Teach the assistant your team already has a license for and will open on Monday morning. If the company runs on Microsoft 365, that is Copilot. If it runs on Google Workspace, Gemini. If it runs on neither, or the work in question is mostly long documents, ChatGPT or Claude. Then teach it in a way that survives a vendor switch, because within two years there will probably be one.
That answer disappoints people who came looking for a benchmark table, and the pages ranking for this question will happily supply one. Context windows, coding scores, security certifications, a feature grid four columns wide. Every one of those pages is answering a purchasing question. You have already purchased. Often twice, which is part of the problem.
The question in front of you is narrower and more useful: which assistant should the room be practicing on for three hours, and how much does that choice change? Having run these sessions, our estimate is about a fifth of the agenda. The other four fifths are the same no matter whose logo is on the screen. That ratio is the thing worth knowing before you book anything, because it makes the AI workshop tool choice a logistics decision rather than a curriculum one.
Start from the license, not the leaderboard
Rank your inputs in this order, and stop at the first one that gives a clear answer.
- What is already paid for. A session built on a tool nobody can open after lunch teaches nothing.
- Where the work lives. If the Thursday reporting job runs through Excel and Outlook, an assistant inside Excel and Outlook has a head start that no benchmark score overcomes.
- What people are quietly using already. This one gets skipped, and it decides more than the first two.
- Model quality. Real, and the last tiebreaker rather than the first filter.
Point three deserves a moment. Microsoft’s Work Trend Index, published in May 2024 from a survey of 31,000 workers across 31 countries, found 75% of knowledge workers using AI at work and 78% of those users bringing their own tools, rising to 80% at small and mid-sized firms. More than half, 52%, were reluctant to admit using it on their most important work. The figures are two years old and the direction has only gone one way since.
So the room you are about to train is not a blank slate. It holds four or five people with real habits on a personal account, a dozen who have tried the suite assistant once and found it underwhelming, and a few who have touched nothing. Ask before you choose. Send a three-question anonymous form two weeks out: which tools have you opened, on whose account, for what task. In our experience that form changes the tool decision about a third of the time. Usually it reveals that the officially licensed assistant is not the one people have been getting results from.
What each of the four is good at in a first session
Published business terms, checked this month, and what each one is genuinely strong at when a group of 12 is doing real work on it.
| Assistant | Where it lives | Strongest first session | Published business price |
|---|---|---|---|
| Microsoft Copilot | Inside Word, Excel, Outlook, Teams, plus a chat surface | “Answer this from our own files and mail” | $23.50 per user per month for Copilot for Business, $30.00 for Copilot for Enterprise, both paid yearly |
| ChatGPT Business | Browser and desktop app, outside the suite | Broad mixed work, and the tool most people already know | $20 per seat per month billed annually, $25 monthly, two-seat minimum |
| Claude Team | Browser and desktop app, outside the suite | Long documents, careful drafting, structured analysis | $20 per seat per month billed annually, $25 monthly, two-seat minimum |
| Gemini in Google Workspace | Inside Gmail, Docs, Sheets, Meet, plus a chat surface | Mail and document work for a Workspace team | Included in paid Workspace tiers at list: $7 Starter, $14 Standard, $22 Plus per user per month |
Two notes that matter in the room rather than on the invoice.
Copilot’s real difference is not its writing. It is that it grounds answers in your own tenant. Microsoft’s own architecture documentation states that it reads organization data through Microsoft Graph and “only accesses data that an individual user is authorized to access.” That makes for a strong exercise, and it also turns the session into a test of your permissions hygiene. We have watched a Copilot demo surface a document the room was surprised anyone could see. That is a useful 10 minutes, though not the 10 minutes on the agenda.
Gemini arrives with the Workspace plan rather than as a separate purchase, since Google folded the old standalone add-ons into the base tiers. Google also sells Gemini Enterprise, a separate per-seat agent platform, which is a different product from the assistant in Gmail and Docs and not what a first fluency session needs.
Market share answers a question you did not ask
Most corporate AI workshop proposals arrive quoting enterprise market share. The numbers are real, and they are measuring the wrong thing.
Menlo Ventures’ 2025 enterprise survey was fielded with roughly 495 US decision-makers in November 2025. It put enterprise LLM API spend at 40% Anthropic, 27% OpenAI, and 21% Google, with Anthropic at an estimated 54% of the coding segment. That is developers wiring models into products. It tells you little about which chat window your accounts-payable clerk should have open.
