Why most corporate AI training doesn't work (and why it's riskier in regulated firms)

More training won’t fix this.

You’ve rolled out Copilot, run lunch-and-learns, and circulated the prompt cheat-sheet, but twelve months later, where is the productivity improvement?

Many organisations now provide some form of AI training, but staff adoption is still poor, AI outputs contain errors, tasks seem to be taking just as long as before, and people are still using their personal ChatGPT accounts rather than the firm’s approved solution.

The problem is that the standard offering breaks every link between training and a business result.

And in a regulated firm, ineffective training isn't just a waste of time and money. It can give staff the confidence to use the tools without equipping them to catch errors, increasing risk rather than capability.

Employees themselves are often sceptical of what is provided, with BCG finding that only 36% of staff considered their training adequate.¹

The training often consists of a rundown of tools, frameworks, and use-case examples. Delivered as a workshop and some self-paced modules, this is a discrete, procurable deliverable that matches the standard idea of what training should look like.

But the return on this type of training is notoriously poor², and the situation is particularly bad in the case of AI.

The lack of impact is often blamed on factors like culture and change management. These are important, but the training also fails on its own terms. For training to really work, it must:

• teach the right things
• in a way that makes them stick
• to people with a reason to change
• into a workflow where the change can land
• with the judgement to know when the tool is wrong.

These are a mixture of broader organisational factors and the design of the training itself, but the standard offering breaks every link.

It teaches the wrong things: prompt engineering rather than judgement

Effective AI use depends on questions like when to use it, when not to, how to spot mistakes and hallucinations, and when to stop iterating. The standard curriculum contains almost none of this.

Instead, training tends to focus on easily packageable techniques like best practices for prompt engineering, even though many of these ideas are outdated, poorly evidenced, or have inconsistent results.³

The curriculum often includes techniques like assigning the LLM a role and using rigid prompt templates, even though the current world of reasoning models, tool use, and agentic workflows makes most of this out of date. Durable skills of evaluation and judgement are skipped in favour of easier-to-teach ideas that are now unnecessary or even counterproductive.

A user who can structure a prompt but can’t spot a hallucinated citation or a mistake in a document summary is missing the skills that matter most.

Even if it taught the right things, they wouldn't stick: the format ignores how people learn

The delivery format itself, whether a workshop, self-paced modules, or a combination of the two, is a near-perfect inversion of what the learning science says makes anything stick.

The most replicated findings in cognitive psychology are a checklist of what the format omits: massed delivery instead of spacing, passive exposure rather than active engagement with the material, and no opportunity to get feedback on the learner’s own errors.⁴

A fluent, well-produced session maximises the feeling of learning while doing little for actual learning. If you’ve ever tried to quiz people a few months later, most will have forgotten much of the content unless they’ve made a concerted effort to apply the ideas in practice in the meantime.

Expertise is built by accumulating exposure to real cases with feedback, building a mental model that allows you to recognise when something is off⁵. A workshop can be a starting point, but it cannot replace this process by itself.

Even if it stuck, it gives nobody a reason to change

Training assumes a motivated learner but does nothing to create one. Support from above and motivation to apply what has been learned are among the strongest predictors of whether training transfers to business impact.⁶

The good news is that the raw motivation to use AI may already exist. BCG found that 54% of staff would use AI tools even if they weren’t authorised⁷. But if sanctioned use is slower and more constrained then a workshop does nothing to close that gap.

Motivation is mostly built outside the training itself. This could be through factors like senior leaders visibly using the tools (leaders at the highest-performing firms are almost twice as likely to model AI use⁸), protected time to practise on real work, and incentives that reward experimentation rather than punishing it. If staff fear looking incompetent or worry that they’re just training their own replacement, they’re unlikely to engage in the experimentation that results in true learning.

Even with a motivated user, it lands in an unchanged workflow

Most training teaches people to use AI in their existing workflows, but the greatest productivity gains come from redesigning the workflows themselves.

In the research cited above, BCG found that the companies creating the most value from AI concentrate 80% of their investment on redesigning end-to-end workflows and inventing new opportunities rather than just deploying tools.

