Value thesis
Which operating decision or workflow should change?
Define who uses the workflow today and what the current baseline is, then the economic mechanism and the evidence it would take to believe AI can change the outcome.
Private equity field guide
AI due diligence for private equity
Test whether an AI value-creation thesis can survive contact with the target company's workflows and data, and with its operating reality.
The diligence standard
Move from possible use cases to testable operating claims.
A value thesis names a specific workflow and the mechanism that creates the value. Diligence tests that thesis against the evidence available before close, and against the constraints that could prevent deployment.
The right output is a set of investment and operating decisions. Which claims are credible now, and which stay conditional? The first 100 days get a plan, and the underwriting gets an honest list of risks.
The framework
Five diligence lenses.
Data and knowledge
Does the workflow have usable, permitted context?
Identify the systems of record and the unstructured knowledge around them. Access constraints and data quality get checked here, and so do retention requirements and subject-matter owners.
Systems and integration
Can the system act inside the real process?
Map the operational systems a production workflow has to use, with their APIs and permissions, and trace how review steps and exception paths fit in.
Evaluation and reliability
How will the team know the workflow is safe and useful?
Specify representative tasks and quality measures, then work out the failure modes and escalation rules, with monitoring and the evidence bar set before rollout.
Ownership and governance
Who can operate and improve the system after close?
Assess engineering capacity and workflow ownership, and how security review happens. Model and vendor decisions need an owner too, along with change management and executive accountability.
Evidence to request
Inspect the workflow behind the roadmap.
- A walkthrough of the current workflow, including exceptions and manual handoffs
- System-of-record and data-access map
- Representative documents and inputs, with the expected outputs
- Existing AI pilots and vendors, with usage data and evaluation results
- Security and privacy requirements, plus retention and human-review rules
- Who owns the workflow in the business, and who owns it in engineering and operations
- A current baseline for cycle time and quality, or for cost and capacity
Red flags
Treat missing evidence as a risk.
A model-first thesis
The plan begins with a model or chatbot and cannot name the operating workflow it should improve, or who owns it.
No production scorecard
The team judges a polished demo by feel, with no representative tasks or failure modes defined, and no decision gates.
Unavailable context
The workflow depends on fragmented systems and restricted data, or on unwritten knowledge, with no feasible plan for access and ownership.
Automation without review design
The thesis assumes full automation where judgment and regulation call for human review, or where the customer impact does.
No post-close owner
The company has a pilot sponsor and no team accountable for running the system or improving it, security included.
Universal ROI claims
Benefits are borrowed from broad market statistics. A credible claim ties a baseline and a mechanism to the target company's own workflow.
Decision-ready output
Connect pre-close evidence to the first 100 days.
Findings should be explicit about confidence and the access available during diligence. Unverified assumptions remain open risks until the company can test them.
- A ranked workflow and opportunity map
- Technical feasibility and access constraints
- Evaluation and security requirements, with governance spelled out
- Build, buy, or partner decision questions
- A first-100-days implementation sequence
- Clear evidence gaps and investment-committee risks
Common questions
Keep conclusions proportional to the evidence.
What is AI due diligence in a private-equity transaction?
AI due diligence tests whether an AI-related value thesis can work inside the target company's workflows and data, under its systems and controls. It should produce evidence and a post-close plan the deal team can act on.
How is AI due diligence different from general technology diligence?
General technology diligence covers architecture and security, along with the team and its technical debt. AI diligence adds task-level evaluation and model behavior, plus how human review is designed and what the system costs to run once the workflow adopts it.
Should diligence recommend building or buying?
It should identify the decision criteria. A standard capability points toward buying, and a differentiated workflow can justify a custom build. The answer turns on data access and integration, and on control requirements and internal ownership.
Can diligence be completed before full data access?
A preliminary assessment can identify hypotheses and evidence gaps. Label the confidence clearly, and keep material conclusions conditional until representative data and systems can be examined, with operators in the room.
Next step
Pressure-test an AI thesis before it becomes the operating plan.
Bring the thesis and the workflow, along with what you can access during diligence.
Discuss a deployment