All our training is closed and in-house, for your people only. Beyond that there are two varieties. One is the programme we have taught before. The other we design around your organisation.
In the simpler way, you know what you need, we’ve taught it before, and we come and deliver it. The session is hands-on, runs on your premises, and the exercises come from your sector.
The more interesting way starts from a theme. Perhaps your finance team should be using AI properly, or your engineers keep shipping prototypes that fail in production. Before designing anything, we interview the people who will attend, because what actually hurts is rarely what the training brief says. If your policies allow us near your data, we build the exercises on it. People leave having done the thing on their own material.
We prefer this second route even when you’re buying the standard programme. An hour beforehand with the people commissioning it, covering what the team does today and which decisions the training should change, lets us swap the stock exercises for ones that are actually yours.
Sessions are closed so that people can be honest about what they don’t know, and groups are kept small for the same reason. Hands-on workshops are capped at twelve; Decisions in the Age of AI takes up to sixteen, since the day runs on discussion. The learning outcomes are tied to the participants’ actual jobs. Sessions run in whatever AI tools your organisation already uses, and what people learn carries over when those tools change. A month after the session we follow up, because the measure of a workshop is what people are still doing thirty days later. Three of the forty learners in the case where we tested that measurement looked like the strongest improvers in the room on the day and had kept 21.9% of their gain eight weeks later.
These are the themes we keep coming back to. Treat them as starting points; every session is adapted to its audience.1
For boards and function heads · half a day
The session opens with current systems running on the screen, showing what they can and cannot do. Most of the rest covers reading a vendor’s claim, what a proof of concept is evidence of, and where responsibility sits when a model gets something wrong. The group then works on the decisions it’s actually facing, such as a budget line, a build-or-buy decision, or a supplier already in the building. People leave with a shorter list of things AI might do for them, and better reasons for the items still on it.
For finance, operations, sales and HR · one day
Everyone brings a weekly task and rebuilds it with AI: a recurring report, a queue of replies, a pile of documents that somebody has to read. Drafting, summarising, extraction and checking each get a turn, along with the plainer business of writing an instruction that returns the same answer twice. A good part of the day covers verification: which outputs need checking, and how to build that check into the workflow. By the end the task is rebuilt and running.
For controllers and FP&A teams · one day
Forecasting, variance analysis, the close and management reporting make up the material, worked on data shaped like a real ledger. The session covers where a model earns its place, where a deterministic rule is better and cheaper, and how to keep an audit trail through either, because a number that reaches the board has to be explainable. The worked examples come from Eva’s eight years in financial planning at Procter & Gamble and Coty. Teams who can bring an anonymised extract of their own ledger work on that, and the target is one step of a real close automated by the end.
For engineering teams · two days
Most of the time goes to evaluation: building a test set, measuring against it, and noticing when a change has made the system worse. Retrieval and agent design are taught as the things being measured, which puts the harness before the features and keeps the argument about numbers. Teams with a system of their own bring it and build the harness around that; the others get a prepared system with faults planted in it. Two days is enough to leave with a running evaluation suite and a numbered list of what is failing inside it.
For senior data scientists and quants · one day
This session teaches the mathematics directly, starting with what a model’s confidence number actually measures and whether it survives a test. Bayesian treatment of model uncertainty follows, then proper scoring rules, then the decision layer, which covers when to act on an output, when to route it to a person, and what each threshold costs. Participants calibrate a model of their own and measure the result. The material comes out of the working paper the two founders are writing on calibrating uncertainty in language-model systems. Uncertainty quantification is Luigi’s research field; the paper is joint work, and so is the session.
For anyone who has to read a vendor proposal · one day
A proposal, a benchmark table and a price list go up on the screen and get taken apart. We work through what the demonstration is evidence of, what the numbers were measured on, what the thing costs to run a year in once volumes are real, and which questions turn a claim into something testable. Contract terms get a look as well, particularly data handling, exit and who owns what gets built. Anyone with a live proposal can bring it, and if they do, the last stretch goes on drafting the questions to send back.
The EU AI Act is the first general law on artificial intelligence, and it reaches further than most companies assume, because it regulates not only those who build AI systems but those who use them. Article 4, in force since February 2025, requires every organisation whose staff use AI at work to ensure a sufficient level of AI literacy, measured against their roles and the systems in use. The general obligations began applying in August 2026, and the rules for high-risk uses phase in through 2027.
Two consequences matter for most of our clients. First, if your teams use AI at all, even just a model drafting text or an assistant inside a workflow, the literacy duty already applies to you. The European Commission’s own guidance confirms that ordinary commercial use makes a company a “deployer”. No certificate exists, no fixed curriculum is prescribed, and an internal record of relevant training is sufficient evidence. Second, some uses carry far heavier obligations. Systems that influence employment decisions, such as screening candidates or evaluating staff, are classed as high-risk, with their own requirements arriving through 2027.
Our programmes satisfy the Article 4 duty, and we provide the session records that evidence it. Where a question crosses from practice into legal interpretation, we say so and point you to counsel.
Both cases below are invented. They show what tailoring means in practice.
A CFO asks for a workshop on AI-assisted forecasting. In the interviews beforehand, the controllers say the forecast itself is fine; the real pain is the week of manual reconciliation that precedes it. The workshop that eventually runs teaches the team to automate their own close, in their own tools, on their own ledger.
Exemplum · hypothetical
An engineering team keeps shipping LLM prototypes that impress in the demo and fail in production. They bring the current one to the two-day evaluation workshop and spend it building an evaluation harness around their own system. They leave with a measured list of what is actually failing, and a way of testing the next claim before believing it.
Exemplum · hypothetical
A public course teaches a fixed syllabus to whoever signs up, and nobody in the room can talk about their own figures. These sessions are closed and designed after we’ve spoken to the people attending, with the exercises built on your material where your policies allow. In the workshop formats people rebuild a task of their own.
They show the method and the marking rather than results for a client. Nine of the ten are built on data we engineered ourselves with the answer key held out, and one on a public contract dataset with a published expert key, so every figure could be scored against something instead of asserted. One of them tests the instrument we measure training with, on a simulated cohort whose learning was planted in advance. The ten cases carry their negative results as prominently as the positive ones.
Two to three weeks for a standard programme, which still includes the scoping conversation beforehand, and four to six weeks if the sessions are built around your own material. Groups are capped at twelve for the hands-on workshops and sixteen for the Decisions in the Age of AI session. If a particular date matters, say so in the first email and we’ll tell you whether it’s possible.
For the Article 4 literacy duty, yes. We provide a session record of who attended, what was covered and when, which is the evidence the duty asks for, and where a client wants more than an attendance record we attach the pre/post measurement, which evidences the duty with what people can do afterwards. There is no certificate to award, because the Act prescribes none; section iv sets out what it does require and where the heavier high-risk rules begin.
If you know what you need, say so and we’ll tell you whether the standard version fits. If all you have is a theme, send that.