If the overall process is unchanged, using AI to assist with a couple of steps is unlikely to have much impact on the overall timeframe or workload. Speeding up the first draft doesn’t help if you just spend longer waiting for the review.

And even past all of that, it builds confidence faster than judgement, creating risk rather than efficiency

Training often focuses on how you can use AI, with less time spent on when and how not to. Even if the tone isn’t deliberately promotional, the illusion of learning discussed earlier means that confidence rises in the workshop even when skill doesn’t, and a curriculum built from successes only points that confidence in one direction.

The most dangerous combination is confidence without calibration, which in a trust company or law firm can result in a misstatement to a client or a hallucinated citation in a court submission.

This is why much of the best training already available in the market includes explicit examples of AI weak points or cases where it should not be used, though it is still too common for this to be skipped.

What training that works looks like: fixing each broken link

An evidence-based programme looks different to the dominant offering at basically every stage.

The first step is to determine what problem you are trying to solve, and whether training is even the right solution.

Where it does form part of the solution, the best training repairs each link in the chain:

• It focuses on the skills that matter, like critical output evaluation in the participant's own domain.
• It does it in a way that sticks: error-spotting drills on real material, decision-forcing cases from the participant's own practice area, calibration loops (estimate the AI's accuracy, then check), and practice distributed over time rather than in a single workshop.
• Rather than assuming motivation, it is used alongside other interventions such as visible leadership modelling of AI use.
• It lands in a workflow that has been rethought to get the best value from AI. This workflow embeds governance and output validation directly, so the gains are captured rather than lost at the review and sign-off stage.
• It teaches the limits explicitly: what AI does badly and when not to use it at all. And it builds the escalation triggers and psychological safety that let someone flag an output as wrong before it reaches a client or a filing.

An Illustrative example

A law firm asks about prompting training for its fee-earners. The diagnostic finds that drafting wasn’t the constraint: juniors were already using AI to produce first-pass notes, but partner review was absorbing the gains as no proper verification layer was in place.

The first step addresses any governance concerns and moves the process onto an approved enterprise tool, with a verification layer to make it easier for the partner to check each authority and quotation for both accuracy and faithfulness to the source.

The training component then covers onboarding support combined with refresher drills built around error-spotting on research notes containing known faults such as hallucinated cases or quotes missing important context (where possible, based on real AI output from the firm’s own use or published cases).

Three months on, the firm is measuring things like partner time per note, the proportion of notes returned for substantive rework, and the catch rate in a refresher drill. Those numbers show whether AI-assisted research is a capability or a professional-indemnity problem.

The best training is tailored to your goals and is accompanied by appropriate organisational and governance considerations, so the first step is knowing where your firm actually stands. That, rather than another workshop, is where the productivity improvement you went looking for begins to come back.

If you’re interested in learning more, Sindri offers a free AI maturity assessment (https://assessment.sindriconsulting.com/), followed by a diagnostic conversation if you want to go deeper.

---

_{1. BCG, “AI at Work” https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain‍}

_{2. Harvard Business School, “The Great Training Robbery” https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2759357‍}

_{3. Meincke, Lennart and Mollick, Ethan R. and Mollick, Lilach and Shapiro, Dan, Prompting Science Report 1: Prompt Engineering is Complicated and Contingent (March 04, 2025). Available at SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5165270‍}

_{4. Dunlosky J, Rawson KA, Marsh EJ, Nathan MJ, Willingham DT. Improving Students' Learning With Effective Learning Techniques: Promising Directions From Cognitive and Educational Psychology. Psychol Sci Public Interest. 2013 Jan;14(1):4-58. doi: 10.1177/1529100612453266. PMID: 26173288.}

_{5. Sources of Power: How People Make Decisions. By: Gary Klein. https://doi.org/10.7551/mitpress/11307.001.0001. ISBN (electronic): 9780262343244. Publisher: The MIT Press. Published: 2017.}

_{6. https://journals.sagepub.com/doi/10.1177/0149206309352880}

_{7 .BCG, “AI at Work” https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain}

_{8. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai}