As you know, having spent my years in marketing, there are some vanity metrics that are often relied upon too heavily. Followers and impressions that go up and to the right look great in theory, but tell you very little about the actual impact you're creating.
Learning and development has its own vanity metric, and it's course completion. Even when organisations do invest in their people, the number they reach for is completions: modules finished, certificates earned, the percentage of the workforce now "trained". That's worth knowing. But it tells you people showed up. It doesn't tell you whether anything has changed in how the work gets done.
AI use is rising, but support is slipping
Three interesting reports landed in the last fortnight that highlight the real gap emerging in how organisations are building the AI skills they need.
nCino's AI in Banking Benchmark found that 84% of banks are deploying AI, but only 55% are actively reskilling their workforce and 33% are hiring for AI experience (Source: nCino, via FinTech Global). PwC's Global Workforce Hopes and Fears Survey found that 64% of people now use AI at work, while the share who say they have the learning and development resources they need has fallen from 59% to 51% (Source: PwC).
Different surveys, different samples, so I won't pretend they add up neatly. The direction is consistent, though. AI use keeps going up. Support for building the skills is flat at best, and slipping in places.
Then there's the number that made me write this issue. Pluralsight found that 96% of organisations have invested in AI tools, yet only 52% say their workforce is AI literate. And 77% of tech executives say they have watched people complete training or certifications and still not apply the skills (Source: Pluralsight Tech Skills Report).
When more than three quarters of tech leaders have seen that happen, completion is a poor stand-in for capability. Yet it's still the number most of us report upwards.
You can tick the box and change nothing
This week Rob Barnett spoke at the FStech Future of AI conference. It was a day dominated by risk, compliance and governance, but Rob's talk was one of the few on the people side.
I've known Rob for years, and he's not afraid of speaking his mind. The line that stood out to me was about people who get "sheep dipped into training, you might be able to tick the compliance box, but it's not going to make a difference to the organisation."
The sheep dip is a good image, because it's exactly what the completion metric measures. Everyone went in. Everyone came out the other side. Nobody can tell you whether they're any better at the job.
Rob also named something that sits as an unknown, unquantified cost on every company's P&L: AI capability debt. Most of us know tech debt, the accumulated cost of old or improper technology that keeps weighing an organisation down. AI capability debt is the people version. It's the capability that already exists in your workforce but goes unrealised.
And it's building up faster than tech debt ever did. It shows up as a widening gap between what AI can do and what people are able to use it for, and it's fast becoming the biggest bottleneck in the business.
Rob wasn't the only one making the case. Yogesh Sholapurkar of Close Brothers, who has spent 30 years in banking technology and served as the bank's CIO, argued that upskilling belongs among the most strategic investments a firm can make, rather than sitting in a support function. Hearing that from someone who has run the technology side of a bank carries weight.
Put the training in the flow of work
Real learning happens through immersion, which means training needs to look very different. Training on factual items, like compliance or a specific business system, where there's a known, clear answer, has its place. AI is different. The real skills are knowing whether to point AI at a task at all, knowing which workflows AI can change the most, and knowing how to judge a good output. Those can only be built in the flow of work.
Rob described it at the conference as finding the ten things in someone's process that cost them the most time or cause the most difficulty, and building the learning around fixing those. The employee gets something back, which matters because, in his words, "You can't transform an organisation without taking people with you." Yogesh took it a step further: capture the patterns, evidence and lessons as you go, so one team's progress becomes the next team's starting point.
I should declare an interest, because this is what we built AcademyAI to do. It builds the AI skills that matter most in a role-specific way, in the flow of work, so an organisation can make capability visible, track it as it grows, and start paying down its capability debt.
Judgement is what makes the rest safe
One of the most important skills to develop, as part of broader AI capability, is judgement. Prompting is the easy part. The skills that hold up as the models change are knowing when to reach for AI, thinking about the work as a system, and judging what comes back. Everyone needs a strong point of view on how outputs get reviewed before they're passed on.
Many of the delegates at the event shared their own examples of workslop: poorly reviewed AI output sent on to line managers to check. That just shifts the bottleneck somewhere else in the business, as we wrote about in AI didn't remove the work.
These skills matter just as much for risk and governance, because people are the first line of defence.
Yogesh made the same case from the CIO's perspective: clear rules, boundaries and ownership, with human review and approval routes proportionate to the risk, so people have the confidence to act rather than waiting to be told.
Rob's take on shadow AI fits the same theme. He wouldn't pick between "ban it" and "people reaching for consumer tools are telling you something", because both are true: "it's banning what you need to ban and it's listening where you need to listen."
Where to start
If the research is even roughly right, the next step for most organisations isn't another tool or another course catalogue. It's three things.
First, assess the gap. Find out where capability actually sits today, by role and by workflow, rather than who has a licence or who has finished the module. Treat it as a living diagnostic rather than a one-off survey, because the tools and the work keep moving. When the frontier shifts every quarter, "fully capable" isn't a fixed destination. Progress against what the role now demands is the more honest measure.
Second, close the gap with role-specific learning. A baseline of AI fluency for everyone, then the judgement and workflows each role actually needs, tied to problems people already own.
Third, evidence the improvement. Look for changes in the work itself, at team or function level, and check that what you reward points the same way. Be careful what you lead with here. Cost-saving case studies are an understandable place to start, but as Rob put it, "you can't shrink yourself to success." If your first wave of AI wins all ends in fewer people, you've made the cultural shift much harder for everyone who comes next.
Yogesh closed his talk with a thought I keep coming back to. Today's model, tool or interface won't be the last one. The biggest value won't go to the organisation with the most AI, but to the one that does the best job of helping its people grow.
Completion tells you people showed up. Capability tells you something changed. The second is harder to measure, but it's the one your board will eventually ask about.
So what are you reporting upwards on AI skills right now, and does it tell you whether the work has changed? I'd love to hear what has and hasn't worked for you.
Thanks, as always,
Andy
PS. If you'd rather talk this through live, we're running a webinar on 13 October. Sign up here.
Sources
- nCino AI in Banking Benchmark, reported by FinTech Global, 28 September 2026: fintech.global
- Pluralsight Tech Skills Report, 30 September 2026: PR Newswire
- PwC Global Workforce Hopes and Fears Survey 2026, 29 September 2026: PR Newswire
- Rob Barnett, session at FStech Future of AI conference, 6 October 2026
- Yogesh Sholapurkar, Close Brothers, talk at FStech Future of AI conference, 6 October 2026
