Evidence on AI adoption, inside your AI assistant
Connect Claude, ChatGPT, Claude Code, Cursor or VS Code to ai-first.company. Your assistant then answers questions about adopting AI from rated practices, named company cases, checked statistics and studies, and links every claim to its source. Read-only, no login.
Questions managers ask before they spend money on AI. Ask them in your own words, in any language: the evidence is in English, and your assistant answers in yours.
Is it true that 95% of AI pilots fail? I want to put it in a board presentation.
You get: Where the figure comes from, who was asked, the verdict of a check against the source, and a sentence you can quote instead.
A vendor says its AI agents will cut our customer support costs by 40%. What should we check before signing?
You get: The questions to ask, companies that put AI into customer service and how it turned out, and the measured results to compare with.
Review our plan: ‘We will roll out Copilot to all staff in Q1. No pilot, baseline or staff training is needed.’
You get: Which do-first practices the plan commits to, rules out or leaves out, each with its evidence rating, and similar cases that went well or badly.
Prepare a one-page board brief: should we give Copilot to all 400 staff?
You get: What measured trials found, how strong the evidence is, what to do first, and companies that tried it, each with a link.
Which law firms got into trouble using AI, and what exactly happened?
You get: Named cases with court and regulator records, what went wrong and what the firm had in place.
We are a 40-person accounting firm. Where should we start with AI?
You get: The practices to take up first at your size, rated by evidence, and a firm like yours that has done it.
The assistant says so when the evidence has nothing on a question, rather than filling the gap from memory. A figure a company reports about itself is labelled as such.
Add the server once; your assistant uses it when a question calls for it.
https://mcp.ai-first.company/mcpGo to Customize → Connectors, click + Add, then Add custom connector. Enter a name and the server address, choose No sign in and click Add. In a chat, turn it on under + → Connectors. Free accounts can add one custom connector; on Team and Enterprise plans an owner adds it for everyone.
On the web, open Plugins, click +, then Add custom MCP server. Enter a name and the server address, choose No authentication, confirm the warning and select Create as a plugin. Install it, then type @ in a chat to use it. In a company workspace, an admin may have to add it.
Run claude mcp add --transport http ai-first https://mcp.ai-first.company/mcp --scope user. Leave out --scope user to add it to the current project only.
Add to ~/.cursor/mcp.json (or .cursor/mcp.json in a project): {"mcpServers":{"ai-first":{"url":"https://mcp.ai-first.company/mcp"}}}
Run MCP: Add Server from the Command Palette, choose HTTP, then paste the server address and name it ai-first.
Your assistant finds its way through the evidence by menus, the way you would on the site, so it does not depend on guessing the right search words.
Company cases by the kind of work AI was used for, six themes of practice, checks of widely repeated statistics, and what every rating means.
A theme, the cases on customer service or legal work, the figures on time saved. Each list shows how many rows it holds.
Figures with their caveats, the evidence rating, the primary source and the page on the site, plus the related studies and cases to follow.
Ten read-only tools. What each list holds and what every value means is in the guide, the same text the start tool returns.
start: Start here for any question about adopting AI in a company: the map of the ai-first.company evidence. Returns the company cases by the kind of work AI was used for (customer service, legal work, office assistant and others), the six themes (with counts of practices, cases and studies), the claim checks of widely repeated AI statistics, the groups of other figures, the company cases by outcome, the assessments by kind, and what each filter value and rating means. Then open the part that fits and fetch records with get_evidence.open: Open one part of the evidence map: a list (practices, cases, studies, figures, claim-checks, assessments) with filters such as theme, outcome, industry, company size, stage, priority, figure group (line) or assessment kind. Returns every matching row in one line with its id, or, when there are more than 25, a menu splitting them by the next filter, with counts. Use the arguments the menus show. Then fetch the rows you pick with get_evidence.get_evidence: Get the full records for ids from the menus and lists (such as "case:klarna-ai-customer-service" or "claim-check:mit-95-percent-of-ai-pilots-fail"): figures, ratings, verdicts, caveats, primary sources, the ai-first.company link, and the ids of related studies, cases and practices to follow. Up to 20 ids per call. Quote and cite only from what this returns.check_ai_figure: Before quoting an AI statistic (e.g. '95% of AI pilots fail', 'only 13% of companies are ready for AI'), check where it comes from, who was asked and what it covers. Returns every claim check of a widely repeated statistic and the groups of other figures; fetch the matching claim check, or open the group the figure belongs to. Use it for figures from surveys, reports and the press. For a claim a vendor makes about its own product, use check_vendor_claim.check_vendor_claim: Check a claim made by an AI vendor or consultant, such as '30% productivity gain', 'cuts support costs by 40%' or 'our agents replace 700 staff'. Returns the questions to ask the vendor, the claim checks, the groups of figures and the company cases by the kind of work AI was used for, so you can find the same result and companies that tried the same kind of tool. Use it whenever someone selling or proposing AI quotes a result. For a statistic from the press or a report, use check_ai_figure.review_ai_transformation_plan: Review an AI transformation proposal, business case or rollout memo, for example before it goes to a board. Returns the do-first practices for the company’s size as a checklist, each with its evidence rating and wording that may show it, and how to find similar cases. You read the plan and decide for each practice whether the plan commits to it, rules it out or leaves it out. Pass the size and industry; the plan itself stays with you. For one figure in the plan, use check_ai_figure.find_practices: List practices for adopting AI that fit a company's size, rated by strength of evidence. Use it for 'Where should we start with AI?' or 'What should a 120-person manufacturer do first?'. Pass the number of employees or a size band, and optionally the theme, stage or priority. Returns the practices in one line each, or a menu by theme when there are many; fetch the ones you use with get_evidence for their studies, cases and links.find_company_cases: List real companies that used AI and how it turned out: measured result, company-reported result, went wrong, or partly walked back. Use it for 'Who has done this?' or 'What went wrong for others?'. Filter by the kind of work AI was used for (task), outcome, theme, industry or company size; returns the cases in one line each, or a menu with counts when there are many. Fetch the ones you use with get_evidence for the story, who produced the figures and the public sources.find_readiness_assessment: List published questionnaires, tests and task maps to assess AI readiness, AI literacy, staff attitudes or which tasks suit AI, best tested first, optionally by kind or free to use. Fetch the ones that fit with get_evidence for how well each has been tested, who answers it, its length, licence and cost. For advice on what to do, use find_practices.board_brief: Prepare a one-page board brief on a question about AI adoption, from rated evidence only. Returns the company cases by the kind of work AI was used for, the themes (with practice counts for the company’s size), the claim checks and figure groups, and how to write the brief: open the cases for the work in question and one to three themes, fetch eight to twelve records and write from them. Use it when someone needs a summary for a board, an owner or an investment committee. To review a written plan, use review_ai_transformation_plan.The server never receives your questions: your assistant sends it only choices from its menus. It counts how often each tool, filter and record is used, and by which assistant app, and sends the totals to PostHog every ten minutes. It keeps no log of requests and stores no IP addresses or user IDs.
The same lists are free to download at ai-first.company/data, under CC BY 4.0. Credit: Data: CC BY 4.0, Paweł Szczęsny, ai-first.company. How the evidence is rated: how we rate.