Which work is worth automating? Calculate the whole effort

The most repetitive task is not always the best first automation. A frequent activity with unstable inputs, unclear ownership and expensive mistakes can require more supervision than expected. The useful starting point is a specific process with measurable effort and an agreed definition of a correct result.

GSA's Eliminate, Optimize, Automate approach places removal of unnecessary work and simplification before automation. That ordering is valuable even for a small team: do not pay to run a step faster before deciding whether the step is needed. GSA: Eliminate, Optimize, Automate.

Describe the job before choosing a tool

Write down the trigger, inputs, decision rules, output, owner and exceptions. “Process invoices” is too broad. “Extract the supplier reference and amount into a review queue, without approving payment” gives a bounded pilot with a clear human decision.

Observe actual cases, including awkward ones. Count missing fields, duplicate submissions, corrections and format changes. If the rule only exists in one experienced person's head, first make that rule explicit. Automating an unexplained judgement can multiply inconsistent decisions.

A worked business case

Assume a fictional team handles 300 records each month. Manual preparation takes four minutes per record: 1,200 minutes, or 20 hours.

The proposed automation leaves one minute of review per record, adding up to five hours. Fifteen exception cases need six additional minutes each, or 1.5 hours. Monitoring and maintenance require another two hours monthly. Total recurring effort is therefore 8.5 hours, and net time released is 11.5 hours.

At an illustrative capacity value of EUR 30 per hour, that time represents EUR 345 per month. Subtract EUR 45 of monthly tool costs to get EUR 300 of estimated net monthly value. If setup costs EUR 1,800, simple payback is six months under these assumptions.

This is not automatically EUR 300 of cash savings. If staffing expenditure stays unchanged, the main benefit may be capacity for other work. State how that capacity will be used. The calculation also excludes taxes, financing and discounting and assumes the measured quality remains acceptable.

Test whether the case survives lower volume

At 150 records per month, the manual baseline becomes ten hours. If review and exception effort halve but maintenance remains two hours, the new workload is 2.5 + 0.75 + 2 = 5.25 hours. Released capacity is 4.75 hours, worth EUR 142.50 under the same assumption; after the tool fee, net monthly value is EUR 97.50.

The same setup now has simple payback of about 18.5 months. Fixed maintenance changes the result substantially. Test a realistic low-volume case and a high-exception case before treating the optimistic scenario as the business case.

The value uses EUR 30 per hour of capacity; it is not automatically a reduction in payroll or cash expenditure.

Include the cost of being wrong

For a reading list, an incorrect classification may be easy to repair. For payment, access rights or contractual communication, a mistake can have much larger consequences. Choose the review and approval boundary according to the impact, not only the confidence score shown by a tool.

Where AI extracts or classifies content, retain the source and let the reviewer compare the proposed result with it. Use examples outside the development sample for evaluation. “It worked on the ten examples we used to build it” gives limited evidence about future inputs.

Measure correction effort and missed errors, not only processing speed. A workflow that saves five minutes and creates twenty minutes of reconciliation is not an improvement.

Run a controlled pilot

• Use representative records with permission to process their data.
• Compare proposed outputs with the current method before enabling consequential actions.
• Define what enters the exception queue and who clears it.
• Prevent duplicate actions when a job is retried.
• Keep a manual fallback and a way to stop the process.
• Assign ownership for changed input formats, expired access and supplier changes.

Make failure visible. A silent run producing no output should not look identical to a successful run with nothing to process. The operator needs a clear status and enough information to investigate without exposing unnecessary sensitive data.

Decide using observed results

After the pilot, replace assumptions with measured volume, review time, exceptions, maintenance and quality results. Decide whether to expand, simplify further or stop. Stopping a weak automation after a small trial is a useful outcome.

Review again when the process changes. An automation is an ongoing operational commitment; its value depends on reliable output and sustained net benefit, not on the number of steps removed from the screen. All figures are fictional teaching examples.