The closer measure is seat spend, and there the Ramp AI Index is more useful. In September 2026, 55.9% of US businesses on Ramp paid for AI tools, with 43.8% paying Anthropic and 40.6% paying OpenAI. Those two shares add to well over the total, which is the finding worth carrying into your meeting: a large share of companies pay both. Multi-vendor is already normal. The comparison pages that present this as an exclusive choice are describing a market that does not exist.
The four fifths that transfer
Here is why the decision is lower-stakes than the search results imply. Strip a good session down and most of it is craft, not menus.
Framing the task. Saying what you want, for whom, in what form, with what constraints. The single highest-value habit, and identical across all four.
Supplying context. Giving the assistant the source material instead of hoping it knows: the actual contract, the actual export, the actual thread. The mechanics differ slightly, the discipline does not.
Verifying. Knowing which parts of an answer need checking and against what. Any assistant will produce a confident number that is wrong. The checking habit is the output of the session.
Data boundaries. What may be pasted into which tool, and on whose account. Taught on the letter someone just drafted, this sticks. Taught as a policy slide, it does not.
Iterating and keeping what worked. Two saved prompts a person wrote themselves, on their own documents, beat a handout of fifty.
None of that is vendor-specific, which is fortunate, because the vendors keep moving. OpenAI lowered its annual Business seat price in 2026. Google retired the standalone Gemini add-ons and put the assistant in the base plans. Microsoft renamed and repackaged its suite assistant, with Copilot now bundled into its top enterprise tier. A curriculum keyed to menu positions expires faster than the purchase order. This is one of the arguments for building the session around your own work rather than a catalog module: your documents are stable, the product UI is not.
The fifth that is tool-specific
The remaining piece is real and worth doing properly. In a three-hour session it is roughly 30 to 40 minutes.
- Where the thing lives. The sidebar in Word, the button in Gmail, the separate app. Whether it is in the flow of work or one tab away changes daily use more than model quality does.
- How material gets in. Upload, paste, link, or automatic grounding in files you already have access to. Suite assistants reach your content by default; standalone assistants need you to hand it over.
- What it can see. Copilot sees what the signed-in user is permitted to see. A standalone assistant sees what you give it. That gap drives its own exercises and its own warnings.
- Where the controls are. Who administers it, what is retained, what is shared across the team.
- What it will not do. Every assistant has a weak spot. Naming it early buys credibility and prevents the quiet disappointment that kills adoption in week two.
Teach that fifth last, after people have already produced something useful. Lead with the tool tour and you spend the energy of the room on software rather than on their own work, which is the most common design mistake in these sessions.
When two tools are right, and when it is noise
Two is sometimes correct. Four never is.
Teach two when a named function has truly different work: engineers whose assistant lives in the editor, legal reviewing 80-page documents, a finance team whose job is entirely inside spreadsheets. Teach two when you are mid-migration and both tools will be live for six months. Teach two when one is the suite assistant everyone has and one is the heavy-lifting tool a small group needs.
In all three cases, separate them. Different rooms, or clearly separate blocks with their own practice time. Comparing two assistants side by side on the same prompt is a fun demo and poor teaching, because the audience watches the scoreboard instead of learning either tool.
The case against four is arithmetic. Every extra assistant costs demonstration time, and demonstration time comes out of practice time, which is where the learning happens. A three-hour session with four tools is a product showcase with laptops open. One function, one primary tool, one task, and the hour of corrected practice that separates a session from a seminar.
Deciding it in one meeting
This should take 45 minutes, not a procurement cycle.
- Pull the license list. Not the plan you think you have. The actual assigned seats, and how many are unused.
- Ask the room, anonymously. Which tools they have opened, on whose account, for what task.
- Name the task the session will teach. One recurring job, weekly or more often, that people dislike.
- Check which licensed assistant reaches the data that task needs. If the source material is in SharePoint and Outlook, that narrows quickly. If it is PDFs people download anyway, it does not narrow at all and you may pick on quality.
- Pick one primary, and name a second only if step three produced two separate tasks. Then tell your facilitator two weeks out and send sample documents.
Step five is where committees go wrong. A group asked to choose an assistant will choose all of them, because nobody wants to tell a colleague their preferred tool is off the agenda. Give the decision to one person who owns the outcome, which is also the right answer for who should stand at the front of the room.
What the choice does to the budget
Worth seeing the two line items side by side, because the ratio surprises people.
