Every AI session for a finance team gets judged on one question, and nobody asks it in the room. It gets asked the following Friday, when an analyst opens the variance file and decides whether to work the way she worked last month. Plenty of sessions are pleasant and change nothing by that Friday. The difference is rarely the trainer’s slides. It is how the hours were divided, what the team was allowed to put into a model, and whether anyone owned the two weeks after.
Finance leaders buy these sessions on a thin description. Event pages publish a time grid with no content: registration at 2:30, program at 3:00, happy hour at 4:30. Curriculum pages publish content with no time, like “generate your AI use-case register.” Here is the other version: a half-day with a finance team, block by block, including the part that fails.
What the session is for
Nobody books this because the team lacks tools. The licenses are already paid for, which is the whole problem. Gartner’s 2025 survey of 183 CFOs and senior finance leaders found 59% reporting AI use in the finance function, against 58% a year earlier and 37% in 2023. The curve went flat. Deloitte’s fourth-quarter 2025 CFO Signals survey of 200 large-company finance chiefs found 87% putting AI in the top two importance tiers for finance operations in 2026. Expectation is near-universal; usage stalled. A seat license closes that gap about as well as a gym membership closes a fitness gap, the argument behind our case that buying licenses is not the same as building capability.
So the session has one job: make four or five recurring finance deliverables faster to produce at the same review standard. A named deliverable, a named person, a measured change in cycle time.
Who is in the room
Eight to fourteen people, about the whole finance function at a company of 50 to 500:
- The controller, who owns the close calendar and the follow-through. If the controller is not in the room, book a different date.
- Two or three FP&A analysts, who write the most per month and get the fastest wins.
- Whoever runs AP and AR, whose work is high-volume and pattern-heavy.
- The tax or compliance lead for an hour, to say out loud what cannot leave the building.
- The CFO for the first thirty minutes and the last twenty. Longer and the junior people stop experimenting.
One room works for a first session even though the roles differ, because the shared skill is reading a model’s output with a finance eye. The split comes later: a second session for FP&A on forecast narrative is a different animal from one for AP on exception handling, which is the subject of our playbook on why one company needs nine different curricula.
Keep vendors and software demos out. The AI in Finance Lab, a practitioner series for controllers and FP&A directors, excludes vendors and consultants from attending, and it is right to. A demo changes what the room talks about, and a finance team gets only a few hours of its own attention.
A half-day, block by block
Minutes matter more than topics, so block length is the design.
| Block | Minutes | What happens | Why that long |
|---|---|---|---|
| Access check | 15 | Everyone signs in to the approved account, own laptop, work network | Cheap first, fatal at minute 90 |
| Where it fails | 30 | The team asks a model to foot a schedule and finds the wrong total | Trust calibrates early or never |
| Variance commentary | 50 | Each analyst drafts last month’s commentary from their own file, then edits | Finance’s highest-volume writing task |
| Break | 15 | Coffee, and the first side conversation about a task we did not plan for | Where real use cases surface |
| Exception chasing | 40 | Reconciliation triage: group open items, draft chase emails, rank by size | Volume work earns its keep |
| The narrative | 35 | One board page and one business-review paragraph, two audiences | Audience control transfers |
| Rules of use | 25 | The data line and the review rule, written down and agreed | Or a memo nobody reads later |
| One task each | 20 | Each person names one task they will do this way next month, in writing | The only part that survives |
Three hours and fifty minutes of program, four and a half hours with arrivals and overrun. Shorter formats teach prompting, not capability: Schneider Downs runs a 90-minute session on variance analysis, reconciliations, reporting narratives, and close support, a fine use of 90 minutes that never reaches the last block.
The exercise that fails, on purpose
Hand a finance team a model and watch what they do first. They ask it to add something up. Foot this schedule. Recalculate this variance. Tie out this balance. It is the most natural request in the building and the one use where a language model is least dependable, for a reason the room needs to see rather than hear: the model is predicting plausible text, and a wrong total is formatted exactly like a right one. No hedge, no red cell. In a function built on numbers that tie, that is the most expensive habit a team could pick up on its own.
