AI Workflow" guide walks through more examples. ### Ready to build agents specific to your business? This guide gives you the foundation. In our 1:1 sessions, we map your actual workflows, identify the biggest time sinks, and build custom agents together using your real tools and data. Most clients walk away with 2 to 3 working automations by the end of the session. [Book a Free Discovery Call](https://calendly.com/practical-ai/30min) --- Cloud AI tools like ChatGPT and Claude are powerful — but every prompt you type is sent to an external server. For sensitive business content like client contracts, financial data, or HR matters, that's a real exposure risk. Running an AI model locally keeps everything on your machine, with zero data leaving your device and no monthly subscription required. ## What you need - A Mac (Apple Silicon M1 or later) or Windows PC with 8 GB+ RAM - **Ollama** — a free tool that runs open-source models locally - About 5 GB of free disk space for the model files ## Step-by-step setup 1. Go to **ollama.com** and download the installer for your operating system. The install takes about two minutes. 2. Open Terminal (Mac) or Command Prompt (Windows) and run: `ollama pull gemma4` This downloads Google's Gemma 4 model — roughly 5 GB. Go make a coffee. 3. Once the download completes, start a conversation: `ollama run gemma4` 4. Type your prompt and press Enter. Your conversation stays entirely on your machine. ## Which model should you use? - **gemma4** — Google's latest model. Best all-rounder for writing, summarising, and Q&A. Start here. - **mistral-small** — Mistral's efficient model. Faster on older hardware and still very capable for business tasks. - **qwen3:4b** — Alibaba's compact model. Excellent if your machine has limited RAM. ## Using the chat interface The Ollama desktop app includes a built-in chat window — no extra tools required. Once Ollama is installed, open the app and you'll find a ChatGPT-style interface where you can type prompts, paste documents, and switch between models. Your conversations stay entirely on your machine. ## Practical uses once you're up and running - **Summarise confidential documents** — paste a contract or board report and ask for a plain-English summary - **Draft sensitive communications** — HR letters, disciplinary notes, internal memos - **Review financial documents** — flag risks or anomalies without sending data to a third party - **Quick Q&A** — ask questions about a document you've pasted in, without it leaving your machine ### Want to put this into practice? Our training sessions walk through this setup live and show you how to use local AI in the context of your specific business. [Book a Free Discovery Call] --- > "You don't need a new email app to save 45 minutes a day. You just need AI to summarise those 40-message threads before you even start reading." Most of us lose close to two hours a day on email. AI tools can help claw back at least 45 minutes of that - summarising long threads before you read them, drafting replies in your tone, and flagging what actually needs you. The best part: you don't have to switch email apps or learn anything new. The tools are already inside Gmail and Outlook. You just need to know which buttons to press. ## The two main approaches You can either add AI to your existing email client via a plugin, or switch to an AI-native client. Both work well - it depends on how much you want to change your setup. Use Gmail? Start with Gemini. Use Outlook? Start with Copilot. Want the most private option, or use both? Use Claude. ## The Easiest Path: Upgrading Your Current Inbox - **Gemini for Gmail** - A solid starting point for AI in email and calendar. Works on personal Gmail accounts out of the box - you have probably already seen this in your account. Click the Gemini icon in any thread to get a summary, a draft reply, or answers to questions about the email content. It can also reference your Google Calendar - try asking "Tell me about my priority meetings today." No setup required. Not ready to use it on your main account? Create a free Gmail account to experiment first before applying it to sensitive data. [Google's setup guide →](https://support.google.com/mail/answer/16831098?hl=en) - **Claude: the privacy-first option** - Worried about AI reading sensitive emails? Claude lets you grant permissions one at a time. Start with "read only" - it can summarise and answer questions about your inbox but literally cannot send, delete, or reply. Upgrade permissions later when you trust it. Works with both