AI’s hidden cost

Today I want to change the tone a little.

No restaurant story.

No behavioural economics anecdote.

No joke about Procurement, although the temptation remains strong.

I want to share a concern.

Most companies think they understand the cost of AI.

A subscription.

An enterprise licence.

An API bill.

A token invoice that someone in Finance will eventually pretend to understand.

But I do not think that is the real cost.

The real cost is quieter.

Every time a legal team uses AI seriously, it gives the system something more valuable than money:

how the organisation thinks.

Not just documents.

Judgment.

Negotiation positions.

Risk appetite.

Preferred language.

Corrections.

Workflows.

What the GC considers “good”.

What the firm considers “client-ready”.

What the business tolerates.

What Legal never wants to concede.

This is where the debate around AI sovereignty becomes important.

Palantir has recently been making the case that institutions must retain control over their data, models, compute and operational advantage.

Satya Nadella has put a very useful name to the problem:

the Reverse Information Paradox.

Kenneth Arrow’s classic information paradox said that the seller of information risks giving away knowledge in order to prove its value.

AI reverses the problem.

Now the buyer risks giving away knowledge just to use the product.

You pay once with money.

Then you pay again with the proprietary intelligence required to make the tool useful.

Nadella’s sentence is worth sitting with:

“In consuming intelligence, you are creating intelligence. And what you create should belong to you.”

For lawyers, this should ring very loudly.

A legal department’s real IP is not only in its contracts repository.

A law firm’s real IP is not only in its precedents.

The most valuable knowledge is often in the invisible layer:

how senior lawyers improve drafts, reject arguments, frame risk, choose trade-offs, brief business leaders, negotiate pressure points and decide when to push and when to move.

That is not generic data.

That is institutional intelligence.

And AI systems learn from that intelligence.

Prompt by prompt.

Correction by correction.

Evaluation by evaluation.

Workflow by workflow.

This does not mean companies should avoid AI.

That would be absurd.

The winners will use AI deeply.

But they should not confuse adoption with surrender.

The question is no longer only:

Is this tool secure?

The deeper question is:

Who owns the learning loop?

If your team corrects an output, does that learning stay inside your organisation?

If your lawyers build prompts, playbooks, evals, negotiation patterns and workflow memory, does that compound for you?

If a model provider changes terms, pricing, access or strategy, can you still operate?

If one model disappears tomorrow, does your institutional capability remain with you?

For GCs and law firm leaders, this is becoming a strategic issue, not a technical footnote.

AI governance cannot be only about privacy policies, acceptable use and vendor questionnaires.

Those matter.

But they are not enough.

The real issue is whether your organisation can use intelligence without giving away the knowledge that makes it different.

That means asking harder questions.

Do we control our prompts, outputs, corrections, memory and evals?

Can our data and usage improve external models?

Can we export what we build?

Can we switch models without losing our institutional knowledge?

Do we have our own definition of what “good” looks like?

Are we creating a private learning loop, or training someone else’s?

This is especially important for legal work.

Because legal AI will not just draft faster.

It will gradually shape how contracts are negotiated, how disputes are assessed, how compliance is monitored, how investigations are run, how regulatory risk is framed and how business decisions are made.

At that point, the question is not whether AI is useful.

Of course it is.

The question is whether the intelligence created through legal work belongs to the institution that created it.

My view is simple.

The next phase of legal AI will be about sovereignty.

Not in the political sense.

In the operational sense.

Control over the data.

Control over the workflows.

Control over the memory.

Control over the evals.

Control over the judgment that compounds over time.

The future will not belong to the legal teams that merely buy the most impressive tools.

It will belong to the teams that know how to use AI without losing ownership of their own intelligence.

Because the hidden cost of AI is not the invoice.

It is waking up one day and realising that the system learned from you.

But not for you.

Stay cool,
Dr. No