When AI automation pays for itself

AI automation is justified when the process repeats at least a few hundred times a month, the result can be checked objectively, an occasional error causes no irreversible damage, and the required data already exists in an accessible form. If one of those conditions is missing, the cost usually exceeds the benefit.

Criterion 1: volume

A process that repeats ten times a month does not justify the effort of building, monitoring and maintaining it. Our rough threshold is a few hundred cases monthly, or a rarer process that consumes expensive specialist time.

Criterion 2: verifiability

There has to be an objective way to say whether the result is correct. Extracting an invoice number is verifiable; "wrote good copy" is not. Without a measurable criterion, you cannot tell whether the system is degrading over time.

Criterion 3: the cost of an error

If a mistake means a five-minute correction, automation is reasonable. If it means a wrong payment, a contractual obligation or legal exposure, you need mandatory human review, which changes the economics.

Criterion 4: data availability

Models need context. If the relevant information lives in three people's heads and in private messages, the first project is not automation but organising the data.

Processes we do not automate

• Commercial decisions: pricing, discounts, contract approval.

• Sensitive communication: serious complaints, legal matters, human resources.

• Any flow where we cannot reconstruct afterwards why a decision was made.

• Processes that change monthly: you automate them faster than you can stabilise them.

How we calculate the return

1. Measure the current time per case and the number of cases per month.

2. Estimate the realistic share of cases the system can handle without intervention.

3. Add up build cost, monthly model cost and the review time that remains.

4. Compare over a 12-month horizon, not a single month.

5. If payback takes longer than a year, we recommend not doing it.

Start small

The best first project is narrow, high-volume and low-risk: extracting data from a single document type, or classifying inbound requests. The outcome gives you real figures for the next decision.