Gmail and Microsoft 365. [Set up Gmail connector →](https://support.claude.com/en/articles/10166901-use-google-workspace-connectors) · [Set up Microsoft 365 connector →](https://support.claude.com/en/articles/12542951-enable-and-use-the-microsoft-365-connector) - **Copilot for Outlook** - Built into Outlook 365. Summarises threads, drafts replies, and coaches writing tone. Requires a Microsoft 365 Copilot licence. Not the sharpest AI on the market, but if you're already in the Microsoft ecosystem it's a low-friction place to start. ## Quick start: Gmail + Gemini (5 minutes) The Gemini panel opens on the right side of Gmail - type your question in the box at the bottom. *Image: Google Workspace* 1. Open Gmail in Chrome and find a long thread you've been putting off. 2. Click the **Gemini star icon** on the right side of the screen. 3. Click **"Summarise this email"** - you'll get a 3-bullet summary in a few seconds. 4. Click **"Reply to this email"**, describe the tone you want, and let Gemini draft the response. 5. Edit if needed and send. You've just processed that email in under a minute. The biggest time saving isn't drafting replies - it's summarising. Use AI to reduce a 40-message thread to 5 bullet points before you reply. This alone saves most executives 30–45 minutes a day. ## Control your calendar from the same chat window Gemini can read and write to Google Calendar without leaving your Gmail tab. Type `@Google Calendar` at the start of your prompt to target it directly: - **Check your schedule:** *"@Google Calendar what events do I have this Friday?"* - **Create an event instantly:** *"@Google Calendar schedule a 30-minute sync for tomorrow at 2 PM - description: Review final assets."* Gemini creates the event and returns a direct link so you can verify or undo it in one click. Not sure which tool fits your company's security policies? [Book a free 15-minute call - no software demo or pitch, just answers to your specific questions →](https://calendly.com/practical-ai/30min) ## What to delegate to AI vs. what to write yourself - **Let AI draft:** status updates, scheduling replies, acknowledgements, routine follow-ups - **Write yourself:** anything sensitive, performance-related, or where the relationship is the point - **Let AI summarise:** every thread over 5 messages before you reply ### Want to cut your inbox time in half? Our training sessions show you how to set this up for your specific email client and build habits that stick. [Book a Free Discovery Call] --- This 5-minute guide turns raw meeting transcripts into structured, actionable data your team can use today. No coding or paid tools required. ## The Problem with Native AI Summaries Replacing a 60-minute meeting with a 3-page AI summary doesn't save time — it just creates another document nobody reads. Native summaries from Teams, Zoom, or Meet are just unstructured data sitting in a silo. They don't update project trackers, assign tasks, or flag moved deadlines. They just exist. To actually reclaim your time, you need a workflow that isolates the signal from the noise. Here is the exact process to turn your conversations into execution. ## What You'll Build By the end of this guide, you'll have a repeatable process that takes any meeting transcript and extracts: - **Action items**: who needs to do what, and by when - **Decisions made**: what was agreed, so nobody re-debates it next week - **Open questions**: things that still need answers - **Blockers**: anything that's going to slow the work down - **Key dates**: deadlines, milestones, and follow-up meetings The output is structured and clean. Ready to paste into whatever you use to manage your work: Trello, Asana, Monday, Notion, a spreadsheet, or just an email to your team. ## Step 1: Get the Transcript You probably already have this and don't realise it. **Microsoft Teams:** After a recorded meeting, go to the meeting chat or the recap page. Click on the transcript tab. You can download it or copy the full text. **Zoom:** If cloud recording is enabled, Zoom generates a transcript automatically. Find it under Recordings in your Zoom account. You can download the .vtt file and open it in any text editor, or copy the text from the Zoom web portal. **Google Meet:** Transcripts appear in your Google Drive after the meeting (in a folder called "Meet Recordings"). Open the Google Doc and copy the text. ## Step 2: The Extraction Prompt This is the step that actually matters. Copy the prompt below and save it somewhere you can grab it before every meeting. Open ChatGPT, Claude, or