Book the session now. A team spread across Microsoft Copilot, ChatGPT and Claude does not need procurement to pick a winner first, because most of what a good session teaches is the same work in all three windows. Teach that part to the whole room at once, split the last stretch by tool, and treat what you watch people do as the evidence that settles the stack next quarter.
Almost every piece of advice on this runs the other way: map the tools, cut from twenty to five, standardize, then upskill the team onto what survives. The order sounds disciplined. In practice it parks AI enablement behind a procurement review nobody owns, while the licenses keep billing and capability stays flat.
The mixed stack is not the problem to solve before training. It is the most useful thing you have to train on.
The order most advice gets backwards
Consolidation content is written by firms that sell consolidation: SaaS management platforms, data catalogs, course libraries. The argument is sound on cost and wrong on sequence. A license report tells you how many seats you bought and who logged in. It cannot tell you which tool is doing real work, and that is the only input the consolidation decision needs.
It is also an incomplete census. Microsoft and LinkedIn’s 2024 Work Trend Index surveyed 31,000 knowledge workers across 31 countries through Edelman. It found 78% of people using AI at work were bringing their own tools, which Microsoft labels BYOAI, rising to 80% at small and mid-sized firms. Another 52% were reluctant to admit using AI on their most important tasks. Part of your stack is not on the license report at all, and a room where people practice openly surfaces more of it in three hours than a procurement questionnaire does in a month.
The same survey found only 39% of workplace AI users had been given any AI training by their employer. Stalling the session to tidy the stack adds three months to that gap for no gain.
What a three-assistant stack usually is
The three tools in a 50-to-500-person company rarely arrived as a plan. They arrived as three purchase events, and the pattern is consistent enough to predict.
| How it arrived | Typical tier | What it is doing | What training must account for |
|---|---|---|---|
| Microsoft Copilot, bundled into the Microsoft 365 renewal | Business, assigned by IT | Summaries in Outlook and Teams, drafting in Word | Sits inside the apps people already have open; answers are bounded by file permissions |
| ChatGPT, started on personal logins, later a Team plan for one department | Mixed consumer and business | Drafting, rewriting, analysis of pasted material | Half the team may be on a personal account under different terms |
| Claude, adopted by two or three heavy users | Often personal Pro | Long documents, structured analysis, technical work | The in-house experts sit here, and they are your co-facilitators |
| Gemini, arrived with Google Workspace | Business, assigned by IT | Docs, Gmail, meeting notes | Same bundling story as Copilot, different suite |
Read that as a market condition, not a governance failure. Menlo Ventures’ 2025 enterprise survey of 495 US enterprise AI decision makers put Anthropic at 40% of enterprise LLM API market share, OpenAI at 27% and Google at 21%, which is 88% across those three. Multi-vendor is how the category works. A company that picked one assistant in 2024 and never revisited it is the unusual case, not the tidy one.
The shared layer is most of the session
Strip the assistants down to what a person does with one and the overlap is heavy. Our case for teaching transferable skill works through the arithmetic; the short version is that roughly four fifths of a competent first session is tool-agnostic, which is why a mixed room is far less of a problem than the seating chart suggests.
What the whole room learns together, on their own documents, in any window:
- Framing the task instead of asking a question. Name the output, the reader, the length and the format.
- Giving the model the source material. The biggest quality jump comes from attaching the real document instead of asking the tool to recall facts it was never given.
- Asking for structure. A table with named columns, a list with a fixed count, a draft with the sections you need.
- Iterating two or three turns. Most people stop at the first answer. Correcting once is where the useful version appears.
- Verifying before it leaves the building. What to check, where these tools invent detail, and which numbers get traced back to the export.
- Knowing when to close the window. The tasks where the tool costs more time than it saves.
None of that changes when the person is in Copilot instead of Claude. Teaching it once to a mixed room is cheaper than teaching it three times, and it lands harder because people watch colleagues reach the same result by three routes.
One more thing belongs in the shared hour. Three assistants means a person re-decides which window to open before every task. Harvard Business Review’s 2022 study of application switching tracked 137 people across three large firms and found around 1,200 toggles a day, close to 9% of annual working time spent reorienting. That research measured all app switching, not AI tools, so treat it as the mechanism and not the measurement. The fix is cheap either way: have the room agree a default tool per recurring task, so the choice gets made once instead of daily.
The stretch that has to split by tool
Budget 45 to 60 minutes at the end for tool-specific work. This is the part a single-tool agenda cannot cover and the part most vendors quietly drop.
