Your AI Stack Has an IP Line Item Now

Most AI business cases still look the same.

They have a model line. A cloud line. Sometimes a headcount line. Occasionally a vague "productivity" upside that nobody can audit.

What they usually miss is a line that just became impossible to ignore: training-data and copyright liability.

Technology is easy. Valuation is hard.

The missing line item

When leadership approves an AI investment, it typically stresses three questions:

1. Does the model work in a demo?
2. Can we integrate it?
3. What does inference cost?

Those are necessary. They are not sufficient.

A serious valuation also asks:

• What data trained the systems we depend on?
• Who carries residual IP risk if that data was dirty?
• What happens to our process if the vendor's economics or legal posture change?

If those questions are not on the page, you do not have an AI valuation. You have a hope sheet.

What a nine-figure settlement signals to buyers

In 2026, public reporting on Anthropic's copyright settlement put a $1.5 billion figure in front of boards and markets - widely described as among the largest US copyright class outcomes. You do not need to litigate Silicon Valley from your boardroom. You need to treat the signal correctly.

The signal is not "AI is illegal."

The signal is:

• Training data has a price - explicit or latent.
• Vendor cost structures can reprice after legal events.
• Precedent emboldens rights holders and changes diligence norms.

For enterprise buyers, that converts into plain questions:

• Are we buying a product, or a balance-sheet contingency?
• Does the contract indemnify training-data claims - or only forward use of outputs?
• If the vendor's unit economics jump, do we have a kill criterion and an exit path?

Ignore those, and your "AI transformation" multiple is fiction.

Execution got cheap. Liability did not.

There is a second shift happening in parallel.

Frontier models make software execution dramatically cheaper. The old moat - "we can code" - is eroding for productized software and internal tools. Distribution, proprietary data, and operational finish lines matter more than ever.

That does not mean enterprise AI is free.

It means the scarce assets moved:

Boards that celebrate cheaper demos while ignoring liability are optimizing the wrong column.

Open source vs closed is a portfolio choice

The public debate often frames open-source AI as ideology.

For operators, it is a portfolio decision:

• Closed vendors: commercial support, product polish, contractual surface - and concentration risk.
• Open models: control and optionality - and operational burden, security, and still-unresolved rights questions depending on stack and jurisdiction.

Neither choice removes the enterprise from residual risk.

If a model sits inside a process that decides credit, medical workflow, industrial control, or customer rights, you own the outcome. Vendor marketing does not.

So the board question is not "open or closed forever?"

It is:

• Which workloads justify which stack?
• What is the residual IP and compliance risk in each path?
• What is our switch cost if the bet fails?

That is valuation language. Use it.

First-principles questions before scale, pause, or stop

Before the next AI budget leaves the room, demand written answers:

1. Economic claim - If this works, what cash, cost, risk, or cycle-time number moves - and by when?
2. Finish line - What does "shipped" mean in one sentence for a real user or process?
3. Kill criterion - What evidence forces pause or stop within 90 days?
4. Data provenance - What do we know about training-data rights for systems we scale?
5. Indemnity and residual risk - Who pays if rights holders succeed - vendor, insurer, or us?
6. Concentration - How many mission-critical flows depend on one model family or one vendor?
7. Exit - Can we reverse the process without theater?

If (1)-(3) are weak, you have a demo problem.

If (4)-(7) are weak, you have a valuation problem.

A practical framework: four columns

I use a simple board sheet for AI initiatives:

Most organizations overweight the first column and underweight the last two.

That is how you get impressive pilot theater and fragile production systems.

What good looks like

A leadership team that takes this seriously does not ban AI.

It does the opposite of hype:

• Fewer starts, clearer finish lines
• Vendor diligence that includes legal and economic stress tests
• Portfolio thinking: open and closed where each earns its keep
• Explicit non-goals so capital is not diluted across twenty half-alive pilots

Purposeful AI is not "more models."

It is investable systems with economics, rights, and accountability that survive board scrutiny.

How to act this quarter

1. List AI systems already in or near production.
2. Score each on the four columns above - red/amber/green is enough.
3. Stop or fence anything red on IP/provenance without a mitigation plan.
4. Rewrite the next business case template to include the IP line item by default.
5. If a bet is material, run a structured valuation review before scale capital.

If you want a structured outside view, that is exactly what a short Tech Valuation Audit or strategy conversation is for: decide scale, pause, or stop with criteria leadership can defend.

Technology will keep getting easier.

Valuation - including IP - remains the work.