55% use AI – only 10% have integrated it into operations

The number everyone quotes, and the number that matters

In 2025, 55 per cent of Norwegian companies said they use artificial intelligence. Two years earlier the share was 24 per cent, according to NHO. That is the figure most frequently cited since the report was published.

The same survey, which covers around 4,000 companies, contains another number that receives far less attention: only about 10 per cent of companies state that AI is well integrated into the business. In industry, roughly half of the companies use AI daily, without that translating into the integration figure.

The distance between 55 and 10 is not a rounding error. It is the difference between a technology that is available to employees, and a technology that has changed the way the business is actually run.

Usage is an activity. Integration is a cost structure.

Using AI changes how a task is performed. Integrating AI changes how the business is staffed, priced and organised around delivery.

This distinction is easy to miss because it is invisible in the employee experience. A caseworker who saves twenty minutes on a draft experiences a genuine improvement. The accounts register nothing. The hour has already been paid for through fixed salary, and time saved that is not reallocated to more volume, higher quality or fewer full-time equivalents is idle capacity – not improved earnings.

"Hours saved" is therefore a gross figure. It describes what was freed up, not what was extracted. The accounts measure only the latter.

Three figures that separate a pilot from operations

If AI usage is genuinely integrated into operations, it will leave traces in three figures a chief executive already has in the reporting pack.

Contribution margin per employee. Contribution divided by the number of full-time equivalents shows whether freed capacity has been converted into revenue. If this figure rises without headcount growing, the time saved has been turned into volume or margin. If it stays flat while AI usage increases, the gain is still sitting unredeemed in people's calendars.

Payroll cost as a share of revenue. This is the most unforgiving test of integration. The share falls in only two ways: staffing is adjusted, or revenue grows faster than the wage bill. Neither happens by itself because a tool has been adopted.

Cost per delivery. Unit cost – per case, per order, per project, per customer month – is the closest a company gets to its own productivity measure. It is also the only one of the three that isolates efficiency gains from price increases and sales growth.

What all three have in common is that they require a basis for comparison from before AI was introduced. Without a baseline, any improvement becomes a matter of belief.

The effect is too small to be spotted through anecdotes

The NHO report draws in part on the OECD's review of the field, which estimates that AI may contribute between 0.2 and 1.3 per cent to productivity growth. That is a substantial estimate at the macro level, and a demanding one at the company level.

An annual effect of that magnitude drowns in everything else happening in an operating account: a changed customer mix, price adjustments, sick leave, a single large contract. It is impossible to isolate in reporting based on what employees feel they are saving, but entirely measurable in a time series for unit cost across four to eight quarters.

That moves the bottleneck. The scarce resource in Norwegian AI initiatives is not access to models, which in practice is a subscription cost. It is the measurement apparatus capable of separating half a percentage point from noise.

Freed-up time is not the same as improved earnings

The report's projections point towards broad AI use delivering the same output with significantly lower time input towards 2045. The wording is worth reading closely, because it describes a release of capacity – not an automatic improvement in earnings.

From the same starting point, two entirely different routes run through the accounts. One reduces headcount and takes the gain on the cost side. The other keeps headcount and takes the gain on the revenue side, with more deliveries handled by the same number of people.

Both can be defended. But they produce different trajectories for the wage share, different capital tie-up and different risk, and choosing between them is an ownership decision – not a consequence of the technology. Companies that have not made the choice usually end up keeping both the headcount and the cost, and calling it innovation.

Why the youngest companies end up last

The report also notes that early-stage companies have adopted AI to a lesser degree. That is counterintuitive, given that they have the least legacy technology to dismantle.

The explanation probably lies in what integration actually requires: repeatable processes to automate, enough history to detect a difference, and unit economics stable enough for the effect to be separated from growth. In a company doubling its revenue, a few percentage points of efficiency disappear entirely into its own momentum.

Samfunnsøkonomisk Analyse describes the relationship as a gradient: the broader and more integrated the use, the larger the gains. Turned on its head, that means partial integration can sit in a range where the cost has been incurred but the gain is not yet large enough to cover it. That is an entirely ordinary position to be in – and far easier to defend in the boardroom if it has been identified as a phase rather than as a result.

From technology project to investment case

The report was launched on 12 January 2026 at the conference "From buzzword to bottom line". The title pinpoints exactly where the 45 percentage points between usage and integration reside: in the chart of accounts.

An AI initiative treated as a pilot is assessed on uptake, enthusiasm and user numbers. An AI initiative treated as an investment is assessed on capital committed against changes in contribution margin, wage share and unit cost – with the same documentation requirements as a machine investment or a newly hired sales team.

There is no technological difference between the two. The difference lies in which numbers are presented, and how long they have been measured.