- Where the file goes. Attaching to a chat, dropping into a project or workspace, or working on a document already in the suite. Each has a different ceiling on how much material it holds.
- What the tool can see unprompted. Copilot reaches into mail, files and calendar through existing permissions. A chat window sees only what the person pastes. That gap changes the prompt.
- Where saved work lives. Saved prompts, projects, custom instructions and chat history differ in every tool, and this mechanic decides whether anything survives to the following week.
- Which admin settings are on. History, retention and sharing are set at the tenant or workspace level, and half the room does not know their own setting.
Run it as breakouts by tool, each group on its own documents. If the team has not yet named a primary tool, this stretch produces the answer, and our guide on which assistant a session should teach covers how to weigh the call.
Three tools means three data boundaries
The real cost of a mixed stack is not skill. It is that nobody can state the data rule for all three tools, so people apply the strictest rumor they heard, or none at all.
The boundaries are published, and the vendors’ own words are the ones to teach. Microsoft’s documentation for Copilot states that “prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs”. It also states that Copilot “only surfaces organizational data to which individual users have at least view permissions”, which makes your SharePoint permission model part of your AI surface. Anthropic’s privacy documentation says it “will not use your inputs or outputs from our commercial products to train our models” by default, with submitted feedback as the exception. Business tiers across these vendors converge on that same default. Consumer accounts do not, which is why the bring-your-own-tool finding above is a curriculum item and not a footnote.
So a mixed room needs one page, not three policies: for each tool in the building, which classes of document may go in, which never do, and what the account tier is. Teach it inside the task people just practiced. As an opening slide it reads as compliance theater and nobody retains it. Taught at the moment someone is about to paste a payroll file, it holds.
This is where a mixed stack beats a standardized one. A team that has handled the same document in two tools reads the boundary as a property of the tool and not a rule from legal.
How the room is set up
We deliver these sessions, so this part is observation and not theory.
Seat by tool for the practice stretches. People troubleshoot a neighbor’s screen faster than a facilitator crosses the room. Mixed tables look collaborative and slow the hands-on work down.
Cap it at 15 per facilitator, and treat three tool groups as the ceiling. A fourth group fragments the room past what one person can correct. Beyond 20 people in a mixed room, the second facilitator is not a luxury.
Demo once, on the most-licensed tool, then everyone works in their own. The demo shows the shape of the task, not the buttons. Demonstrate every exercise in all three tools and the session becomes a product tour with no practice time left. The wider mechanics of building that agenda sit in our workshop design guide.
Insist on a facilitator who works in all three. This is the question that separates vendors on a mixed-stack brief. Ask what they shipped last month and in which tool. Someone fluent in one assistant and briefed on the other two will teach your Copilot group from a help page.
Recruit your heavy users beforehand. Give each a named exercise to lead and the session gains a facilitator. Leave them unbriefed and they answer every question from the back row.
Book the follow-up before the first session runs. Ninety minutes, same group, two to three weeks out. In a mixed room this is where tool choice gets decided on evidence, because by then people know which window they reach for. The case for a short session plus a clinic is in our review of formats for non-technical teams.
What the mixed stack adds to the price
Less than most buyers expect. Market rates for a private session of hands-on AI training run about $2,000 to $15,000 depending on how many roles need their own material. A mixed stack adds prep time and, past roughly 20 people, a second facilitator. It does not add a second curriculum, because the shared layer is the curriculum.
Here is what the corporate AI workshop catalog publishes, checked this month.
| Provider | Published format | Published price | Covers a mixed stack? |
|---|---|---|---|
| Certstaffix | Half-day business overview and four one-day live online titles | $350 per person per title | Single-tool live titles |
| Certstaffix AI Prompting bundle | 17 self-paced courses | $375 per person | Yes, names ChatGPT, Gemini and Copilot |
| Section | 1.5 hours of live instruction plus recordings | $195 list, or membership at $62.50 per month billed annually | Promises the major models, names none |
| Private group session, market rate | 3 to 3.5 hours plus a clinic | $2,000 to $15,000 | Built around whatever is in the building |
One item in that list addresses a mixed stack directly, and it is self-paced. Nobody sells the mixed room, which is why the brief has to be written instead of ordered.
The decision pack the session should hand you
Treat the workshop as the cheapest input to the consolidation decision, and ask for four things out of it:
- A census by task, not by seat. Which tool people reached for, for which recurring job, when nobody was enforcing anything.
