Hidden AI risks in a deal: data debt, vendor dependence, the AI Act

AI risk as a line item in the deal model

Most AI risks are discussed in technical language, which is exactly why they end up outside the deal model. That's a mistake. Each of them has a price, a probability and an owner after close — everything you need to put it into the valuation. The question isn't "is this system well built," but "what does it cost me that it isn't, and when does that cost materialize."

Three categories survive the close regardless of how good the demo looked: data provenance, vendor dependence and regulatory exposure. The rest is usually operating cost that can be fixed in the first year.

Data debt: a liability you buy as a package

Training data with no documented provenance is the most common hidden liability. A model trained on scraped content, on customer data used in breach of contract, or on datasets sourced from "somewhere" transfers the risk to the buyer in full. Economically, it's the twin of an undisclosed tax liability: invisible in the demo, surfacing at the first audit or lawsuit.

On top of that sits data privacy. If personal data went into training, or leaks in the model's responses, you have a compliance problem you can't remove without retraining. Ask for the legal basis of every dataset and whether it can be proven. Treat the absence of proof the way you'd treat a missing legal title.

Vendor dependence: a margin sensitivity

A company built entirely on a single external model has its margin tied to a price list it doesn't negotiate. This isn't a theoretical risk — vendors change prices, limits and terms, and sometimes retire models. Without an abstraction layer over the model, migration means rewriting and re-evaluating the whole system.

Price it with two numbers: the cost of migrating to an alternative and the impact of a realistic price increase on EBITDA. If both are high and there's no exit plan, lock-in is a line item in the model, not a footnote.

AI Act exposure and governance

The AI Act ties obligations to the use case, not to who trained the model. A company using someone else's model in a high-risk area has its own documentation, oversight and disclosure obligations. Failing to meet them is an exposure that grows as enforcement tightens.

A signal of maturity is AI governance: who is accountable for the system, how the model's decisions are logged, how human oversight works over high-stakes cases. A company without it isn't so much breaking the rules as unable to prove it isn't — and in an audit, that's the same cost line.

The attack surface nobody talks about

AI systems carry security risks that classic software doesn't. The most common is prompt injection — crafted input that hijacks the model's behavior, for instance coaxing it into revealing data or carrying out an unintended action. If the model has access to tools, email or a database, this isn't a curiosity but a real leak vector.

The line of defense is guardrails: input and output controls, restricted model permissions, a human approving high-stakes actions. Their absence doesn't raise a single risk — it turns an incident from a one-off cost into a recurring liability, because every new integration opens a fresh hole.

How to frame it in the decision

The board doesn't need a technical report, just a table that ties each risk to a number.

Risk	What materializes it	How to price it
Data debt	Audit, lawsuit, deletion request	Cost of retraining plus legal exposure
Vendor lock-in	Price increase, model retirement	Cost of migration plus EBITDA sensitivity
AI Act exposure	Enforcement, regulator audit	Cost of closing governance and documentation gaps
Prompt injection	Data leak, unintended action	Cost of deploying guardrails plus incident risk

Operator's rule: an AI risk you can't attach to an amount and an owner after close hasn't gone away — you simply haven't priced it yet. Each of these line items either lowers the price or goes into the first-100-days plan.