In its sixth month, a well-kept workflow is better than the one handed over. It has met cases nobody thought to test, taken in new rules, and still passes every test it passed on day one, plus the ones added since. That is how 10X the ROI of your existing AI keeps growing after the build.
Who owns an AI workflow after the pilot? names the roles behind a workflow, and the rollout checklist covers its first month. This piece covers the months after: what changes, who sees it first and who keeps the workflow current.
Four things change around a workflow.
A workflow is built against one moment: one version of the procedure, one export format, one model and one team. Expect each of them to move.
- A procedure.The company changes a rule, a threshold or a required fact. Whoever made the change knows first; the workflow knows nothing until someone tells it. Make “tell the workflow’s owner” a step in every procedure change.
- A system’s export.An update to the accounting system, the CRM or a web form renames a column or adds a field. The reviewer notices first, as a blank or odd field in an output. A workflow that reads a system through a connection gets the same check; Fully Powered Connectors explains how connections are chosen.
- The AI platform.ChatGPT, Claude, Copilot and Gemini each ship new models, and a new default model can change how a workflow writes or what it picks up. Whoever administers the workspace sees the announcement first; the test set shows whether the work changed.
- The people.The experienced user changes role, a new colleague starts, the owner goes on leave. The owner notices. The workflow seldom changes; the job aid and the responsibility table do.
The owner’s routine.
Three habits catch all four kinds of change.
- Keep the exception log.Users record every case where they stopped, corrected the output or asked: what the workflow did and what they did instead. It records cases, never people. The owner reads it weekly.
- Retest after every change.After any change to the instructions, the reference files, an export or the platform, every test case runs again before the team uses the new version. Each repeated miss in the log becomes a new case. Build the test set that proves an AI workflow is ready shows how to write them.
- Write a version note.One line per release: version, date, what changed, why, which tests ran and who approved it. When an output looks different, the note says whether anything changed on purpose.
Six months of one workflow.
An invented law firm built an inquiry handoff workflow in a 2-Week Workflow Sprint. It turns each new inquiry into one packet for conflict review and the responsible attorney. The firm administrator owns it, the intake coordinator runs it, and the attorneys decide which clients the firm takes. This is the administrator’s log from the end of the support window. Copy its three columns for your own.
Illustrative example with invented data. An invented 35-person employment and commercial litigation firm. Every case, count and role is invented.
| Month | What changed, and who noticed | What was done, and the version |
|---|---|---|
| 1 | Two inquiries named the other side’s parent company only in an attached letter, and the packet listed it as missing. The intake coordinator logged both. | A defect: reading attachments was agreed. Corrected, both cases added to the tests, all sixteen run again. Version 1.1. |
| 2 | The administrator added a required fact for employment inquiries: whether the person still works for the employer they describe. The change was the administrator’s own. | Checklist updated, workflow checked against it, employment cases run again with two new ones. Version 1.2. |
| 3 | The website’s form service changed its export: the call-back number moved and a “How did you hear about us?” field appeared. The coordinator saw blank phone numbers in a packet and logged it that morning. | Field mapping adjusted, every case run again on the new export. Version 1.3. |
| 4 | The firm’s AI platform announced a new default model two weeks ahead. The outside IT provider told the administrator. | All eighteen cases run on the new model before the switch. One packet listed missing items out of the firm’s order; the instruction was tightened and the cases rerun, with model and result recorded. Version 1.4. |
| 5 | The experienced coordinator became a paralegal and a new coordinator started. The administrator updated the responsibility table that week. | No change to the workflow. The new coordinator’s first five packets were reviewed in full; two questions became lines in the job aid. Version 1.4 stands. |
| 6 | The managing partner asked for a monthly summary of where inquiries come from, by practice area, using the new form field. | A new requirement, not maintenance: quoted as new work with a fixed fee for the agreed scope. The handoff runs on 1.4 meanwhile. |
In six months the workflow had four releases, and each was retested before anyone used it. The workflow in month six handles cases the original never met, and the administrator can show what changed and why. Every decision about a client stayed with people.
Where Keep it working comes in.
Each sprint includes support for delivery defects: five business days after handover for the 3-Day Workflow Sprint, ten for the 2-Week Workflow Sprint. Keep it working is our offer for the months after. It covers three things.
- Defect correction.When a workflow does not do what was agreed, we correct it, add the case to its tests and retest.
- Maintenance when something around it changes.A procedure, an export format, a system or the AI platform moves: we check the workflow, make the agreed adjustments and retest it.
- New requirements, quoted as new work.A new output, team, step or system connection is scoped and quoted as a fixed fee before any work starts.
The fee is agreed once your workflows are in use and set out in your agreement before the work starts. Your team keeps the routine: the exception log, the owner’s decisions, every approval. We do the checking, adjusting and retesting, and every retest goes in the workflow’s test log, misses included. The workflow prepares; your people check and decide, in month six as on day one.
All of it runs on the platform you already pay for. If you change platforms, What moves between AI platforms sets out what carries over.