Sixty seats at $20 to $30 per user per month is $14,400 to $21,600 a year, recurring. A private hands-on workshop for the same group runs roughly $2,000 to $15,000 at market rates, once. The AI enablement line is the smaller of the two, and it is the one that decides whether the larger one returns anything, and BCG’s AI at Work 2026 survey, fielded with close to 12,000 employees and managers in more than a dozen markets, still found only 36% feeling adequately upskilled.
One rule follows from that: do not buy new licenses to match a workshop. Build the workshop on the licenses you already pay for, see what the team does with it over a quarter, and let real use decide whether a second platform earns its seats. Buying a tool because a vendor’s curriculum assumes it is backwards, and it is how companies end up with three assistants and no capability in any of them.
Frequently asked questions
Which AI assistant should we train our team on?
The one you already license and the team will open without being reminded. Microsoft 365 companies should start with Copilot, Google Workspace companies with Gemini, and companies on neither, or whose work is mostly long documents, with ChatGPT or Claude. Confirm the choice against what people are already using on personal accounts before you book anything.
Should a workshop teach more than one AI tool?
One primary tool per room. Two is defensible when a named function has truly different work, such as engineers or a legal team reviewing long contracts, and even then teach them in separate blocks with their own practice time. Four assistants in one session is a product showcase.
Does the choice matter if we might switch vendors later?
Less than you would think, as long as the session is built on your own work rather than on menus. Task framing, context supply, verification, data boundaries, and saving what worked carry across all four assistants. Budget roughly a fifth of the agenda for the tool-specific part, which is the only piece a switch invalidates.
We run on Microsoft 365. Is Copilot enough for a first session?
Usually yes, and it has one advantage the standalone assistants cannot match: it answers from your own files and mail, scoped to each user’s permissions. Two cautions. Its answers are only as good as your SharePoint and Outlook hygiene, and a grounded demo occasionally reveals that permissions are looser than anyone thought.
Half our team already uses ChatGPT on personal accounts. What then?
Treat that as the most useful thing you know going in. People are getting results somewhere, and a session on a different tool asks them to abandon a working habit. Either teach what they use, on properly licensed business seats, or be explicit about why you are moving them and what the new tool does better.
Is Claude or ChatGPT the better choice for a workshop?
For a general fluency session, close enough that price and familiarity should decide it, and both sit at $20 per seat per month on annual billing. Claude tends to suit long-document and careful-drafting work; ChatGPT has broader built-in extras and is the tool more people have already seen. Pick the one the group will keep opening.
Do these tools train on our data?
All four vendors state in their published business and enterprise terms that customer business data is not used to train their models by default. Read the current terms for the exact plan you hold before repeating that in a room, since these pages change. Then make the gap between business seats and personal accounts explicit in the session. That difference is the one people get wrong.
Who should make the call, IT or the function that needs the training?
The person accountable for the outcome, with IT holding a veto on security and licensing rather than a vote on preference. Tool choice made purely by IT lands on whatever is cheapest to administer; made purely by the function, it lands on whatever one vocal person likes.
What if we have no business licenses yet?
Buy two seats of one assistant, give them to the two people who will use them hardest, and wait a month. A room trained on a tool nobody has yet retains nothing, and a month of real use will tell you more about which platform fits your work than any comparison, including this one.
Can one session cover a team that uses different assistants?
It can if you teach the craft and treat the tool as the variable. Run the exercises on the same task, let people use whichever licensed assistant they have, and cover the tool-specific mechanics in small groups. Weaker than a single-tool room, better than cancelling.
Key takeaways
- Pick the assistant you already license and the team will open unprompted. Where the work lives beats benchmark scores, and both beat vendor preference.
- Ask what people are quietly using before you choose. Microsoft’s survey put bring-your-own AI at 78% of AI users and 80% at smaller firms, and the answer often changes the decision.
- Roughly four fifths of a good session transfers across every assistant. Keep the tool-specific part to about 30 minutes of a three-hour session, taught after people have produced something real.
- Enterprise API market share, where Menlo put Anthropic at 40% and OpenAI at 27%, is a developer metric. Seat data is the relevant one, and Ramp’s September 2026 numbers show most companies paying more than one lab.
- Build the session on licenses you already hold. Sixty seats cost $14,400 to $21,600 a year, while the AI enablement work that makes them pay off costs $2,000 to $15,000 once.
Open the admin console, count the assigned seats nobody is using, and name the weekly task you want moved. A free AI-readiness call will tell you which of your existing licenses the first session should be built on, and what to leave out of it.
Dirk Jan van Veen, PhD
