Two quotes for a corporate AI workshop land on your desk. One trains your team on Microsoft Copilot, because that is what you license. The other covers ChatGPT, Claude, Gemini and Copilot, and costs more. The honest answer is that tool count is the wrong axis: a session built on one assistant can produce people who pick up the next one in twenty minutes, and a session that tours four can produce people who cannot do anything on Monday with any of them.
What separates the two is how much of the day goes to work that survives a tool switch. Deciding whether a task suits an assistant, assembling the context it needs, catching a confident wrong answer, knowing what may not be pasted anywhere: all of that holds when your company moves from Copilot to Gemini next year. Where the side panel lives and which file types upload does not. Price the proposals on that ratio and the comparison resolves quickly.
What each format puts in the room
Single-tool training is the default because the tool vendors fund it. Google’s AI Essentials program runs under five hours on a $49 monthly subscription and has you “practice using a conversational AI tool like Gemini.” Microsoft Learn carries Copilot learning paths at no charge, including a seven-module path covering Copilot in Word, Excel, PowerPoint, Teams and Outlook, plus a paid Copilot and Agent Administration Fundamentals exam. Both sets of pages read that way on October 22, 2026. Neither program is bad. Both are built to teach one product, because one product is who paid for them.
Paid multi-tool training sits on the other side. Correlation One, one of the larger enterprise providers, states that its programs “are designed to work with any generative AI foundation model.” The same page then says “every program is built for a specific platform,” naming ChatGPT, Microsoft Copilot, Google Gemini and Anthropic Claude as four separate builds. Engagements run three weeks to four months, with no published rate. Section markets coaching and names no AI product at all in the curriculum copy on its homepage. Across the rest of the market, the shape repeats: four modules covering AI basics, prompt practice, department use cases and data handling, offered as a half day, a full day, or a multi-week program.
Read enough of these pages and the pattern is hard to miss. Every one of them describes what it sells. None helps you choose.
The two layers of AI fluency
Split any curriculum in two and the decision gets easier.
| Layer | Contents | Survives a tool switch |
|---|---|---|
| Transferable | Picking tasks worth handing over, assembling context, giving the model the source material, checking output against the thing it claims, spotting fluent nonsense, knowing what cannot be pasted, rewriting after a weak first answer | Yes |
| Tool-specific | Menu paths, side-panel behavior, file size and type limits, how the memory or project feature is named, admin settings, which integration is turned on | No |
The transferable layer is most of the skill and almost none of the marketing, and it is where an AI enablement budget earns its return. The tool-specific layer is what screenshots show well, which is why it fills so much of the material on both sides of this comparison.
Training research has a name for the distinction. Blume, Ford, Baldwin and Huang’s meta-analysis of transfer of training, published in the Journal of Management in 2010, separates closed skills, which have one correct procedure, from open skills, where the trainee has to decide how to apply what they learned. Software instruction is their standard example of a closed skill, and the factors that drive transfer, including motivation and workplace support, tie more strongly to transfer for open skills than closed ones. Teaching menus is a closed skill. Teaching someone to judge when an assistant helps is an open one, and the open version is what reaches the work.
What transfers, according to the vendors’ own guides
Compare how the model makers tell you to write a prompt and the layers come apart on their own.
Anthropic’s public prompt-engineering documentation has a page devoted to wrapping parts of a prompt in XML tags, so instructions, context and examples stay separate. Google’s Gemini for Workspace guide teaches a four-part formula: persona, task, context, format. Microsoft’s own Copilot training module teaches its own four-element version: goal, context, expectations, source. Three vendors, three vocabularies, one instruction. All of them say to state the role or goal, hand over the real material, name the task, and say what shape the answer should take.
A workshop that teaches that structure and then shows the three dialects as house conventions produces someone who can write a usable prompt in any of the four assistants. A workshop that teaches one vendor’s house style as the rule produces someone who thinks their skill stopped working when the company switched. That is the whole argument for teaching the structure first, and our guide to designing a session that changes Monday morning works through how to sequence it in a room.
The genuinely tool-specific residue is real but thin. Context limits differ. Which assistant reaches your email and which does not is a live constraint. Admin settings decide what people can upload. Budget an hour for it, not a day. Which assistant to run that hands-on hour in is a separate call, and our guide to choosing the assistant a session should teach takes it apart.