whichever AI tool you prefer. Paste the prompt, then paste your meeting transcript underneath it. The prompt: copy and save this You are a project coordinator. I'm going to give you a meeting transcript. I need you to extract structured data from this meeting - not a summary, not a narrative, just the specific items listed below. For each section, if nothing relevant was discussed, write "None identified." --- MEETING OVERVIEW - Date: - Attendees: - Meeting purpose (one sentence): ACTION ITEMS For each action item, provide: - Task: [what needs to be done] - Owner: [who is responsible - use their name from the transcript] - Deadline: [if mentioned; otherwise write "Not specified"] - Priority: [High / Medium / Low - infer from context and urgency of discussion] DECISIONS MADE - List each decision that was agreed upon during the meeting - Include who made or approved the decision if clear from context OPEN QUESTIONS - List any questions that were raised but not resolved - Note who raised them if clear BLOCKERS - List anything discussed that is currently preventing progress - Note the impact if it was mentioned KEY DATES & MILESTONES - List any dates, deadlines, or milestones mentioned - Include what they relate to FOLLOW-UP - Next meeting date/time (if discussed): - Items to prepare before next meeting: ## Step 3: Review the Output The AI will return a structured breakdown of your meeting. Scan it for accuracy. This should take about 2 minutes. **Things to check:** - **Are the action item owners correct?** The AI is usually good at matching names to tasks, but occasionally it'll assign something to the wrong person if the conversation was unclear. - **Are the deadlines right?** If someone said "end of next week" the AI will interpret that based on context. Make sure it matches what was actually meant. - **Did it miss anything?** If a key decision or task is missing, you can just tell the AI: "You missed the discussion about [topic]. Please add any action items from that section." No need to re-paste the transcript. ## Step 4: Put It to Work Now you have structured data instead of a document. This is where it gets practical: **If you use a project management tool** (Trello, Asana, Monday, Notion, etc.): Copy the Action Items section. Each item already has an owner and deadline. Paste them straight into your board or task list. **If you use a shared spreadsheet:** Ask the AI to reformat the action items as a table with columns: Task, Owner, Deadline, Priority, Status. Copy and paste into your sheet. **If you just need to email your team:** Ask the AI: "Reformat this as a brief email to the team summarising what was agreed and who owns what. Keep it under 200 words." You'll have a ready-to-send follow-up in 30 seconds. ## Making This a Habit The whole process (copying the transcript, pasting it with the prompt, reviewing the output, and distributing it) takes about 5–8 minutes after each meeting. Compare that to the 20–30 minutes most people spend writing up meeting notes manually. Or, more realistically, compare it to the current approach of not writing anything up and hoping everyone remembers what was agreed. **Two tips to make it stick:** 1. **Save the prompt somewhere accessible.** Pin it in your notes app, save it as a text snippet, or bookmark it. The friction of finding the prompt is usually what stops people from using it consistently. 2. **Send the output within 10 minutes of the meeting ending.** The value of meeting notes drops dramatically with time. If you send a structured follow-up while the conversation is still fresh, you look organised and your team has clarity. If you send it two days later, nobody reads it. ⚡ Pro method ## The 10x Method: Automate the Pipeline with a Claude Skill If you use Claude Pro (from £17/month), you can skip the copy-paste step entirely by turning this extraction process into a **Claude Skill**. A Skill is different from saving a prompt. It's a small instruction file that teaches Claude a repeatable workflow. Once it's installed, you don't need to open a special project or remember to paste a prompt. You just paste your transcript into any new Claude conversation and Claude recognises what it is, loads the Skill, and runs the extraction automatically. Set it up once. Never think about it again. **What you'll need:** - A Claude Pro, Max, Team, or Enterprise account - Code Execution enabled in your Claude settings (Settings > Features) - About 5 minutes **Step 1: Create the Skill file** Don't worry about reading this. Create a blank text file called `SKILL.md`, click