The hidden cost of AI: human oversight

An AI investment thesis can depend on faster delivery, lower service costs or greater capacity. Each claim needs evidence from the workflow it is meant to improve. What happens between a model producing an answer and the business delivering an acceptable result?

From my diligence work

Buy-side diligence for a $4B AUM fund

My role

Product & AI advisor. Data-room review, operating meetings and assessment of the team behind a proposed acquisition.

Finding

The plan underestimated human oversight costs.

Decision

The deal was stopped following product and AI diligence.

The fund and transaction remain anonymous in the public case.

The lesson I would take into another assessment is straightforward: include the people needed to make AI output usable in the economics of the proposed operating model.

Follow the work through to completion

Choose a recurring workflow that materially affects the thesis. Map the current process, including its volume, people, delivery time, quality standard and exceptions. Then map the proposed process against the same scope.

For each step, ask who produces the output, who checks it and who handles a failure. Include the cases that require specialist judgement or a return to the original process. A sample of easy cases cannot establish the cost of the full workload.

Review time also needs a denominator. Is it measured per generated answer, per customer request or per successfully completed job? If several attempts are needed to complete one job, counting only the final attempt understates the work.

Use an agreed quality threshold for both processes. A faster draft that needs more correction may still be useful, but the assessment should show the complete result.

A framework for assessment

Measure the accepted job

1. Model output

   Provider usage across every attempt.

   →
2. Human review

   Time spent checking the result.

   →
3. Correction & fallback

   Rework, escalation and the original process when needed.

   →
4. Accepted job

   A result meeting the agreed quality threshold.

Recurring cost per accepted job

Provider & infrastructure + delivery labour + support divided by Accepted jobs in the same period

What to request in the data room

Request evidence for the complete operating process:

Current workflow volumes, time records and delivery costs

The baseline and which costs the thesis expects to change.

A representative evaluation sample, with selection method and quality criteria

Whether measured performance covers the work the business actually receives.

Review, correction, retry and escalation records

The human work required to reach an acceptable result.

Provider usage, infrastructure costs and support effort

The operating costs beyond generating the first output.

Rollout assumptions, staffing plan and accountable owner

How an observed improvement could become an operating result.

Record the source and period for each number. Label missing inputs as unknown. A useful diligence conclusion can identify which assumptions remain unproven without replacing them with a confident estimate.

Separate capacity from cash savings

Hours freed create capacity. A claim of lower expenditure needs a separate explanation of which spending changes, when it changes and who owns that decision. The business might use the capacity to serve more customers, reduce overtime or avoid an additional hire. Those are different outcomes and require different evidence.

A practical cost comparison should include:

• Remaining delivery work, including human review, correction and escalation.
• AI provider and infrastructure costs, including retries and fallback paths.
• Ongoing support, evaluation and maintenance.
• Implementation, integration and training costs, shown separately from recurring costs.

Use the same volume, period and quality threshold for the comparison. Keep categories mutually exclusive so that review work or infrastructure is not counted twice. Show assumptions that could change the result, such as workload mix, review effort, adoption and provider costs.

BCG makes a related point in its analysis of AI in private equity: turning productivity improvements into P&L impact requires changes to how the organisation operates. That supports examining the rollout plan alongside the technology.

Turn findings into decisions

For each material assumption, write down what is supported, what remains unknown and what evidence would resolve the uncertainty. Identify the owner and the decision that depends on it.

Where the acquisition proceeds, those findings can inform post-close priorities: establish the baseline, test a representative workflow, measure complete delivery costs and agree the conditions for expanding use. The pace should follow the evidence and the consequences of failure.

The output of product and AI diligence should help the deal team understand what it is buying and what must happen for the AI plan to work.

EY-Parthenon's due diligence framework likewise considers AI across commercial, technical and financial dimensions, including its use in operations.

Five questions before you underwrite AI savings

1. What counts as an accepted job?

   Agree the quality threshold. Walk through a normal case and an exception with the team responsible for delivery.

2. What does the complete result cost?

   Include review, correction, retries, fallback and support. Count unsuccessful attempts against the same period and workload as the baseline.

3. Who can accept the result or stop the process?

   Name the operating owner, the acceptance authority and the person responsible for exceptions.

4. Which operating decision turns capacity into value?

   Specify whether the plan changes throughput, delivery time or a particular expense. Identify what must change, who owns it and when.

5. Which assumptions remain unproven?

   Record the missing evidence, its owner and the investment or rollout decision that depends on it.