Certify your people on their own work, not on an AI course

Nearly a quarter of finance firms now tie raises, bonuses or promotions to employees’ AI proficiency, according to a Selby Jennings survey of 500 finance-industry HR professionals, reported by Reuters on October 14. About 30% require people to disclose their AI use in formal performance reviews. 42% have written formal criteria for evaluating AI skills.

Dylan Gomez of Selby Jennings also told Reuters there is no established formula for how much pay should depend on AI use and performance.

So companies are starting to pay for a skill they don’t yet have a good way to measure. The fix is practical, and it doesn’t need a new vendor.

DataFinance firms and AI skills, per Selby Jennings

Formal AI skill criteria42%
AI use disclosed in reviews30%
Pay tied to AI proficiency25%

Firms are writing AI skills into reviews and pay before agreeing on how to measure them. Source: selbyjennings.com

Training is shrinking while use is growing

The pressure to measure is landing at an awkward moment. PwC’s 2026 Hopes and Fears survey of 49,364 workers found that 64% used AI at work in the past year, up 10 points. Over the same period, the share who said they could access the learning and development they need fell from 59% to 51%, as The Next Web reported on September 29.

The split inside that number is worse. PwC’s largest group, which it calls the “engine room,” makes up 56% of workers. Only two in five of them said they could get the learning they need.

So the people expected to show AI skill in their next review are getting less help building it. If the measure is “did you complete the course,” you’ll get completed courses. If the measure is “did you use the tool,” you’ll get usage. Neither tells you whether someone does their job better.

Why generic AI courses don’t prove much

A general AI course teaches general things: how to write a prompt, what a model is, what not to paste into a chatbot. That’s worth knowing. It doesn’t tell you whether your accounts receivable clerk can use AI to cut the time on collections emails without making a mistake a customer will notice.

Skill at work is specific. The same tool is a different job in the hands of a recruiter, a support lead and a controller. A certification that means something has to test the person on the work they actually do, with the standard their manager already uses to judge that work.

That’s also the only kind of certification you could fairly tie to pay. A score on a generic quiz is easy to game and hard to defend. A record showing someone handled ten real cases from their own job, faster and at the same quality, is evidence.

How to certify someone on their own work

This is the method I’d use. It needs no software and no training budget.

Pick two or three real tasks per role. Choose tasks the person does every week, where AI can plausibly help: drafting client updates, reconciling a report, summarizing call notes into the CRM. Let the manager pick them, not the training team.

Write the standard before the test. For each task, write what a good result looks like and what an acceptable time is. Use the standard the manager already applies. If there isn’t one, write it now, because you need it anyway.

Have the person do the task with AI on live or recent work. Real cases, not practice ones. They choose how to use the tools. You want to see judgment, including when they decide not to use AI for a step.

Have a reviewer check the output against the standard. The reviewer is someone who knows the work, usually the manager. They check the facts, the numbers and the tone. They also note the time.

Certify, or send it back with a note. If the work meets the standard, record the certification with the task, the date and the reviewer. If it doesn’t, the person gets specific feedback and tries again on a new case.

Recertify when the tools or the job change. A certification from a year ago on a tool you’ve since replaced doesn’t mean much. Set a review date.

Here is what that looks like in practice, as a hypothetical. Say you run a 40-person accounting firm, and one of the tasks you pick for staff accountants is the monthly variance note: the short explanation that goes to a client when an expense line moves more than expected. The manager writes the standard. The note names the line, the amount and the likely cause, it checks the cause against the ledger, and it reads like the firm wrote it. An acceptable time is whatever the manager would accept from a strong accountant today.

The accountant writes five real variance notes for this month’s clients using whatever AI tools the firm approves. The manager reviews each one before it goes out, as they would anyway. Four pass. One cites a cause the ledger doesn’t support, so it goes back with a note. The accountant redoes it on a new case and passes. That person is now certified on that task, with a record anyone can read.

LoopCertifying one task

1Pick a real task from the job
2Do it with AI on live work
3Reviewer checks the standard
4Certify or retry with a note

The test is the job itself, judged by someone who knows the work.

It’s the idea behind AIXE Operator, the certification we run at AIXE, where team members prove AI skills on tasks from their own jobs. You don’t need a program to start, though. A spreadsheet and a manager who cares will get you most of the way.

What to watch for

Don’t certify usage. Logging into a tool every day is not a skill. Some of your best people will use AI less than others and get better results because they know when it helps.

Keep the reviewer close to the work. A central training team can run the process, but the person judging the output should know what good looks like in that role.

Record time honestly. Include the time spent checking the AI’s work. If someone saves twenty minutes drafting and spends twenty-five fixing, the task did not get faster.

Separate the certificate from the pay decision at first. Run a quarter of certifications before tying anything to compensation. You’ll learn which tasks were the right ones and which standards need adjusting, and nobody’s raise will depend on your first draft.

ChecklistBefore your first certification round

Two or three tasks per role
A written standard for each
A reviewer who knows the work
A place to record results
A recertification date

What to do this week

Pick one team and ask its manager for the three weekly tasks where AI could help most.
Write the standard for each task: what good output looks like and an acceptable time.
Ask two people on the team to do each task with AI on this week’s real work.
Have the manager review the results against the standard and record pass or retry.
Collect the retry notes. They tell you exactly what training the team needs next, which is better information than any course catalog.