AI Delivery Manager
You own the date, the budget and the delivery system for work whose scope moves under you: models change, evals surprise, stakeholders oscillate between fear and magic. This is the playbook I use, anonymised from delivery in regulated and engineering organisations.
5 daysgoverned knowledge hub, idea to live
200+person global team supported to award-winning delivery
30+coaches and portfolio leaders directed across regions
12% → 87%adoption in 90 days on a global programme
The job in one sentence
Ship governed AI value on a predictable rhythm: something visible every week, judgement on the quarter, and gates that match how reversible each decision actually is.
The operating rhythm
- Deliver weekly. Something visible ships every week: a working slice, a measured result, a retired assumption. Not a status deck.
- Review fortnightly with the owners. Use-case owners, not proxies. Each slice gets judged against its value hypothesis, not against activity.
- Judge quarterly. Direction changes happen on the quarter, on evidence. Weekly noise does not steer the programme.
- Keep one visible backlog. Scores, owners, stage, value hypothesis. Leadership sees one honest list. How the backlog gets built is the use-case engine in the enablement playbook.
- Gate by reversibility, not by size. See below: the one-way-door rule does more governance work than any board pack.
Reviewing everything and moving slowly looks safe. It is usually the risky option: the queue grows, trust drops, and unapproved AI use fills the gap.
One-way doors: the only gate that matters
- Reversible decisions move fast. Prompt changes, pilot scope, tool trials inside the approved tier: decided at team level, logged, reviewed later.
- Hard-to-reverse decisions slow down. Customer data exposure, model commitments in contracts, architecture that locks a vendor in, anything regulated: full evidence, named approver, written trail.
- Classify at intake. The single question "can we undo this in a sprint?" sorts 90-plus percent of decisions in seconds and saves the committee for the few that deserve it.
Hard deadlines with uncertain scope
- Fix the date, flex the scope. AI work makes scope estimates soft; the thin-slice discipline keeps the date hard. Every slice is shippable, so the deadline ships whatever is proven by then.
- Pin scope to workflows, not models. Models change monthly; the user's workflow does not. Contract and plan against the workflow outcome, and let the model be an implementation detail you can swap.
- Make evals the acceptance criteria. "The model feels better" is not done. An agreed eval set, run on every change, is: it turns model risk into a number the programme can manage. The eval architecture is in the architect playbook.
- Keep the human gate in the critical path and staff it. The commonest AI schedule slip is not the build; it is an unstaffed review queue.
When the programme is already red
- Stabilise the rhythm first. Before re-planning anything, get one visible thing shipping weekly. Cadence rebuilds trust faster than any recovery plan document.
- One honest list. Merge the competing trackers into a single backlog with real statuses. The act of merging surfaces the truth the programme has been avoiding.
- Re-baseline with the sponsor on evidence, not optimism: what is proven, what is pilot, what is hope. Cut the hope out loud.
The metrics that matter
| Leading (predicts delivery) | Lagging (proves delivery) | |
|---|---|---|
| Flow | Weekly ship streak, slice cycle time, review-queue depth | Committed date performance, slices shipped vs planned |
| Value | Eval scores per change, pilot usage in week one | Realised benefit vs the value hypothesis, adoption sustained at 90 days |
Status is reported from this table, not from sentiment. A green programme with a growing review queue is not green.
What not to claim
- No committed dates on unproven scope: the date is committed, the scope inside it is evidence-ranked.
- No "production ready" before deployment validation, whatever the demo looked like.
- No velocity theatre: slices shipped and benefits realised are the only counts that leave the team.
Questions to ask before you start
- Which decisions here are genuinely one-way doors, and who approves them?
- What does the sponsor expect to see weekly, and can we ship it?
- Who staffs the human review gate, and what is their capacity?
- What is the eval set, and who owns keeping it honest?
Running an AI programme, or rescuing one? Start with the advisory services or email hello@krishnachodipilli.com.