AI Strategy and Transformation: Workflows
What arrives, what leaves, the discover-decide-prove loop, and the handoff contracts with the PMO, Data and Analytics, HR and the CEO.
Markdown. No sign-up, no email.
This function sits between everyone else. It does not deliver changes and it does not measure them; it decides which ones happen, in what order, and whether they paid. Those handoffs are where the work goes wrong, so they are written down here.
What arrives#
| From | What | Becomes |
|---|---|---|
| Any function | "This process is painful" | A candidate, once its hours are measured rather than asserted |
| Finance | Cost and headcount by function | The hours side of the ranking |
| Data and Analytics | Cycle times, volumes, error rates | The baseline |
| Research and Innovation | A capability that now works | A candidate, or a note that it still does not fit |
| Legal and Compliance | Regulatory constraints | Entries on the refusal list |
| HR | Where roles would change | A gate, not an input. Nothing schedules past it |
A complaint is not a candidate. It becomes one when someone has measured the hours. This is the most common place the pipeline is short-circuited, because a loud complaint feels like evidence.
What leaves#
| To | What |
|---|---|
| PMO | Defined, sequenced changes with a baseline already running |
| CEO | The portfolio, the capital ask, and the kill recommendations |
| AI Engineering and Product | Changes that need something built |
| Every function | The refusal list, and the published benefit results |
| Knowledge Management | The record of what was tried and what happened |
The main loop#
The single link that makes the whole thing work is baseline, change, same measure, published result. Break any one of the four and the function reverts to counting pilots, which is the failure mode it exists to avoid.
Handoff contracts#
To the PMO. A change is handed over only with: the measure named and running, the rollback path written, the accountable person named, and the exception route defined. A change handed over without these comes back.
From Data and Analytics. The baseline is theirs, not ours. We say what to measure; they produce the number and they publish it. If both roles ever sit in one place, the benefit figures stop being evidence.
To HR. Any change that alters what people do all day goes to HR before it is scheduled, not before it ships. By the time something is in a delivery plan, the conversation about roles has already been decided by default.
To the CEO. Kill recommendations go up with the evidence and a recommendation, not as an open question. The portfolio review is where they are decided, in the meeting, rather than deferred, and a deferred kill is how a failing programme survives another quarter.
Cadence#
| Weekly | Pipeline standing: what is stuck, and on whom |
| Monthly | Latency and cost per changed process |
| Quarterly | Portfolio review with kill decisions taken in the room; refusal list re-read and dated |
| Annually | Full process inventory rebuilt; 12-month benefit re-checks |
The failure this design is built against#
The common pattern is a transformation function that becomes an internal marketing department for AI: it collects enthusiasm, launches pilots, publishes an adoption dashboard, and cannot say whether the work got cheaper. Every element above, the hours-first ranking, the frozen baseline, measurement owned elsewhere, published refusals, kill decisions taken in the meeting, exists to make that outcome structurally difficult rather than merely discouraged.