Tool-specific knowledge has a short shelf life
Stanford’s 2025 AI Index put a number on how fast the products converge. On the Chatbot Arena leaderboard, the Elo gap between the top model and the tenth fell from 11.9% to 5.4%, and the gap between the top two models from 4.9% to 0.7%. Whatever one assistant did better last year, the others mostly do now.
Feature parity moves the same way. Document upload, project folders, web search, memory and side-panel access inside an office suite all arrived on one product first and then on the rest. Material built around a feature gap is dated within a few quarters, and material built around judgment is not.
This is the part a procurement conversation usually misses. You are not choosing between two curricula of equal durability. You are choosing between one that holds its value and one that needs reprinting every time a vendor ships, which is also the core of the case against treating a license rollout as AI enablement.
A day is a fixed budget
An AI workshop for teams buys a fixed number of hours, and that is the real constraint. There is evidence for where the threshold sits. BCG’s AI at Work 2025 survey, covering more than 10,600 leaders, managers and frontline staff across 11 countries, found 79% of people who had received more than five hours of training were regular AI users, against 67% below five hours. Its 2026 edition, with close to 12,000 respondents, found 74% of frontline employees now call themselves AI users while only 36% feel they have had adequate upskilling.
Five hours is roughly one working day with breaks. Spend that day on four product tours and each assistant gets about seventy minutes, enough for a demo and one exercise. Nobody leaves with a habit. Spend it on one assistant, on the team’s own documents, and then give the last hour to the differences, and people leave having done real work and seen that the method is not tied to the logo.
A tour is not hands-on AI training, whatever the proposal calls it, and the failure mode has a tell you can hear from the back of the room. In a tour, the facilitator drives and the room watches four screens. In a working session, the room is quiet because everyone is typing, and the questions are about their own documents rather than about the software. Which format a vendor is selling becomes obvious in the first thirty minutes, and the distinction holds whether the room is technical or not, as our look at which formats work for non-technical teams covers.
Half your room is already using a second assistant
The single-tool plan has a hole in it that usage dashboards do not show. Microsoft and LinkedIn’s 2024 Work Trend Index found 78% of AI users bring their own AI tools to work, rising to 80% at small and medium companies. People have a personal ChatGPT account and a work Copilot license and they use whichever answers faster.
Two consequences for the training decision. A Copilot-only session leaves the second tool much of your team drafts in untouched, including the risky part, since the assistant outside your tenant is the one with no retention controls. And a session that acknowledges the other assistants gets a more honest room, because nobody has to pretend. Teaching the shared method covers the sanctioned tool and the unsanctioned one in the same breath, which is the cheapest governance win available in a one-day format.
Four cases where single-tool is the right buy
The argument cuts the other way in four situations, and vendors selling breadth rarely say so.
A credential someone is asking for. When a client, an insurer or an auditor wants evidence of training on a named platform, a vendor certificate on that platform is the deliverable. Buy the thing that produces the document.
Deployment and administration. Configuring Copilot across a tenant, setting retention rules, deciding who gets which agent: all of that is Microsoft-specific work for two or three people, not team fluency. It belongs in a separate session with a smaller invite list.
A genuinely standardized small team. Twelve people, one tenant, strict data rules, no shadow accounts. Breadth buys little. Depth on the one tool, built from their files, buys a lot.
The free option is good enough. For baseline literacy on a single product, the vendor already published the course. Google AI Essentials runs under five hours on a $49 monthly subscription; Microsoft Learn’s Copilot paths cost nothing. Paying a facilitator market rates, $2,000 to $15,000 for a workshop, to deliver the same material is hard to defend. Make the free course pre-work with a date on it and spend the paid day on your own documents, which is the same information-gain test applied in our comparison of custom and off-the-shelf curricula.
If you are unsure which of these describes your team, a free AI-readiness call will sort it in half an hour.
What to demand from either vendor
Five questions, askable before you sign, that separate transferable instruction from a product tour.
- What fraction of the day is spent inside tool menus rather than on our work? Anything over a quarter is a tour.
- What will each person produce by the end, using their own documents as input?
- If we switch assistants in eight months, what in this material stops being true?
- How do you teach prompt structure so it holds across assistants, and where do you show the differences?
- Who facilitates, and have they done our kind of work with these tools rather than only taught them?
A vendor who answers question three with “nothing much” is describing transferable material. A vendor who cannot answer it is quoting a curriculum with an expiry date they have not thought about. Question five is the one most often skipped and it predicts the rest, as our comparison of practitioner-led and trainer-led delivery sets out.
