Practical guide · Genitechs
Which process should you automate with AI?
A useful first candidate is frequent, narrowly defined and measurable, with errors that can be caught before they cause harm. AI is useful only where conventional rules are insufficient.
Published October 4, 2026
Start with an observable problem
Choose a task for which your team can describe the inputs, steps and expected output. “Use AI” is a possible method, not an operational need. “Reduce the time spent sorting incoming email requests” gives you a specific starting point.
Record workload, turnaround time, errors and exceptions over a representative period. Separate processing time from waiting time: automating five minutes of data entry may not solve a week spent waiting for approval.
Choose between rules, integration and AI
| Situation | Approach to consider | What to validate |
|---|---|---|
| Structured data and stable rules | Rule-based automation | Exceptions and access rights |
| The same information copied between tools | System integration | APIs, identifiers and the system of record |
| Variable text or document formats | AI assistance with human review | Output quality and error handling |
| Sensitive decisions or unavailable data | Clarify requirements before a pilot | Accountability and authorised data |
Check data and exceptions
Identify where data lives, who can authorise its use and which information must not be shared with a provider. Running a pilot does not remove this requirement.
Test difficult cases: incomplete documents, ambiguous text, duplicates, contradictory information and unexpected formats. Provide a manual path when the system cannot process a case correctly.
Measure net improvement, not just speed
Compare a baseline and a pilot under similar conditions. Measure total time, including review, corrections and follow-up. Also track errors that escape review.
A simple estimate of monthly time saved is: processed volume × (average manual time − average assisted time, including review). Compare the value of that time with implementation, usage, maintenance and oversight costs. This calculation is a working hypothesis, not a return-on-investment promise.
Agree when to proceed, revise or stop
Set acceptance criteria before the pilot: minimum quality, acceptable review time, unacceptable error types and the person responsible for deciding. Define a fallback.
A pilot may show that AI is unsuitable or that a simpler integration is sufficient. That conclusion can prevent a poorly directed investment.
Information to bring to a discussion
- The task and its owner.
- Current volume, frequency and turnaround time.
- Tools involved and how they exchange information.
- Representative examples, anonymised where necessary.
- Possible errors and their consequences.
- The expected outcome and how it will be checked.