The AI produces a result in seconds. A teammate then spends several minutes checking it, correcting it, and moving it into the right system. The demonstration looked fast; the working day does not feel much shorter.
Measure the complete task, including review and recovery. That tells you whether to improve the automation, narrow its responsibilities, or keep a different process.
Count time until the result is usable
Here is an invented example. A team completes 100 tasks a week, taking six minutes each: 600 minutes. With AI, preparing and handling each task takes one minute, and reviewing each result takes three more. Twenty tasks also need five minutes of correction.
The new total is 100 + 300 + 100 = 500 minutes. That saves 100 minutes before maintenance, interruptions, and other overhead. If maintaining the workflow takes two hours a week, it currently consumes more staff time overall.
This calculation does not settle every business decision. Faster responses, improved consistency, or extra capacity may still matter. It does show why model response speed alone is a weak measure of operational value.
Find out why review takes so long
Watch someone check a real, appropriately authorized example. Are they hunting for the source material? Reformatting the output? Recalculating numbers? Correcting missing fields? Deciding whether the result should have been generated at all?
These are different problems. Showing the supporting source beside a proposed answer can reduce searching. A fixed output structure can reduce copying. Required-field and arithmetic checks can catch specific errors. None of those checks establishes that every claim is true.
For a hypothetical invoice workflow, checking that totals add up is useful. It does not establish that the invoice belongs to the right supplier or that payment is authorized. Keep those business decisions explicit.
Give automation a narrower responsibility
A workflow may be useful when it drafts a response for approval, even if it cannot safely send that response automatically. It may reliably process one document format while routing unfamiliar formats to a person.
Define the conditions for each path and assess them with representative examples. An AI model saying it is confident is not, by itself, evidence that the result should bypass review.
Your evaluation should include supported answers, missing information, unusual inputs, and cases where the correct action is to ask for help. Anthropic's evaluation guidance describes defining measurable success criteria and building evaluations around them. Apply those criteria to the whole workflow, including the person's remaining work.
Measure a business result and assign an owner
Track a small set of outcomes consistently:
- Completed tasks whose results meet the agreed standard.
- Total staff time, including review, corrections, and maintenance.
- Errors that escaped review and the consequences of those errors.
- Unresolved exceptions and how long they wait for an owner.
A green execution status only tells you that the software reached its defined endpoint. If a required business condition is missing, the workflow needs to recognize that and route it appropriately. Tools such as n8n's error handling can support this, but the business rules still need to be designed.
Compare the revised workflow against the current process on similar tasks. Keep, narrow, or stop the automation based on useful outcomes and operating effort.
Five prompts to use with your AI coding assistant
Before you begin: Use the same project conversation for all five prompts. Inspect first, approve the plan yourself, and never test against live customer data or secrets.
These prompts assume your assistant can inspect the relevant project. Ask it to say clearly if it cannot. Use the prompts one at a time, reviewing each response before moving to the next step.
1. Map
Map the end-to-end workflow, including AI decisions, ordinary rules, human review, retries, corrections, and final business outcomes. Distinguish automated actions from tasks truly completed.
2. Audit
Identify where people spend time checking, fixing, or finishing AI work. Define measurements using real authorized operational data where available; do not invent savings.
3. Plan
Propose a narrower or redesigned workflow with deterministic rules, validated selective review, exception handling, and clear ownership. Specify where human decisions must remain.
4. Challenge
Compare the plan with a baseline of staff time, errors, rework, operating cost, and completed tasks. Test plausible failure cases and identify where automation might add work.
5. Implement
After approval, implement a limited pilot and measurement dashboard in a safe environment. Report observed time saved, remaining human work, and a recommendation to expand, revise, or discontinue.
DitchNow can help with measurement and iteration and practical team enablement. Bring one repetitive process, including the checking work that happens after the AI finishes.
About DitchNow
DitchNow is based in Haarlem, Netherlands, and helps founders and teams fix and extend software and put AI to work. Meet DitchNow.