One more move costs nothing. Ask both vendors to run the same exercise on an assistant they did not plan for. The multi-tool vendor who can do it on the spot has the method. The one who needs a week has four decks.
Frequently asked questions
Should an AI workshop teach one tool or several?
Teach one assistant in depth, on the team’s own work, and spend the last hour on what differs across the others. Depth produces people who can do the work; the closing hour produces people who know the skill is not locked to one product. Giving each of four assistants a quarter of the day produces a tour nobody applies.
Is multi-tool AI training worth paying more for?
Only when the breadth buys method rather than coverage. A session that teaches prompt structure once and then demonstrates it across assistants earns the premium. A session that runs four product demos is paying for slides. Ask what fraction of the day sits inside menus; that number tells you which one you are buying.
We standardized on Microsoft Copilot. Does multi-tool training still make sense?
Standardizing changes what gets demonstrated, not what gets taught. Run the hands-on work in Copilot, since that is where the work happens, and keep the method vendor-neutral so the skill survives your next procurement cycle. Add a short segment on the personal accounts people are already using, because those sit outside your tenant controls.
How much of prompting actually transfers between ChatGPT, Claude, Gemini and Copilot?
Most of it. Anthropic’s documentation recommends XML tags to separate the parts of a prompt, Google’s Workspace guide teaches persona, task, context and format, and Microsoft’s Copilot training teaches goal, context, expectations and source. Three vendors describing the same move: state the role or goal, supply the real material, name the task, specify the output shape. The vocabularies differ; the structure does not.
Does a single-tool certificate mean anything to a client or an auditor?
It evidences attendance on a named platform, which is sometimes the exact thing being asked for and a fair reason to buy one. It is not evidence that anyone’s work changed. If a contract or an insurer names a certificate, buy the certificate and run the capability work separately.
Won’t covering several assistants confuse people?
It confuses people when the session teaches four interfaces. It does not when the session teaches one method and treats the interfaces as a closing comparison. The order matters: method first, built on real work, differences last.
How many tools can one day realistically cover?
One in depth, plus a demonstration of two or three others in the final hour. BCG’s 2025 survey found regular use climbing past the five-hour mark, which is about one working day. Dividing that day four ways leaves each tool roughly seventy minutes, which covers a demo and one exercise.
Can we just use the free vendor training instead?
For baseline literacy on one product, yes, and you should. Google AI Essentials runs under five hours on a $49 monthly subscription and Microsoft Learn’s Copilot paths are free. What neither contains is your documents, your failed outputs, or your review path, which is the only material worth paying a facilitator for.
What does multi-tool AI training cost?
Workshops run about $2,000 to $15,000 at market rates, driven by room size, how much design happens before the session, and whether anything follows it. Tool count moves that less than most buyers expect, because the design cost sits in building exercises from your material. Enterprise providers quote by engagement, often three weeks to four months, and publish no rates.
What should we send a vendor before the session?
Five to ten documents your team handles weekly, redacted where needed, plus one AI output that went wrong and why. A vendor building transferable instruction will tell you how those change the exercises. A vendor running a tour will ask for a headcount, a date and your logo.
Key takeaways
- Tool count is the wrong axis. Judge a proposal by what fraction of the day survives a tool switch: task selection, context assembly, output checking and data judgment all do; menu paths do not.
- Single-tool training dominates because the tool vendors fund it. Google AI Essentials runs under five hours at $49 a month and Microsoft Learn’s Copilot paths are free, so any paid Copilot-only proposal competes with free.
- Tool-specific material depreciates fast. Stanford’s 2025 AI Index put the Chatbot Arena Elo gap between the top model and the tenth at 5.4%, down from 11.9%.
- The vendors’ own prompting guides prove the method transfers. Anthropic’s XML tags, Google’s persona-task-context-format formula and Microsoft’s goal-context-expectations-source framework are one instruction in three vocabularies.
- Hours are fixed. BCG found 79% of people with more than five hours of training were regular users against 67% below it, and a five-hour day split four ways leaves seventy minutes per tool.
- With 78% of AI users bringing their own tools to work, a single-tool session ignores the assistant part of your team already drafts in, including the one your controls do not reach.
- Single-tool is right when a credential is required, when the work is tenant configuration, when the team is small and truly standardized, or when the free vendor course covers it.
Pull both quotes and mark every agenda item as transferable or tool-specific before you look at the prices. If the cheaper proposal has more transferable hours in it, that is your answer, and a free AI-readiness call will tell you which hours your team is short of.
Arthur Wandzel