So we trigger it on purpose, in the first half hour, on a schedule where we already know the answer. Someone catches the error. The room goes quiet for a second. Then we draw the line that holds for the rest of the day: the spreadsheet computes, the model writes, restructures, classifies, and chases.
The data agrees. In the same Gartner survey, the most common finance use cases were knowledge management at 49%, accounts payable process automation at 37%, and error and anomaly detection at 34%. Retrieval, routing, pattern-spotting. Not arithmetic. Teams already getting value had learned in private what this block teaches in public.
The four task families that hold up
After the failure block, the day runs on four families: the ones that still look good a month later, once the novelty is gone.
| Task family | What the model does | What stays with the analyst | Control that makes it safe |
|---|---|---|---|
| Variance commentary | Drafts the explanation from the variance table and last month’s text | The causal claim: why the number moved | Every driver traced to a source line |
| Reconciliation triage | Groups open items, drafts chase emails, ranks by size and age | Deciding what is a real break | Items stay in the system of record |
| Reporting narrative | Rewrites one set of facts for the board, the business unit, the all-hands | What gets emphasized, what gets left out | Facts pasted in, never recalled |
| Policy and process text | Drafts close checklists and memos, answers “what does our policy say” | Judgment and approval | The policy is pasted in and cited |
Two hours on four families beats a tour of twenty features. Prompts written by a trainer go stale by the next close; prompts an analyst wrote while swearing at her own variance file do not.
One scope note: this is a fluency session, not a software build. Nobody leaves with an agent wired into the ERP; they leave able to use ChatGPT, Claude, Gemini, or Microsoft Copilot well on work they already do. Anything built is a project, scoped as one.
Settle the data line before the invite goes out
A finance team holds material non-public information for part of every month, so the access question is a control question, and it gets answered before anyone is invited rather than negotiated in the room.
One widely attended hands-on finance lab tells registrants to bring a personal email as a backup if their company restricts certain tools. For a controller sitting on unreleased results, that is an instruction to move company data outside the control environment, and it should rule the format out. Three questions, answered in writing, before the date is confirmed:
- Which account? An enterprise or business tier whose terms exclude training on your data, administered by your own IT. Not a personal login.
- Which data? Our default is post-close actuals from a period already reported, plus policy and process documents. Pre-release results, compensation detail, customer contracts, and anything under a non-disclosure agreement stay out until a written rule lets them in.
- Who decides? One named person, usually the controller or the CFO, who can approve an exception during the session without a committee.
The third question is the one most companies have not answered. Deloitte’s second-quarter 2026 CFO Signals survey, fielded May 22 to June 7 with 200 North American finance chiefs, found 59% calling the balance between fast AI adoption and risk their top governance challenge and 43% citing limited insight into which AI tools are in use. A team that cannot say which tools are in use already has quiet usage, and a session that ends with a written data line is often the first time the rule got written down. The bar is higher again under a regulator’s eye, which we cover in our guide to AI training for advisory firms.
Where output has to stop
The review rule fits on a sticky note, and it is what lets a controller sign off on the whole experiment: no model output enters a reviewed deliverable until a human has traced every number in it to a source. Three habits carry it, all taught in the rules block:
- Paste, do not ask. Facts come into the conversation from your file. A figure the model produces from memory is a guess wearing a suit.
- Name the source line. Commentary that says “driven by the timing of the Q3 media spend” has to point at the line that shows it. The analyst does that, not the model.
- Keep the prompt. If a drafted paragraph reaches a board deck, the prompt and pasted input belong in the working file. Auditors ask how a number was produced, and “the model wrote it” is not an answer.
Book it around the close
Finance has a monthly shape and most training vendors ignore it. Days one through eight belong to the close, and a session booked in that window leaves half the room answering messages, which is worse than not attending. Use the second half of the month. For teams with a busy season the rule is sharper still: train in the trough, leave the peak alone, as we set out in our piece on what to cover before busy season.
The four weeks after
Programs decay. A team gets a good session, goes back to the close, and reverts inside three weeks. The fix is boring and costs the controller about an hour a month.
