AdapData

Established 2015 Notes

Notes · III Independent review

Technical due diligence for AI and data assets

In short

Technical due diligence is an independent examination of the systems a business depends on, carried out for a buyer, a lender or a board. It covers architecture, code, data, security and the people who maintain them. Where value rests on models and data, the examination extends to provenance, model behaviour over time, dependence on external providers, the cost of producing each output, and who answers for an output that turns out to be wrong.

Note

What is technical due diligence?

Technical due diligence establishes what a buyer is acquiring, what it will cost to keep running, and which of the seller’s claims rest on something the buyer will still own after completion. It is not an award of marks against a maturity model. It is a set of findings, each attached to a consequence for price, for the plan, or for the conditions of completion.

Technical due diligence and technology due diligence are used interchangeably. Where a distinction is drawn, technology due diligence takes in the product and the roadmap, and technical due diligence concentrates on the systems, data and engineering practice beneath them.

Where did the data come from, and may it be used?

Provenance is the first question, because it is the one that cannot be repaired after completion. Records assembled from public sources, from customer activity or under a third-party licence carry different rights, and those rights are not always what the seller believes them to be.

Three matters warrant a written answer. On what terms was each material dataset obtained. Do the customer contracts permit the use to which the data is now put, including the training or tuning of models. And does any material dataset depend on a licence that terminates, or reprices, on a change of control. Where the answer is unclear, the value attributed to the data is unclear with it.

What is the model doing, and how would anyone know?

A model in production is a component whose behaviour shifts with its inputs. The question is not whether it performs well on the evaluation the seller has chosen to show, but whether anyone would notice if it stopped performing well.

Look for an evaluation set that was not selected after the results were known, a record of how performance has moved over the past year, and a defined procedure for what happens when a model is replaced or a provider changes a version beneath it. Where none of that exists, the target does not know how its own product behaves. This is a common finding and a manageable one, provided it is priced.

How much of the system belongs to somebody else?

Most systems now rest on external providers: model interfaces, cloud infrastructure, data feeds and open components. Dependence is not itself a fault. Undocumented dependence is.

The buyer should establish which providers the product cannot operate without, what a plausible price change from each would do to gross margin, and how long a substitution would take in weeks and in named staff. Where the product is a thin layer over a provider that also sells directly to the same customers, the report should say so rather than imply it.

What does the system cost to serve and to maintain?

Cost to serve is a commercial figure with a technical origin. It is the marginal cost of producing one more unit of output, including inference, retrieval, storage and any human review attached to it. In a business priced per seat or per subscription, a rising cost to serve moves margin quietly and is often discovered late.

Maintenance cost is the second figure, and it is measured in people rather than invoices. How many engineers are required simply to keep the present system working, and how much of their time is absorbed by defects arising from decisions taken years earlier.

Who answers for the output?

Where a system produces an output that a customer relies upon, somebody must be accountable for it. The examination establishes who reviews what, on what basis, and what record survives that review. It also establishes what the target has told its customers, because commitments about accuracy, human oversight or data handling become the buyer’s commitments at completion.

This is the point at which technical and legal diligence meet. The two workstreams should be made to compare notes rather than report in parallel.

Questions

What is technical due diligence?
It is an independent examination of the systems, code, data, security and engineering organisation that a target depends on, conducted for a buyer, a lender or a board. It establishes what is being acquired, what it costs to operate, and which claims rest on assets that actually transfer.
What is technology due diligence in M&A?
The examination of a target's technology as part of a transaction: architecture, code, data, security, engineering practice, and the cost of operating what is being bought. It runs alongside the commercial and financial workstreams and reports to the same timetable.
How does technical due diligence differ from technology due diligence?
The terms are largely interchangeable. Where a distinction is drawn, technology due diligence includes the product and roadmap, while technical due diligence concentrates on the systems, data and engineering practice beneath them.
How long does technical due diligence take?
Two to four weeks is usual for a mid-market target, running alongside the commercial and financial workstreams. Unscripted access to a senior engineer matters more to the quality of the work than the length of the window.
What are the most common findings on an AI-based target?
Unclear data rights, no record of model performance over time, undocumented dependence on a single external provider, and a cost to serve that has never been calculated for a single unit of output.
Is access to source code necessary?
Read access to representative repositories is preferable and is normally granted under a restricted arrangement. Where it is refused, the examination proceeds on documents and interviews, and the report states the limitation.
What belongs on a technical due diligence checklist?
Architecture and its single points of failure; code ownership and licence position; data provenance and permitted use; model evaluation and monitoring; external provider dependence; security and access control; cost to serve per unit of output; and engineering headcount, attrition and key-person risk.
How is artificial intelligence used in due diligence itself?
Chiefly for document review and extraction in data rooms, where it improves coverage of large volumes within a short window. It does not supply judgement, and its output requires verification against the source. Any finding that bears on price should be traced to a document a person has read.

Engagement

We are usually engaged by an investor, a board or an owner who requires a disinterested account of a system rather than confirmation of a decision already reached. Technical origination and examination are carried out with our research partner. Engagements are senior-led and held in strict confidence. We reply personally to every enquiry; write to [email protected].

Related