AI did not fire anyone. We simply did not hire the replacement

A few weeks ago, one of the lawyers in our department left after more than ten years with us.

The departure had nothing to do with artificial intelligence. There was no plan to reduce headcount, and nobody had decided to replace a lawyer with a language model.

A vacancy simply appeared.

And with it came a fairly traditional business question:

Who do we hire now?

Our case presents an additional difficulty. We needed someone with sufficiently strong knowledge of Spanish law who could also work professionally in Norwegian.

There are not many of them.

We started recruiting.

We did not find anyone who convinced us.

We then considered a different solution. Perhaps we did not need to replicate the departing profile exactly. We could strengthen the administrative side of the department and redistribute the legal work among the existing lawyers.

We interviewed candidates for that role too.

Again, we did not find the right person.

And then something interesting happened.

While searching for someone to hire, we still had to do the work.

The reorganisation that was supposed to be temporary

We could not wait indefinitely for the perfect candidate.

We redistributed responsibilities, examined workflows and began using artificial intelligence more systematically to absorb particular categories of work.

Not to replace legal judgement.

To reduce the work surrounding that judgement.

Document preparation. First-pass review. Comparing texts. Summarising. Extracting information. Organising files. Preliminary research. Draft preparation. Classification. Translation. Structuring facts. Turning disorganised information into something a lawyer can check much more quickly.

All these tasks existed before.

They now require less human time.

Initially we regarded the arrangement as temporary while we continued searching.

Around two weeks have passed.

So far, something has happened that I did not expect to be quite so clear:

the department's workload, both in working hours and in mental load, does not appear to have increased.

We need to be cautious. Two weeks is not an experiment. Work volumes may change and a particularly busy period may reveal limits we cannot yet see.

But the organisational observation is worth considering.

One person left.

We have not replaced that person.

And the system continues to work.

Nobody decided to replace a job with AI

This distinction seems fundamental to me.

Discussions about AI and employment often imagine a deliberate decision:

“We employ ten people. We install AI. We now need eight. We dismiss two.”

That will happen in some organisations.

But I suspect that a large part of the transformation will be considerably less spectacular.

Someone retires.

Someone changes employer.

A department that would ordinarily recruit two additional people decides it needs one.

A vacancy appears.

And before publishing the job description, management asks:

Do we actually need to recreate this position exactly as it existed?

That may become one of the most important labour-market questions of the coming years.

In our case, AI did not eliminate a job.

The job was already vacant.

AI changed the calculation about whether it had to be filled again.

Automation may enter through vacancies

Businesses are not static structures.

Every month there are departures, retirements, reductions in working hours, growth, changes of role and reorganisations.

AI therefore does not need to produce waves of visible redundancies to alter labour demand substantially.

It can change the flow of hiring.

A company of eight that previously would have grown to ten may grow to nine.

A five-person department losing one employee may discover that four people can sustain the output.

A business may keep the same number of employees while doubling its capacity.

And expanding companies may hire different profiles from the ones they would have recruited three years earlier.

The effect could take years to become obvious in unemployment statistics while already altering the structure of the labour market.

Evidence compatible with this mechanism is beginning to emerge.

A recent Stanford Digital Economy Lab working paper analyses 1.25 billion job postings and 154 million employment records across 41 countries and infers AI adoption from job advertisements involving generative AI use. It estimates that foreign affiliates of AI-adopting firms reduce the relative share of junior workers compared with senior workers, relative to comparable control affiliates. Significantly, the measured effect comes largely from stronger growth in senior employment rather than dramatic absolute junior job losses. Stanford Digital Economy Lab

Another study, using vacancy data from Flanders (Belgium) for 2021–2025, observes that, at peak adoption, greater exposure to generative AI is associated with fewer entry-level vacancies, while the estimate for experienced professionals is small and positive. The pooled estimate is smaller and less precise. papers.ssrn.com

These studies do not prove what is happening in our department.

But they describe a mechanism consistent with it:

AI may first change whom we stop hiring before it changes whom we dismiss.

Small businesses may notice this effect sooner

There is another reason why this dynamic may be particularly important for small and medium-sized organisations.

For a multinational, adding or removing one person barely changes the structure.

For a small organisation, one person can represent 10, 20 or 30 per cent of a department.

Each hiring decision matters more.

SMEs are also particularly affected by shortages of certain profiles. They cannot always compete with large organisations on pay, and in specialised sectors the exact combination of knowledge they need may simply not exist in the local labour market.

The OECD has identified precisely this dynamic: some SMEs using generative AI mention the ability to compensate for skills gaps and labour shortages among its benefits, alongside improving productivity. OECD

That describes our situation rather well.

We did not arrive at this reorganisation because consultants presented a cost-reduction plan.

We arrived there because work needed to be done and the labour market did not immediately provide the person we wanted.

