Do you know whether your AI programme worked, and what would show it

Whether organisations are adopting AI is settled. Whether they know what it gave them is not.

The adoption has happened

In Tieto's Nordic AI survey, conducted in February 2026 among 623 respondents in Finland, Sweden and Norway, 31 per cent say AI is now in production across the business. Three in ten use AI solutions to a large extent, and nearly a third have already begun experimenting with their own AI agents.

The first two figures measure different things, one at organisation level and one at employee level, but they come from the same 623 responses and therefore do not corroborate one another. The survey is built on an online panel run through Norstat, and Tieto states itself that the sample does not represent the workforce but consists of people who take part in or support AI decisions. The figures describe what decision-makers report, not what is happening inside the organisation.

Two years ago those numbers would have been news. Today they mostly say that adoption is no longer what separates one organisation from another.

But one in four has nothing to measure against

The same survey shows that 24 per cent of organisations still lack formal metrics for their AI use. The share varies by country, from 31 per cent in Norway to 17 per cent in Finland.

One organisation in four has therefore adopted something that cannot be evaluated. That does not mean it is not working. It means nobody can say whether it is, and that the decision to continue or to stop is consequently made on impressions.

A programme without metrics cannot fail. It can only stop being mentioned.

That is the uncomfortable property of unmeasured initiatives. They rarely end with a decision. They thin out, change name, become part of something else, and nobody ever has to say that it came to nothing.

Where somebody measured, the outcome can be settled

In November and December 2025 Gartner asked 782 infrastructure and operations leaders how their AI use cases had fared. Twenty-eight per cent succeeded fully and met the return expectations that had been set. Twenty per cent failed outright. Gartner reports neither countries nor job titles nor sampling method, and that caveat belongs with the number.

The population is a different one from Tieto's and the question is a different one, which makes the figures an independent addition. But the interesting part is not the twenty-eight per cent. It is what is left over: a good half of the use cases land neither among those that succeeded nor among those that failed.

Where somebody has measured against a stated expectation, the outcome can be settled in either direction. Where nobody has, the initiative lands permanently in the middle, neither confirmed nor written off, and that is not a technical outcome but a measurement outcome.

The commonest metrics are the hardest

Among those who do measure, the two commonest metrics are productivity, cited by 33 per cent, and cost savings, cited by 31 per cent.

Both are reasonable things to want to know. Both are also the hardest of all to attribute to a single initiative. Productivity moves with staffing, season, case mix and whatever else the organisation did that quarter. Costs move with pricing, contracts and volume. When one of those two numbers moves in the right direction it is rarely possible to show why, and when it moves the wrong way it is just as rarely possible to defend the initiative.

The result is a metric that sounds important and that nobody ends up trusting.

The reported obstacles are not technical

The survey also asks what is slowing things down. Security concerns are cited by 45 per cent as the main barrier, and 39 per cent say the skills gap is slowing adoption.

Neither is a technical obstacle, and that is worth pausing on. Both security concerns and a shortage of skills are the kind of thing negotiated with evidence: a risk assessment, a boundary, a plan. Evidence in turn presupposes that somebody measured something. Organisations that do not measure therefore find it harder to get past the very obstacles they themselves report as the largest.

What a usable metric looks like

A metric will do if it meets three conditions.

1. It can be measured before the initiative. Either the history already exists, or it is measured for a few weeks before anything changes. A metric that comes into existence at the same time as the solution is not a metric, it is a description of the solution.
2. It concerns a bounded step. The lead time for a defined case type, the share of cases that must be corrected afterwards, the number of steps requiring manual handling. Not the productivity of a whole department.
3. It is not moved by ten other things at once. If the metric moves when somebody leaves, when volume rises or when a contract is renegotiated, it is not measuring the initiative.

Those conditions deliberately rule out productivity and cost as primary metrics. Both may well be the reason for the programme. They are simply unsuited to serving as evidence for it.

If the programme is already done

The commonest situation is not that the metrics are poor, but that they were never decided. There is a way back from that which works surprisingly often.

Write down what was assumed would change, in plain words and without figures. Then work through the list and see which of the assumptions has a counterpart in data already being collected: in the case management system, in time reporting, in the logs. There are usually two or three, and they can be compared backwards because the history is already recorded.

What cannot be reconstructed are quality metrics nobody recorded before the implementation. There the only route is to start measuring now and accept that the comparison begins here. That is worse than having measured from the start and considerably better than continuing to guess.