The demo went well. The model summarized the contracts, drafted the collection emails and answered the policy questions. Leadership was impressed. Six months later, nothing in the business works differently. This is the most common AI story in mid-market companies, and it is rarely a technology problem.
In our work we separate two states that most status reports merge: built and running. A pilot is built when it works in a test. It is running when the business depends on it every week, with an owner, a measure and a support process. Most AI programs report the first and assume the second.
1. Nobody owns the outcome
Pilots are often owned by the person who championed them, not by the leader whose numbers they are supposed to change. When the champion moves on, the pilot stalls. Fix: name an accountable business owner for every pilot before it starts, and write down the measure they will report.
2. The process was never written down
AI automates a process. If the process lives in three people's heads and differs by team, the model has nothing stable to follow. Fix: document the current workflow, agree the future one, and only then automate. This step is usually faster than leaders expect, especially with AI-assisted drafting.
3. The data has no single source of truth
Customer IDs in different formats, balances asserted in five places, a field nobody maintains. Models trained or prompted on inconsistent data produce confident, inconsistent answers. Fix: set one system of record for each fact and one identifier per customer before scaling anything.
4. There is no operating cadence
Running systems need a rhythm: a weekly check on accuracy and exceptions, a monthly review of results, a named person to call when it breaks. Pilots launched without that rhythm decay quietly. Fix: add each AI workflow to an existing operating review, with a short scorecard.
5. People were trained on the tool, not the work
A general AI course does not tell a credit controller when to trust the draft and when to rewrite it. Fix: role-based playbooks that show the review step, the escalation path and what good looks like.
What leadership can do this quarter
- List every AI pilot and mark it built or running.
- Name one business owner and one measure for each.
- Retire the pilots nobody will own.
- Add the survivors to your monthly operating review.
That single exercise usually cuts an AI portfolio by half and doubles what actually gets used. If you would like an independent view, our AI Readiness Assessment does this in three to four weeks.