Copy, and paste it in. Copy SKILL.md --- name: meeting-data-extractor description: Extracts structured project data from meeting transcripts. Use when the user pastes a meeting transcript, video call notes, or conversation log and wants action items, decisions, blockers, and deadlines pulled out in a structured format. --- # Meeting Data Extractor ## When to use this skill The user has pasted or uploaded a meeting transcript, call notes, or conversation log and wants structured data extracted from it. They may say things like "extract the actions from this meeting", "what were the decisions", "pull out the deliverables", or simply paste a large block of transcript text. ## Instructions You are a project coordinator. Extract structured data from the meeting transcript provided. Do not write a summary or narrative. Extract only the specific items listed below. For each section, if nothing relevant was discussed, write "None identified." ### Output format **MEETING OVERVIEW** - Date: - Attendees: - Meeting purpose (one sentence): **ACTION ITEMS** For each action item: - Task: [what needs to be done] - Owner: [who is responsible - use their name from the transcript] - Deadline: [if mentioned; otherwise "Not specified"] - Priority: [High / Medium / Low - infer from context and urgency] **DECISIONS MADE** - Each decision agreed upon during the meeting - Include who made or approved the decision if clear **OPEN QUESTIONS** - Questions raised but not resolved - Who raised them if clear **BLOCKERS** - Anything currently preventing progress - Impact if mentioned **KEY DATES & MILESTONES** - All dates, deadlines, or milestones mentioned - What they relate to **FOLLOW-UP** - Next meeting date/time (if discussed): - Items to prepare before next meeting: ## Guidelines - Use names exactly as they appear in the transcript - If a task has no clear owner, flag it as "Owner: Unassigned" - Keep each action item to one clear sentence - If the transcript is unclear or ambiguous, note that rather than guessing - After producing the structured output, ask the user if they want it reformatted for a specific tool (Trello, Asana, email, spreadsheet, etc.) **Step 2: Create the folder and zip it** Create a new folder on your computer called `meeting-data-extractor`. Put the `SKILL.md` file inside it. Then zip the folder: - **Windows:** Right-click the folder > Compress to ZIP file - **Mac:** Right-click the folder > Compress You should now have a file called `meeting-data-extractor.zip`. **Step 3: Upload to Claude** 1. Go to [claude.ai](https://claude.ai) 2. Open **Customize** (click your profile icon, then Customize, then Skills) 3. Click the **+** button, then **+ Create skill** 4. Upload your `meeting-data-extractor.zip` file 5. Toggle the skill on That's it. The Skill is now active on your account. **Step 4: Test it** Open a new conversation in Claude. Paste any meeting transcript. Claude should automatically recognise it as a transcript, load the Skill, and produce the structured extraction without you needing to add any instructions. If it doesn't trigger, check that the Skill is toggled on in your Customize > Skills list, and that Code Execution is enabled in Settings > Features. You just built a custom AI extraction engine. From now on, your meeting admin takes 30 seconds. You can build separate Skills for different meeting types. Use the variant prompts from the "Going Further" section below as the basis for a `client-meeting-extractor` Skill, a `standup-extractor` Skill, and so on. Same process: create the SKILL.md, zip it, upload it. **What about ChatGPT?** ChatGPT Plus ($20/month) offers Custom GPTs, which serve a similar purpose. You save the instructions once and reuse them, but you need to remember to open the right GPT. The Claude Skill approach is more hands-off because Claude loads it automatically based on what you paste. Don't have a paid plan? The copy-paste method from Step 2 works perfectly on every free tier. ## Going Further: Custom Prompts for Different Meeting Types Once you're comfortable with the basic extraction prompt, you can create variations for different meeting types. Here are three to get you started: **Client meetings:** add this to the prompt: Also extract: - Client concerns or objections raised - Commitments we made to the client - Any scope changes discussed (and whether they were agreed or need approval) **Team standups / check-ins:** simplify the prompt to: Extract from this meeting transcript: - What each person is working on - What's blocked - What needs to happen before next check-in