AI Training For Employees is Lacking

I’ve noticed a few problems with AI training for employees. Before I go deeper, I should mention that its great firms ARE doing training, and the speed of generative AI adoption is fantastic.

Most employees have had some level of AI training. And most are experimenting with it in some form. Unfortunately there’s some recurring issues - all ironically easy to fix with minimal investment.

The most surprising part to me is not the poor training itself - its the opportunity cost of not doing it properly.

1: Vendor Training is Too Biased

Many vendors are offering AI literacy training bundled with their product. At face value this is great, but keep in mind that most vendors have an agenda: they want you to buy - and use - more of their stuff.

All the little decisions they make thereafter in regard to you happen with that in mind:

• What frameworks and tools is the sales engineer fluent in and willing to teach you about?
• What are the customer success manager’s KPIs? Many will say “to help you succeed,” but I guarantee their manager is asking them each week how your renewal / upsell / implementation is going.

I’ve also heard more than a few professionals say they felt the vendor did not understand their domain very well and it seemed more of a box-ticking exercise.

2: Safety isn’t properly addressed in most AI training for employees

There’s a ton of research on risks of generative AI use. Cognitive atrophy of users. Loss of ownership. Training tends to only partially cover this.

The most pernicious hazard of LLMs - which some refer to as “a feature” - is hallucinations. AI training for employees, in my experience, rarely goes into this deep enough. The solution, all too often, is “check AI outputs.” Which, if you’ve read the Pinsent Mason’s case (or the hundreds of others like it), doesn’t solve the issue.

Hallucinations are arguably an artifact of how LLMs work. The machine predicts the most likely acceptable answer. As we all know from math, when a system predicts the most likely outcome, a % of the predictions will be wrong. Which in turn leads to bad answers, and the LLMs unfortunately confidently state this as fact.

Referring back to point 1, I think sometimes in the industry we’re guilty of overhyping the tech - “they’ll get more accurate over time” (this is only partially true) or even when it goes wrong “the user should have used a better model / tool.”

The irony is that in many cases, the users are actually using a top-shelf solution. Which brings me to my next point.

3: The standard AI story for non-techos is confusing

AI will take your jobs. But you should use it at work.

AI is dangerous. But you can use it to help you complete complex - even high stakes - initiatives.

AI is super powerful and can make you more efficient. But you need to babysit it to make sure it doesn’t screw up simple tasks (its your fault if it does).

The irony of the last point is one I hear a lot of people complain about. If the AI is all powerful why do we need to babysit it so much?

Those who have seen the heralded productivity gains from AI tend to hold a more nuanced view.

4: AI users are not created equal

Several data points suggest that AI adoption is much faster in some fields than others.

Anthropic’s own Economic Index report in 2025 shows very high usage among software professionals and some reasonably high usage among creatives. Workday’s 2026 Future of Work report showed AI productivity gains (once rework has been factored in) are much higher among IT employees.

I suspect employees with a technical bent have a deeper understanding of it's limitations - and thus where and how to use it (or how not to).

5: Most AI Enablement is too generic

I’ll be brief on this one since it tends to apply to all types of training. But a frequent gripe from employees I hear is that the AI training they had was irrelevant to their day-to-day job.

It is not difficult to come up with some specific use cases for AI specific to your role. You can run a prompt yourself - or even throw your JD at a tool like Jobotron.

6: There are serious compliance and reputational risks

This isn’t the most exciting topic, which is perhaps why its often overlooked.

The EU AI act enforcement deadline looms on the horizon*. What that actually looks like (in terms of fines) is unclear. However, they have the OPTION to fine you millions of euros (minimum) for breaches.

Most people aren’t worried about the Act because they think “well I’m not an AI vendor, and I’m not making killer robots or totalitarian surveillance so I’m cool, right?” Others in countries like the UK or US will say “well, I’m not in the EU.”

The messy bit is the High Risk Use cases, and the Deployer rules.

Under the Act, if you use AI in your business, you are likely to be considered a Deployer. Deployers are liable under the Act. I keep thinking of the analogy of self-driving cars: if you’re in a Tesla and it mows down some people on Autopilot, the person in the driver seat is liable.

High Risk cases are more common than you’d think. The most obvious to me is anything to do with “screening.” If you are using AI to screen loan applicants, or job candidates, or help grade students’ answers… you’re probably High Risk. Which means you can expect more scrutiny by the regulators.

*Note: Originally, the act was set to be enforced from August 2026. As of last month, it appears they may defer this to 2027. In any case, there are other oversight groups who care significantly more. For instance, many professional services providers are finding their indemnity insurers auditing their AI usage to ensure proper governance and risk management.

There is an opportunity cost to all of this

When organisations avoid doing AI enablement properly, I find they end up in one of a handful of scenarios:

• Wild west, where people use their own tools with zero governance. This is high risk. All you need is one employee to accidentally put PII into their personal ChatGPT account and your firm committed a major privacy breach - or worse.

1. There’s a less extreme version of this situation which I’ll call “partial enablement” where people are given training on an official tool they find insufficient for their needs. They then go and quietly use another tool (eg Claude or a specialist one in the background). In tech land we call this Shadow IT, and it can lead to all kinds of compliance nightmares.

• Blanket bans, where employees can’t use AI at all, leaving you at an unfair disadvantage. There are firms who actually do this by the way.
• Hiring expensive workflow redesign consultants to tell you how to use AI in your business. This is actually not an awful outcome, and proper workflow redesign is often the biggest success enabler of all. And I’m a big fan of a lot of these consulting businesses. BUT what you really really want, is at some point, people driving change internally. There are very clear studies (pun intended) on how incredibly effective this is compared to change that’s only driven externally.

Give an employee a fish via AI, and you’ll feed them for a day. Teach them to fish with AI and you’ll feed them for life.

OK, so, what does good AI enablement look like?

AI training needs to include safety considerations - the actual concerns like:

• Hallucinations - and how to avoid them
• Cognitive atrophy and loss of critical thinking
• Sedation risk - running on autopilot, leading to overlooked errors

It also needs to feature a few other basics, like:

• Why should they use this tool - what’s in it for them? [I argue, most important thing]
• What should they be doing differently now that this tool is available - and what stays the same?
• Processes / tools that are needed to supplement any shortfall. For example, a legal citation checker to help reduce hallucinations
• What are some things they can start doing right now that will be beneficial to them?
• An incentive or follow-up to ensure they continue using it. I’ve seen some interesting approaches here, from the aggressive (a CEO who told employees they wouldn’t be eligible for promotion without clear demo of AI use in their next review) to the chill (show-and-tell workshops).

Separately, its always helpful to review KPIs and incentives to make sure there’s not an obvious conflict. An example we see right now in many law firms is that the billable hours model is creating a perverse motive to not be TOO efficient with AI. The sad part is clients don’t feel the same way - or as Richard Susskind recently put it at a LegalGeek event - “the market's not going to show any loyalty for my old ways of working.”

For what its worth, the bar isn’t very high to make an impact. Google’s own research found that a 2H investment in AI training for employees is likely to yield a 2X increase in AI usage. I’ve seen similar stats from other players who aren’t necessarily vendors.

We’re piloting all the above - and more - at monboard. Head over to our services page to find out more - or drop us a ping on the Contact section.