Andrew Radosevich, Founder and Managing PrincipalManagement describes an AI capability that processes customer requests in minutes. The demonstration runs smoothly. The investment presentation shows an attractive margin opportunity.
Then someone asks how many requests the system completes without correction. The answer is less certain.
That uncertainty deserves attention. It may reflect a promising capability still under development, an incomplete measurement process, or an assumption that has traveled too far into the financial model.
For a PE sponsor, the concern is the distance between what management claims and what the business can demonstrate. Seven warning signs help identify where deeper work is warranted.
1. The savings calculation starts and ends with hours
A target reports that AI saves employees hundreds of hours each month. Management multiplies those hours by an average compensation rate and presents the result as annual savings.
The calculation estimates the value of capacity released. Whether it represents an earnings improvement depends on what happened next.
Ask which expenses declined, which planned expenditures were avoided, or what additional profitable work the team completed. Include the time employees spend reviewing outputs, correcting mistakes, and handling exceptions.
Evidence to request: A reconciliation from the original workload and cost baseline to the claimed benefit, supported by operating records and finance review.
Without that connection, the figure remains a productivity estimate. It may represent a worthwhile opportunity, but the work required to realize it belongs in the investment case.
2. A successful pilot is described as a company-wide capability
A pilot may perform well with a small group of experienced employees, carefully prepared data, and close attention from the implementation team.
Those conditions can differ substantially from daily operations across locations, customer segments, or acquired businesses.
When management describes a system as "fully deployed," ask what share of eligible transactions actually passes through it, how many employees use it regularly, and how long it has operated at that level.
Evidence to request: Usage records showing eligible volume, actual adoption, completion rates, and exceptions over a representative period.
A functioning pilot can justify further investment. An earnings assumption covering the whole business requires evidence that the capability can operate at the necessary scale.
3. Accuracy figures exclude the difficult work
"Accuracy exceeds 95%" sounds reassuring until the team asks what was measured.
Did the test include unusual requests and incomplete information? Were failed or abandoned tasks counted? Did employees correct the answers before they were scored? Were the test examples also used to develop the system?
A single percentage can also conceal differences in consequence. A formatting error and an incorrect customer commitment should not carry the same weight.
Evidence to request: The test method, representative samples, definitions of success and failure, and results separated by error type and severity.
Review unsuccessful cases alongside successful ones. They reveal where the business needs additional controls, specialist judgment, or a different process.
4. "Proprietary AI" has no clear explanation
A target can create substantial customer value using a commercial AI model. The diligence concern arises when management attributes defensibility to technology it neither owns nor controls.
Ask what is proprietary. The answer might be software, exclusive data rights, specialized testing, customer integrations, or accumulated operating knowledge. Each supports a different argument about competitive advantage.
Evidence to request: A clear account of owned and licensed components, relevant agreements, and the effort required for a competitor to reproduce the capability.
The investment question is whether the differentiation can support customer retention, pricing, or delivery advantages over time. The word "proprietary" needs enough substance to inform that judgment.
5. Possession of data is treated as permission to use it
Customer records may be central to the target's AI story. Yet the company may not have established whether its intended uses are consistent with its agreements and obligations.
An assurance that "the data is ours" leaves important questions unanswered. Where did the information originate? What uses were permitted? What passes to external providers? What restrictions govern retention or reuse?
Evidence to request: Documentation of data sources, permitted uses, access, and vendor handling, reviewed by the appropriate legal and technical specialists.
Unresolved permissions can affect deployment scope, cost, and timing. Where a projected benefit depends on that data, the uncertainty belongs in the underwriting.
6. Human oversight exists only as an assurance
Management says an employee checks every important output. The team cannot explain who performs the review, what the reviewer checks, or how exceptions are recorded.
That leaves both the control and its cost uncertain.
A credible review process gives employees sufficient time, information, and authority to intervene. It includes a route for escalating problems and stopping the workflow when necessary.
The NIST AI Risk Management Framework offers a voluntary reference for assessing how organizations manage risks throughout the design, use, and evaluation of AI systems.
Evidence to request: Review procedures, assigned responsibilities, examples of escalations, incident records, and the labor included in the operating model.
If the projected margin depends on removing manual work, the diligence team should understand precisely how much review remains.
7. The forecast assumes an easy handoff after closing
An AI initiative may depend on one employee, an outside consultant, or a vendor relationship that has received little scrutiny.
Ask who can maintain the system, investigate failures, approve changes, and train new users. Establish whether management has budgeted for ongoing support and whether the responsible people have capacity to deliver.
Evidence to request: Named operating owners, maintenance documentation, support terms, a transition plan, and a budget covering implementation and ongoing operation.
A capability that works today may still require substantial investment to remain reliable as volumes, systems, and personnel change.
What should a sponsor do with these findings?
Each warning sign should lead to a specific diligence action and an investment consequence.
An unsupported savings claim may require a lower earnings assumption. A promising pilot may remain outside the base case until adoption and costs are established. A material data-rights issue may require specialist review before the sponsor can assess the capability's value.
For each finding, document the claim, the missing evidence, the financial exposure, and the work required to resolve it.
Management's response matters. A team that acknowledges limitations and provides access to evidence gives the sponsor a firmer basis for planning. Repeated assurances without supporting records warrant closer attention, particularly when AI is central to the acquisition thesis.
A useful diligence conclusion makes clear what the business can demonstrate today, what still requires investment, and how much confidence the sponsor should place in the projected result.
Common questions
What are the most common red flags in a target's AI claims?
Savings that cannot be reconciled to financial results, pilots presented as company-wide capabilities, accuracy figures that exclude difficult cases, unexplained "proprietary AI," data used without confirmed permission, human oversight that exists only as an assurance, and forecasts that assume an easy handoff after closing.
Why are hours-saved calculations a red flag?
Multiplying hours saved by an average compensation rate estimates the value of capacity released, not an earnings improvement. A sponsor should ask which expenses declined, which planned spending was avoided, or what additional profitable work was completed, net of review and correction time.
How should a sponsor test an AI accuracy claim?
Ask for the test method, representative samples, definitions of success and failure, and results separated by error type and severity. Confirm whether unusual requests and failed tasks were included, whether corrected outputs were scored, and whether test examples were also used to develop the system.
What makes an AI capability genuinely proprietary?
It could be owned software, exclusive data rights, specialized testing, customer integrations, or accumulated operating knowledge. The test is whether the capability supports retention, pricing, or delivery advantages that a competitor could not easily reproduce.
What should a sponsor do when it finds a red flag?
Tie it to a specific diligence action and an investment consequence. For each finding, document the claim, the missing evidence, the financial exposure, and the work required to resolve it. An unsupported claim may lower the earnings assumption or stay outside the base case until evidence improves.