Format as a table: Person | Working On | Blocked By | Next Steps **Strategy or planning meetings:** add this to the prompt: Also extract: - Goals or objectives discussed - Options considered (and which were chosen vs rejected) - Risks identified - Resource or budget implications mentioned ## Why This Works Better Than AI Meeting Summaries The built-in AI summaries from Teams, Zoom, and Meet give you a readable narrative of what happened. That's fine if you missed a meeting and need to catch up. But a narrative summary doesn't tell your project manager what to do next. It doesn't tell your client what was agreed. It doesn't tell you what's blocking progress. The difference is between **collecting information** and **extracting data you can use**. A summary collects. The extraction prompt produces outputs you can drop straight into your workflow. The transcript is your raw material. The AI summary packages it nicely. The extraction prompt refines it into something your team can actually work from. That's the difference between a document and a data pipeline, and the difference between AI that feels useful and AI that genuinely saves you time. ## Quick Reference Standard method: any AI tool, free tier | Step | What to do | Time | | --- | --- | --- | | 1 | Copy your meeting transcript from Teams, Zoom, or Meet | 1 min | | 2 | Paste the extraction prompt + transcript into ChatGPT or Claude | 1 min | | 3 | Review the structured output for accuracy | 2 min | | 4 | Paste action items into your project tool or email your team | 2 min | | Total | | ~6 min | Claude Skill method: Pro plan, set up once, auto-triggers | Step | What to do | Time | | --- | --- | --- | | 1 | Copy your meeting transcript | 1 min | | 2 | Paste it into any new Claude conversation | 30 sec | | 3 | Review and distribute | 3 min | | Total | | ~4 min | ### Want to do this with your own meetings, live? Our hands-on training sessions walk through this process with your real tools and your real data, so you leave with a workflow you'll actually use. [Book a Free Discovery Call](https://calendly.com/practical-ai/30min) --- An AI workflow is a chain of automated steps that runs without you. You set it up once and it runs forever — processing emails, generating summaries, filing information, sending alerts. No developer required, and the tools to build them are either free or very cheap. ## What is an AI workflow? A workflow is triggered by an event — a new email, a form submission, a scheduled time — and then takes a sequence of actions. For example: a new email arrives from a client → AI summarises it and extracts any action items → the summary is posted to your Slack channel. You see the key information without touching the email. ## The best no-code tools - **Zapier** — Connects 6,000+ apps with built-in AI steps powered by OpenAI. Has a free tier. Best starting point for most businesses — simple to use and well-documented. - **Make (formerly Integromat)** — More powerful than Zapier for complex multi-step logic. Steeper learning curve but much more flexible once you're comfortable. - **n8n** — Open-source and self-hostable. Best for teams that want full control and privacy, or want to avoid per-task pricing at scale. ## Build this in 20 minutes: daily email digest This workflow sends you an AI-generated summary of all emails received in the last 24 hours, every morning at 8am. Start your day knowing what matters without opening your inbox first. 1. Log into **zapier.com** and click *Create Zap*. 2. **Trigger:** Choose "Schedule by Zapier" → set to daily at 8am. 3. **Step 2:** Add a "Gmail" action → "Find Emails" → filter by last 24 hours. 4. **Step 3:** Add an "AI by Zapier" step. Use this prompt: `Summarise the following emails in 5 bullet points, highlighting anything urgent or requiring a decision: [email content]` 5. **Step 4:** Add a "Gmail" or "Slack" action to send the summary to yourself. 