- The one-page data rule, per tool and per document class, including which accounts are personal.
- The committed tasks. Each person names one real job they will do this way before Friday, with two saved prompts in a shared place.
- The honest read on consolidation. After three weeks of real use, one of the three is usually doing nothing. That is a license you can cut with evidence instead of a theory.
Then handle the part that decides whether any of it reaches the numbers. BCG’s 2026 AI at Work survey covered nearly 12,000 employees, managers and leaders. It found 36% felt adequately upskilled while 72% reported that skill expectations had shifted, and 66% still got limited or no guidance on what to do with the time AI saves them. Unassigned hours go straight back into the inbox. Naming the use of recovered time is a ten-minute agenda item and the difference between a well-reviewed session and AI enablement that moves a number.
Frequently asked questions
Should we standardize on one AI tool before running a workshop?
No. Run the session first. Most of what it teaches works the same in every assistant, and three weeks of corrected practice tells you which tool is doing real work better than a license report does. Consolidate afterward, with evidence, and cut the seat nobody reached for.
Can one workshop cover a team using ChatGPT, Copilot and Claude?
Yes, and it is the normal brief for an AI workshop for teams now. Teach the shared layer to the full room on their own documents, then break out by tool for the last 45 to 60 minutes to cover file handling, saved work and admin settings. Keep it to three groups; a fourth fragments the room past what one facilitator can correct.
How much does an AI workshop for a mixed tool stack cost?
Private group sessions run about $2,000 to $15,000 at market rates. A mixed stack adds prep and, past roughly 20 people, a second facilitator, not a second curriculum. Per-seat catalog options are cheaper and published: $350 a head for a one-day live class, $195 for 90 minutes of live instruction, $375 for a self-paced bundle.
Does teaching several tools confuse people?
Not when the session is built around the task instead of the product. Confusion comes from product tours shown in three windows. When each person works in the tool they will open on Monday, the mixed room is an advantage, because people see the skill is not the software.
What if half the team is on personal accounts?
Treat it as the finding it is. Microsoft’s 2024 survey put 78% of workplace AI users on their own tools, 80% at smaller firms. Teach the account-tier gap explicitly, because default data handling differs between consumer and business plans, then use the session’s census to size the licenses you need.
Who should facilitate a session for a mixed stack?
Someone who works in all three tools daily, not a trainer briefed on two of them. Ask the vendor what they shipped last month and in which tool. A facilitator teaching from a help page cannot fix a Copilot permission problem while a Claude group waits.
Do these tools train on our documents?
On business tiers, the published default is no. Microsoft’s documentation states that prompts, responses and Graph data are not used to train foundation models, and Anthropic’s says commercial inputs and outputs are not used for training by default, with submitted feedback as the exception. Consumer plans differ by vendor and by setting, which is why account tier belongs in the curriculum.
How many people can be in a mixed-tool session?
Fifteen per facilitator, and 12 is better. The limit is correction time and not seating: the facilitator switches context between tool groups, so the ceiling sits lower than for a single-tool session of the same length.
Should the heavy users attend?
Yes, with a job. The two or three people already fluent are the fastest route to the rest of the room. Brief them beforehand and give each a named exercise to lead.
What should we send the vendor before the day?
The tool and tier for every attendee, three or four real documents redacted as needed, and the recurring task you want finished faster. A vendor who never asks for the tier list will teach one tool and hope the rest translates.
Key takeaways
- Do not hold AI enablement behind a consolidation decision. Train the mixed room now, then cut licenses using what the session showed you.
- Roughly four fifths of a competent first session is the same in every assistant. Teach that to the full room on their own documents and split only the last 45 to 60 minutes by tool.
- Part of your stack is invisible to procurement. Microsoft’s 2024 survey found 78% of workplace AI users on tools they brought themselves, 80% at smaller firms, so account tiers belong in the curriculum.
- The mixed stack’s real gap is the data boundary, not the skill. Publish one page covering each tool, each document class and each account tier, and teach it inside a live task.
- Multi-vendor is the market norm. Menlo Ventures put 88% of enterprise model spend across Anthropic, OpenAI and Google, so three assistants is a condition to train through, not a mess to clean up first.
List every tool and tier in the building, name the weekly task you want finished faster, and brief a session against both. A free AI-readiness call will tell you which stretch of the agenda splits by tool and which license the follow-up will probably let you drop.
Arthur Wandzel
