Buy the Certificate Only When Someone Outside Asks for It

What does a foundational AI certificate actually test?

Recognition of concepts under time, inside one vendor's product line. The exam guides are unusually candid about this. AWS says the target candidate for its AI Practitioner exam "uses but does not necessarily build" AI solutions and lists developing or coding models as out of scope ¹. Google's Generative AI Leader is 90 minutes of multiple choice with no prerequisites, aimed at "anyone in any job role, with or without hands-on technical experience" ².

Then there is the credential most often listed as the entry-level AI qualification, which is not a business credential at all. The Microsoft Azure AI Fundamentals exam page states that the candidate is at the beginning of a career in AI solution development and needs knowledge of Python syntax and programming techniques, with the majority of the exam weighted toward implementing solutions in one platform ³. Anyone recommending it to a non-technical team has read the name and not the page.

None of these are bad instruments. They are accurate instruments for something other than the question being asked. A pass shows that a person can pick the right answer from a list, about a defined body of vocabulary, on one day, under time pressure. The AWS and Google guides describe no performance task at all, which means no moment where an assistant produces something plausible and the candidate has to decide whether to check it.

That gap is the entire subject. What a team is short of is not vocabulary. It is the habit of stopping when the output looks right.

When is a certificate the right purchase?

When someone outside the organization is asking for it by name. A client contract that specifies certified staff, a partner tier that counts credentials, a procurement questionnaire with a line for it, a regulator or professional body that requires one. In those cases the badge is doing a real job: it satisfies an external requirement, and no amount of internal evidence substitutes.

Outside that, the test is simple. Ask who reads the credential and what they do differently once they see it. If the answer is a hiring manager two years from now, that is a benefit to the employee rather than to the employer paying. Say which one the budget is buying.

The compliance argument is the one most often invented, so check it before repeating it. For EU-facing employers, the European Commission's own Q&A on the AI Act literacy duty, as it read in August 2026, states there is no obligation to measure employees' AI knowledge and "no need for a certificate", with organisations advised only to keep an internal record of training ⁴. The Q&A is guidance, not law, and the Commission has already rewritten that page once, so read the current version and put the compliance question to counsel before it reaches a board paper. It still means a vendor selling a badge as the route to compliance is selling something the regulator did not ask for.

Where a credential is genuinely required, buy the specific one named and stop there. Certification programs are designed to ladder, and the second exam almost never has an external buyer behind it.

Where should the certification budget go instead?

Into protected hours and a senior reader. The vendor curricula are free, so the real cost of training has always been the time people are not doing their jobs, and that is the line most programs leave unfunded while paying for exam vouchers. Fund the hours first. Then fund the person who reads what came out of them.

The reader is the part no course contains, because it requires knowing which claim in your specific work is expensive to get wrong. Four hours a month of a senior practitioner's time, spent reading finished deliverables and writing three sentences about each, produces something an exam voucher cannot: evidence tied to the work you actually ship. That is the spine of a program, which is why it belongs in the design from the start. Build the upskilling program around that comparison and the certification question mostly answers itself.

A rough allocation that survives scrutiny, for a team of thirty.

• Access and licences. Whatever the tools cost, plus one page saying what may go into them.
• Protected practice hours. Four sessions of ninety minutes per person, costed at loaded salary. This is the largest number and the one that gets cut first.
• Reader time. Two hours per participant across the program.
• Credentials. Only the ones with an external buyer, named individually.

The order matters more than the amounts. A program funded in that sequence still works at half budget. Funded in reverse, it produces a spreadsheet of badges and no evidence that anything changed.

How do you answer the person who asks for one?

Say yes to the time and be specific about the badge. Someone asking for a certification is usually asking for two things underneath it: permission to spend work hours learning, and something durable to show for it. The first is cheap and worth granting immediately. The second can be met better than a badge meets it, and the conversation goes badly only when the request is refused without the substitute.

Employees are not wrong to want the credential. Postings mentioning AI skills advertise salaries 28% higher than postings that do not, roughly $18,000 a year on Lightcast's reading of its posting database, and more than half of the 2024 postings asking for AI skills sat outside IT and computer science ⁵. That is a raw comparison of advertised salaries between two groups of postings rather than a like-for-like wage premium, and AI-mentioning postings skew senior and urban, so it does not mean a badge is worth $18,000. It does mean the market signal your team is reading is real.

Offer the substitute plainly. A written reference from the senior reader describing specific work, naming the deliverable and what the person did with it, travels further in an internal promotion case than any exam result, and it is the same artifact the program already produces. Where the promotion criteria have not been written down yet, that is the gap to close first, and what counts as good AI work in a promotion case is the conversation to have before the certification one.

If the ask is specifically about prompt engineering, the same rule holds and cuts harder, because the outside buyer that would justify a credential almost never exists for that one: a prompt engineering course is usually not the purchase, and a prompt certificate compared against a work sample is the sharper version of the same question. Grant the hours, ask to see the work, and keep the fee.

See the benchmarks >