6. Click *Test*, then *Publish*. Done — it runs automatically from now on. ## Other workflows worth building - **New form submission → AI summary → CRM entry** — when someone fills in a contact form, AI extracts the key details and creates a contact record automatically - **Meeting transcript → action items → task manager** — pipe your Otter or Fireflies summary into Notion, Asana, or Monday automatically - **New invoice received → categorised → filed** — AI reads the invoice, categorises the spend, and files it in the right folder Start with one workflow that eliminates a task you do manually every day. Get it working, get comfortable with the tool, then expand. The most valuable workflows are the ones you don't have to think about. ### Want to automate the repetitive parts of your role? Our sessions map your specific workflows and build the first one together — so you leave with something working, not just a plan. [Book a Free Discovery Call] --- Most people type a vague request into ChatGPT, get a vague response, and conclude that AI isn't that useful. The problem isn't the AI — it's the prompt. The good news: there are two ways to fix it. One takes 10 seconds and requires no learning. The other gives you full control once you're ready. ## The shortcut: let AI interview you The lowest-effort way to get a better output isn't to craft the perfect prompt — it's to let the AI ask you the right questions first. Instead of trying to think of everything upfront, add one line to whatever you're working on: You type "I need to write a proposal for a new client. Before you start, ask me the questions you need to do this well." The AI will respond with something like: AI asks 1. Who is the client and what industry are they in? 2. What specific problem or project are you proposing to solve? 3. What is the approximate scope or value of this engagement? 4. What tone do you want — formal pitch or more conversational? 5. Are there any sections you need, such as a timeline, pricing, or case studies? You answer those questions in plain English — no formatting required — and the output you get will be dramatically better than anything you'd have got from "write me a proposal." This works because you already know your business context. The AI just needs to extract it from you. The interview short-circuits the most common failure mode: giving the AI too little to work with. Use this whenever you're not sure how to start. Paste: "I need to [your task here]. Before you start, ask me the questions you need to do this well." ## Want more control? Use the four-part framework Once you've had the AI interview you a few times, you'll notice a pattern: it's always asking for the same four things. Once you recognise them, you can skip the interview and give the AI everything upfront. 1. **Role** — Tell the AI who it's acting as. *"You are a senior commercial solicitor..."* 2. **Task** — Be specific about what you want. *"...review this contract clause..."* 3. **Context** — Give the relevant background. *"...for a £500k SaaS deal with a UK enterprise client."* 4. **Format** — Specify what the output should look like. *"Respond with bullet points: risks, questions to raise, and a plain-English summary."* Here's what that looks like in practice: Weak prompt "Write me an email about the project delay." Strong prompt "You are a project manager at a consulting firm. Write a professional email to a senior client — the CFO of a retail company — explaining a 2-week delay to our data migration project. The cause was a third-party API issue outside our control. Tone: accountable but confident. Include next steps and a revised go-live date of 12 May. Keep it under 200 words." The strong version took 20 extra seconds to write and will save you 10 minutes of editing. Notice how it already contains everything the AI would have asked you? That's the pattern — once you see it, you can't unsee it. ## Three prompts to save right now - `Summarise this in 5 bullet points for a non-technical audience: [paste text]` - `What are the 3 biggest risks in this document? [paste text]` - `Rewrite this in a more confident, direct tone. Remove any hedging language: [paste text]` - `I need to [task]. Before you start, ask me 5 questions that will help you do this well.` ## How to fix a bad output without starting over If the first response isn't right, don't delete everything and try again. AI responds well to follow-up instructions: - *"Make it shorter"* - *"Less formal"* - *"Focus only on the financial implications"* - *"Add a section on next steps"* Treat it like editing a first draft from a junior colleague — direct, specific feedback gets a much better second version. And if the output still isn't right after your follow-up, let the AI diagnose the gap itself: *"What else do you need to know about my situation to improve this?"* The most common mistake is being too polite. You don't need to say please or sorry. Be direct: "Rewrite the third paragraph. It's too wordy." The AI won't take offence. ### Ready to use AI in your specific workflow? Our sessions go beyond prompting basics — we work through your actual use cases so you leave with prompts that work for your role. [Book a Free Discovery Call] ---