Pick a problem that matters, and check the non-AI fix first
Updated
BeginnerWhat you'll learn
~13 min- Compare a candidate problem against the four-part filter: frequent, slow, bounded, and someone knows what good looks like
- Ask the people who do the work to name the delay before any tool is chosen
- Decide whether fixing the workflow without AI is the better first move, and record why
You leave this lesson with: a one-line problem statement that passes the four-part filter and names the users who helped write it.
Four questions that pick the problem
A demonstration shows that a tool can do something. It does not show that the something is worth doing. Four questions help a leader choose where to spend the first pilot.
Frequent. Repeated work lets a small gain accumulate and gives more cases to compare. For a first pilot, prefer work that happens often enough to evaluate within the review period. Frequency alone does not establish value.
Slow. The work takes real elapsed time, counted from request to finished result, including the waiting. Ask for the calendar days, not the hours of effort.
Bounded. The work has edges: a defined input, a defined output, a defined set of people and data. A bounded problem can be tried, measured and stopped. An unbounded one cannot.
Someone knows what good looks like. A named person can look at a result and say whether it is right. Without that person, nobody can tell a faster result from a worse one.
A problem that fails any one of the four is not ruled out forever. It is a poor first pilot.
Ask the people who do the work
The people who will use the result know where the time goes. Ask them before choosing a tool, and ask them again after the first demonstration.
Take a university research group whose core facility takes in sample-analysis requests from other labs. The facility director assumes the delay is in the analysis. Staff who handle the requests say the delay is earlier: requests arrive missing a required field, and each one goes back and forth by email.
Neither view is careless. The director sees the part of the process that costs the most staff time. The staff see the part that costs the most calendar time. A pilot aimed at the analysis would have optimized the wrong step.
Two questions to put to users: where does a request wait the longest, and what do you redo most often. Write down the answers. They tell you where to measure, and they give you the problem statement in one line, for example: “Sample requests from other labs wait days for a missing form field; the work is frequent, slow and bounded, and the intake lead can judge a complete request. Written with the intake staff.”
What the baseline records
A baseline is a recorded measure of the current workflow before the change, with its source and the period it covers. Record three things for the chosen workflow:
- Elapsed time from request to finished result, including the waiting.
- Error and rework rate: how often a result is corrected or sent back.
- Cost: the staff time and spending the work takes today.
Give each figure its source: a system export, a timed sample of the cases that arrive before the pilot starts, or an estimate labelled as an estimate. State the period each figure covers.
If the team cannot produce a baseline, that is a finding, not a dead end. Ask them to record the current workflow over an agreed observation period before the funding decision, and record the comparison as unavailable until then. Recording today’s process can also show which step is actually slowing the work down before you pick any tool. AI Assurance (Lesson 5.1, a separate training in this series) shows how an assurance team writes that baseline. An assurance team checks that a system is safe and works as claimed before release.
A pilot has its own cost. Estimate its staff hours, for collecting the baseline, writing the pilot’s plan (Lesson 5.1: five written lines, one owner), reviews, training and support, and set that estimate beside the saving the pilot is expected to deliver.
Check the non-AI fix first
Before asking whether AI could help, ask whether the workflow can be repaired without it. Check missing form fields, hand-offs between inboxes and conflicting report versions. If nobody owns an approval’s queue, name someone to resolve requests while keeping what the approval protects (Lesson 3.1: every queue needs a named owner).
Compare the cost and effort of that repair with the proposed AI pilot. If the repair is feasible, measure the remaining delay before deciding whether a pilot is worth funding.
Fixing the workflow first does not mean AI is rejected. It means the leader can say, after the fix, which delay remains and whether it meets the four-part filter.
A university core facility has a sample-request backlog, and a vendor has offered a demonstration of an AI assistant that reads and sorts incoming requests. Staff say most requests wait because the form lacks a required field, so each one needs an email exchange before work can start.
- Check recent requests to confirm the missing field is the cause, then add the missing field and a single named inbox to the form, measure the calendar days from request to start for the next several weeks, then check the four-part filter again.
- Approve a pilot of the AI assistant on all incoming requests.
- Leave the form as it is for now, and put staff time into clearing the current backlog by hand.
Pick one before you open What happens.
ℹWhat happensclick to expand
- Check requests, add the missing field: It costs a form change and several weeks of measuring, shows how much of the wait was the form, and leaves a measured baseline.
- Approve a pilot of the AI assistant: It starts quickly and the demonstration may look good, but the pilot would be measured against a process that is slower than it needs to be, so any gain is unclear and the form problem stays.
- Clear the backlog by hand: It clears the requests already waiting, which helps the labs in the queue now, but the next requests arrive with the same missing field.
Choice 1 holds up, because it fixes a known cause before paying to work around it. If the wait is still long after the fix and the remaining work is frequent, bounded and has someone who can judge a good sort, an AI pilot is then a fair next step.
A program team proposes an AI pilot for a monthly report that takes two staff a week to assemble. No one has checked why it takes a week. What is the first move?
Key takeaway
- LEADERSHIP DECISION
- Choose the first pilot by the four-part filter and the users' account of the delay, not by the demo
- IMMEDIATE ACTION
- Ask the team which non-AI fix to the workflow they would try first, and what it would cost
- SUCCESS MEASURE
- Calendar days from request to finished result, against the baseline recorded before the change, with rework rate no higher