Week one. Each person does their one named task the new way, once. Not a program. One task.
Week two. Fifteen minutes at the front of the existing team meeting. Two people show what they did, including what did not work. The failures are the useful part.
Week three. The controller collects the surviving prompts into one shared file. Four or five good ones beat a course library nobody opens.
Week four. Measure cycle time on one named deliverable, before and after, in hours. If the variance pack took eleven hours in September and seven in November, you have your budget case.
Finance leaders say this is the part that works. The AICPA and CIMA surveyed 1,446 senior finance leaders and managers in August and September 2025 and found 61% ranking on-the-job training the most effective approach to technology upskilling, with 56% naming generative AI their most prominent skills gap. The session starts the work; it is not the work.
What a session cannot fix
It cannot fix bad data. If the chart of accounts is a museum and three systems disagree on revenue, a fluent team will write better descriptions of a mess. Data literacy and data quality were the two largest obstacles in the Gartner survey, and training is not the lever for either. It cannot fix a broken process either: a close that takes twelve days because four approvals sit with one person will still take twelve days.
It cannot replace a credential. The AICPA and CIMA have launched an AI Accelerator Skills Program in three tiers, costing roughly $495 to $1,000 per person and piloted at KPMG, Cisco, and Stanley Black & Decker. Good buy for what it is, and a different purchase from this one; the two are compared in our breakdown of the options for finance and accounting teams.
And it cannot convert someone who was volunteered. The best outcomes we have seen came from skeptical controllers who tested the thing hard in the failure block. The worst came from people sent by a manager who did not attend.
Frequently asked questions
How long should an AI workshop for a finance team run?
About four hours of program, which lands as a half-day. Under two hours you can teach prompting but not reach a committed task per person. Full days lose the room after lunch.
Should the whole finance team attend one session or should we split by role?
One room for the first session, split for the second. The shared skill in session one is reading model output with a finance eye, which is the same for AP and FP&A. The work diverges after that, and the deeper gains sit in role-specific sessions.
Can the team use real financial data in the session?
Yes, with a rule set in advance: post-close actuals from a period already reported, plus policy and process documents, inside an enterprise account your IT administers. Pre-release results and anything under a non-disclosure agreement stay out unless a named approver says otherwise.
How much does a session like this cost?
Live team sessions in the US market generally run $2,000 to $15,000 depending on length, tailoring, and how much pre-work the provider does on your files. Credential programs sit far below that and executive programs well above. Get the scope in writing: who runs the room, how many attendees, whether your own documents are used.
Do people need prior AI experience?
No, and mixed rooms work better than uniform ones. The skeptics ask the questions that calibrate everyone else, and the two people already using these tools privately become the champions by the afternoon.
Does this count for CPE?
Only if the provider is registered to issue it or partners with a firm that is. A fluency session from an engineering firm usually is not, so if CPE hours are the goal, buy them from an accredited program separately.
What does our controller actually have to do afterward?
Fifteen minutes at the front of an existing team meeting in week two, collect the surviving prompts in week three, measure cycle time in week four. About an hour over a month, and it is the difference between a good afternoon and a changed close.
How do we know whether it worked?
Pick one deliverable before the session and time it: hours to produce the variance pack, days to close, items cleared per hour in the reconciliation queue. Compare one cycle before with one cycle after. Usage dashboards tell you people logged in, which is a different question.
Key takeaways
- The session is judged by what changes in the next close. Design every block backward from one named deliverable per person.
- Trigger the arithmetic failure early and on purpose. The spreadsheet computes; the model writes, restructures, classifies, and chases.
- Four task families carry a first session: variance commentary, reconciliation triage, reporting narrative, and policy or process text.
- Settle the data line in writing before the invite goes out: which account, which data, who approves an exception.
- No model output reaches a reviewed deliverable until a human traces every number to a source.
- The four weeks after decide whether it survives: one task each, a show-and-tell, a shared prompt file, one measured cycle time. And book it after the close, never during.
Weighing a session for your own finance team? We offer a free AI-readiness call. Bring your close calendar and one deliverable you would like back an afternoon of.
Arthur Wandzel