Why AI workflows stall after the demo
From a convincing demo to a report your team can trust. Six practical checks, shown through a weekly trading-report example.
The AI demo produces a polished trading report. On Monday, the finance manager rebuilds it by hand.
One export is late. A category has changed. The commentary sounds convincing, but nobody can trace the explanation to the numbers.
This hypothetical example gets to the real test: does the workflow make the whole job easier, including the checking?
1. Map the work people actually do
Walk through the last completed report with its preparer. Look at the source files, corrections and awkward handovers. Agree which data is authoritative, who supplies it and what happens when it is missing.
Question the steps, too. If three people reformat the same information, agree one useful output before automating the duplication.
2. Give AI a clear role
Use fixed rules for agreed calculations and checks. Let AI draft from verified figures and supporting notes. Keep judgement and approval with someone who understands the work.
- 01 / RulesPrepare & check
Collect inputs, calculate movements, flag gaps.
- 02 / AIDraft & explain
Use checked figures. Show sources and uncertainty.
- 03 / PeopleReview & approve
Challenge the explanation. Decide what to do.
A rise in sales is a fact to check. Why it happened needs evidence.
Permission to prepare a draft is separate from permission to change records or send the report. Agree access with the people responsible for the data and systems.
3. Test a difficult Monday
Choose approved historical or synthetic examples that reflect real problems. Define the expected response before running each test.
Illustrative responses for this report. Agree the rules for your own process.
4. Measure the accepted output
Count preparation, chasing, checking, corrections and distribution before and during the trial. Track errors and whether the report arrives when it is needed.
A quick draft is no saving if somebody has to recreate it to trust it. Source references and visible exceptions should help the reviewer check the work without starting again.
5. Give it an owner and a fallback
Name someone who can judge the output, resolve exceptions and pause the process. Give them time and support. Rehearse a normal run and a failed one with the preparer and reviewer.
If old and new methods run side by side, agree when that comparison ends. Otherwise, the team inherits two permanent jobs.
Before handover, document who maintains the workflow, handles changes and pays for software. Keep the test cases for future model or system updates, and a safe way to complete urgent work if the workflow fails. Agree any external support separately.
6. Expand only when the evidence supports it
Can the team trace the inputs, spot an incomplete result and review the commentary efficiently? Does the report arrive reliably with less overall effort?
- ExpandThe process works
Extend the scope carefully. Keep the controls.
- ReviseThe gaps are fixable
Improve the weak step, then test again.
- StopThe case does not hold
Return to the safe process. Keep the learning.
Clerq's AI opportunity assessment examines one process and agreed sample files. You receive a process map, bottlenecks, an assessment of AI and automation fit, one recommended improvement, and an outline build scope with estimated cost.
Implementation is agreed separately. Start with a free introductory conversation about the work your team keeps repeating.
Original hypothetical example, developing themes from Alex Lieberman's AI-transformation discussion. Further reading: the voluntary NIST AI Risk Management Framework.
Want this built in your business?
This is the kind of system Clerq builds inside UK businesses - in your tools, around your controls, handed over working. Tell us what is slowing you down and the founder will reply personally. See our published service prices. Scope and price are agreed before work starts, with no obligation to proceed.