What a three-hour AI crash course looks like

Most corporate AI training fails the same way: ninety minutes of slides about what AI is, a demo on toy data, and a feedback form. Everyone leaves mildly interested and nothing changes. The version that works is almost the opposite — minimal theory, maximum reps, and every exercise built on the team's own documents, models and inboxes. That last point isn't a nice-to-have. It's the whole design.

The format below is one we've run repeatedly with investment teams. It assumes a room of smart, busy, sceptical professionals — the kind who'll give you three hours once, and never again if you waste them. Worth noting the stakes: 75% of UK financial firms now use AI^[1], but capability across teams is wildly uneven, and unofficial usage runs far ahead of official training^[2]. The training gap is the adoption gap.

01The shape of the three hours

02Part 1 — orientation, not theory

The first job is configuration, done together, live: custom instructions so the AI knows who you are and how you like output; memory switched on and reviewed; connectors to the document store (respecting existing permissions — the AI only sees what you can already see); a project set up per deal or workstream so context persists. Ten minutes of setup makes every subsequent hour more useful — a default install and a configured one are different products.

Then the mental model. We keep it to a six-and-three: six strengths to use deliberately, three watch points with a fix for each.

The six strengths — first drafts, research and synthesis, format translation, ideation and counter-argument, data analysis, structured output — get demonstrated rather than described, each on a task from the room. The counter-argument one always lands hardest with investment teams: nobody at committee is incentivised to push hard against a thesis the room already likes, but a model will play the sceptical IC member all day without ego.

03The prompt clinic: same question, three versions

The centrepiece of Part 1 is watching one question improve in front of you. We use IC prep because everyone in the room has lived it:

V3 works because it's scoped to real materials, demands structure, forces honest gap analysis, and anticipates the follow-up. The meta-lesson matters more than the example: the quality of the output is set by the quality of the ask, and the fastest way to improve an ask is to have the AI improve it for you. Anthropic's own prompt-engineering guidance runs on exactly these principles — role, context, constraints, examples, explicit permission to say "I don't know"^[3].

04Part 2 — an hour inside the Microsoft suite

This is the block that converts sceptics, because it happens inside the four applications where the team already spends its life:

• Excel. We take a real financial model, plant first-year-associate mistakes in it, and run a model-audit workflow: every formula error, broken link and off-benchmark assumption flagged, commented in-cell, and summarised in an issues sheet — cell reference, severity, recommended action. Then a two-variable sensitivity table on command. The room goes quiet at this one.
• PowerPoint. A deck with deliberately introduced formatting and data inconsistencies gets a "onceover" review, then an exec-summary section generated from the deck's own content, in house style, flagging any section where source data is missing.
• Word. An NDA redlined against the firm's standard position — every material deviation flagged, explained, with counter-language proposed as tracked changes plus a summary table.
• Outlook. Thread triage (decisions made, open questions, actions by owner with deadlines, what needs your reply) and drafting in your own tone from thread context.

Each exercise ends the same way: save it as a skill. The audit prompt, the exec-summary prompt, the redline prompt — iterated once until right, then saved as a one-click template the whole team can run. Repeatable prompts are how a team's IP compounds instead of evaporating^[4]. One rule travels with every exercise: prove a workflow on a case where you already know the answer before you trust it on one where you don't.

Rule of thumb we teach on modes: thinking or creating → chat. Processing or automating → the agentic desktop. And every factual output is a first draft to verify, not a fact to forward.

05Part 3 and the close

The final block raises the ceiling: dedicated projects that hold a deal's entire document set — so a CIM can be interrogated with page-cited answers, committee prep can be stress-tested against the actual materials, and any team member can get up to speed without a briefing call. Then multi-document extraction: CIM, model and legal docs in one place, one consistent output format, every data point cited back to its source, "Not Found" when it isn't there. We also point at the growing open libraries of pre-built skills for financial services, because starting from a tuned template beats starting from a blank prompt^[4].

The close is deliberately practical — a cheat sheet of habits that compound (screenshots as context; voice input instead of typing; meta-prompting — "take this prompt and improve it"; fresh chats per phase; prompts saved to shared projects; explicit guardrails on every research prompt) — and then a commitment round. Everyone leaves with three things written down:

1. One thing you'll do differently this week.
2. One use case you'll try.
3. One prompt you'll take from today.

For attendees who want the deeper technical grounding afterwards, we send exactly one piece of homework — an hour, and better than most paid courses^[5]:

06Why three hours, and why this order

Because attention is finite and behaviour change is the goal. Orientation first, because a configured tool feels different from a default one. The Microsoft suite second, because meeting people inside their existing tools removes the biggest adoption excuse before it's voiced. Higher-value use cases third, briefly, as a glimpse of the ceiling. And the commitment round last, because training that ends without a named next action is entertainment. A single session doesn't make a team expert — it makes a team started, with saved skills as proof. What keeps them moving is a place where all of it lives and channels that keep the momentum visible — both covered in their own posts.

If your team needs the started part — get in touch.