Norway may be a particularly interesting laboratory

This mechanism may be particularly relevant in Norway.

Statistics Norway’s data for 2026 show that 48% of Norwegian enterprises with at least ten employees, in the sectors covered by the survey, now use one or more AI technologies, up from roughly three in ten in 2025. Even among the small enterprises included in the survey, with 10–19 employees, the figure is four in ten; among large enterprises, with 100 or more employees, it is eight in ten. SSB

Norway combines that rapid adoption with a high-cost labour market and, for some specialised professions, a relatively small pool of candidates.

When finding the right person is difficult and each professional hour is expensive, the economics of automating preparatory tasks change quickly.

AI does not have to “replace a lawyer”.

It merely has to make it possible for four lawyers to manage reasonably the work that previously required five people across legal and supporting tasks.

That is a much lower threshold.

And probably a far more realistic one.

Tasks are not professions

None of this means that a language model can replace a lawyer.

Much of legal work depends on capabilities that models still exercise imperfectly or simply should not exercise independently.

Understanding what a client actually wants.

Recognising that an apparently minor fact is legally decisive.

Choosing a strategy where several answers are defensible.

Negotiating.

Accepting professional responsibility.

Deciding which risks are reasonable.

Knowing when a technically correct answer is a poor practical decision.

And checking.

AI lowers the cost of many tasks.

That does not necessarily eliminate a profession.

It may instead increase the relative value of its most professional elements.

The transition may therefore be less:

lawyer → AI

and more:

lawyer + auxiliary tasks → AI-augmented lawyer.

The uncomfortable junior problem

There is, however, an uncomfortable consequence.

Many of the tasks being automated are exactly the tasks traditionally performed by junior professionals.

Drafting the first version.

Summarising a file.

Researching case law.

Comparing documents.

Structuring facts.

Building a chronology.

Those tasks may have little economic value when performed by an experienced lawyer.

But historically they served another function:

they taught people how to become lawyers.

An important paradox emerges here.

If we automate all junior work because a senior professional with AI can complete it faster, we may greatly increase current productivity while simultaneously destroying the mechanism through which we train the senior professionals of the future.

That is not an argument against automation.

It is an argument for deliberately designing professional training.

Organisations may need fewer juniors to produce the same output. But those they recruit will have to learn differently and will probably take on higher-level tasks much sooner.

We may still hire someone

I also want to make something clear.

I do not know whether we will conclude three months from now that we need another lawyer.

Work volumes may increase.

Some tasks may prove less scalable than they currently appear.

Mental load may return even if hours remain controlled.

Or we may meet an exceptional candidate tomorrow and decide that the additional human capability is worth far more than its cost.

This is not an article about a final staffing decision.

It is about how the question itself has changed.

Previously, when somebody left the department, the almost automatic response was to replace the person.

Now there is an intermediate stage:

What work did this person actually perform?

Which tasks still require professional judgement?

Which can be simplified?

Which can be automated?

Which can the existing team absorb with better tools?

And what human capability is still missing afterwards?

Only then do we know what we need to hire for.

From org charts to capability systems

Perhaps that is one of the most profound organisational shifts.

Companies traditionally think in roles:

lawyer, administrator, accountant, assistant, developer.

AI forces us to think first in capabilities and tasks.

A department needs a certain capacity for legal analysis, document handling, client communication, translation, research, billing, administration and coordination.

Until recently, almost all of that capacity had to come from people.

It can now be distributed across:

people + conventional software + automation + AI models + agents.

The organisation chart no longer fully describes how the organisation produces work.

AI may enter companies without anyone “implementing AI”

This may be the most interesting part of our experience.

We did not hold a strategic meeting about reducing headcount through artificial intelligence.

We did not hire consultants.

We did not set a target for labour savings.

We had a problem.

Someone left.

We could not immediately find the person we wanted.

Work continued arriving.

We reorganised.

We used a technology we already had more intensively.

And, for now, we discovered that the system continues to operate with fewer people.

AI entered through the gap left by a vacancy.

I believe a large part of its impact on small and medium-sized organisations may occur exactly this way.

Not through spectacular automation announcements.

Through hundreds of small, rational decisions:

“Let's wait before we hire.”

“Let's see if we can reorganise this first.”

“That preparatory work no longer needs as many hours.”

“Perhaps we do not need the same profile.”

Individually, each decision seems insignificant.

Multiplied across millions of companies over several years, they could completely transform the labour market.

THE SIGNAL

We often ask how many jobs artificial intelligence will destroy.

Perhaps that is not the best question.

A more useful question could be:

How many jobs that exist today will be recreated when they become vacant?

Our small experiment cannot answer that question.

But two weeks ago, we thought we needed to hire someone.

Today, we are no longer certain.

And that uncertainty, silently repeated across millions of companies, may prove far more important for the labour market than any dramatic announcement about machines replacing workers.