The RACI for a deployment
Fill a deployment's RACI with decisions and named people, one accountable person per row, move your Rs to the customer by week 8, and cost the owner's hours.
On this page
A RACI is a grid of decisions against people, one letter per cell: R does the work, A decides and answers for it, C is asked before, I is told after. For a deployment it says who decides once you have gone. Written in week 7 as an afterthought, most rows have already defaulted to the engineer; written with the scoping doc and read again at handover, it finds the decision nobody at the customer owns.
The grid as it is usually written
The version most pilots produce lists activities against teams:
| Activity | Vendor | Claims | IT |
|---|---|---|---|
| Build the integration | R, A | C | C |
| Prompt engineering | R, A | I | |
| Testing | R | R | |
| Go-live | R | A | R |
| Support | A | R |
Three things are wrong with it. The rows are activities, not decisions: "testing" has no moment at which anyone says yes or no, so nothing in the row can be late or wrong. The columns are teams, and a team cannot be paged or asked why it approved a change; "Claims" owns nothing at 02:00, the point the definition of done makes about its fifth line. And the vendor holds A on the prompt and on support, which is true for eight weeks and false from the Monday after handover.
Six decisions, five people
The claims-triage pilot's grid has one row per decision someone will have to make again after week 8, and one column per named person: the VP of Claims as sponsor; claims operations, held by the deputy since the champion left in week 4; the claims QA lead; the IT platform on-call lead; and you.
| Decision | VP of Claims | Claims ops | QA lead | IT on-call | FDE |
|---|---|---|---|---|---|
| Approve the eval set | I | C | A | R | |
| Approve a prompt change | I | A | I | R | |
| Change a threshold | I | A | C | C | R |
| Move a segment through a gate | A | C | C | I | R |
| Answer an alert out of hours | I | A, R | C | ||
| Own the prompt after week 8 | I | C | A, R | I |
Two rules make it work. Exactly one A per row: R can be shared, A cannot, because two people accountable means each can wait for the other. And every A sits at the customer by week 8, because an accountable person outside the building cannot approve a change at the customer's pace or be asked about it by the customer's auditor.
Reading it row by row
Who approves the eval set. The QA lead. She owns the gold set, the two senior handlers' labels as adjudicated, and the 95% line is measured against it. You draw the stratified sample and prepare the adjudication, so you are R; claims operations is consulted because the segments are theirs and its head rules on each labelling disagreement; the VP is informed, because he cannot judge 200 cases and should not be asked to.
Who is consulted on a threshold change. A threshold looks like a configuration value and behaves like a policy. Lowering the confidence floor below which no clause is shown puts suggestions on more claims, and if the added claims are the hard ones, the miss rate rises. The QA lead is consulted because she re-runs the eval at the new value. The IT on-call lead is consulted too: the rollback trigger was derived from the current miss rate, so a threshold change moves how often the alert fires, and the pager is theirs. Claims operations is accountable, because the floor lives with the result.
Who moves a segment through a gate. The VP, his only A, because a gate move changes the pilot's scope and date; the week-7 meeting where he tried to move one is why it is written down.
Who answers the page. The IT on-call rota, A and R in one cell, with you consulted until your access ends. Not the QA lead: a stalled queue at 02:00 is an infrastructure fault with a fallback, and a drop in quality is never a page. Nothing in this pilot pages before 07:30, so out of hours the rota owns the ticket and, from 07:30, the page.
Reading it by column: no R or A of yours after week 8
Read the FDE column: R, R, R, R, C, blank. Each R is work that continues after week 8 and needs a name in another column before you leave; an R left in the vendor's column is a decision made by default, by nobody until something breaks. From week 9 the eval set, prompt changes, threshold changes and the gate reading move to the QA lead as R. Your C goes when your access does, at week 8 plus 14 days. The handover doc carries the post-week-8 grid, dated.
The cell that is empty in most deployments
Row six, the QA lead's column: the prompt after week 8. It is empty for a reason other than negligence. During the pilot the vendor's engineer does the work and the customer approves it, so nobody at the customer has ever been R on a prompt change. The sponsor assumes the vendor's support contract covers it; the vendor's support team assumes the customer owns its configuration. Both assumptions hold until the first change.
That comes sooner than anyone expects. The insurer issues new endorsement wording about twice a month, and each issue is a re-index and an eval re-run; now and then one changes which clause governs a class of claims, and the prompt must follow. With the cell empty the re-index still runs, because it is a scheduled job on IT's side, and nobody re-runs the eval. Agreement drifts down a point at a time with no reading to show it, and the first to notice is a handler in week 14 who has stopped trusting the panel.
A name without hours is a blank, so the cell gets both. Reading an automated eval re-run's disagreements, about 20 cases at three minutes, is an hour per wording issue, two a month. A monthly drift check on 50 fresh claims needs a senior handler's labels, 50 × 11 minutes ≈ 9 hours, plus an hour of the QA lead's adjudication. An occasional prompt change adds an hour. Call it four hours a month of hers and nine of a senior handler's, approved by the VP and written into the handover doc, which is why he is informed on the row.
When the grid changes
People leave, and the grid is dated for that reason. When the champion resigned in week 4, the claims operations column went to the deputy that week, and his name went beside hers on the definition of done before her last day, the move champion, blocker, executive sponsor describes. A column whose person has left is the empty cell again, one row at a time.
Your own PM reads the grid for rows that should be product defaults rather than a customer's decision, such as a threshold every customer sets by hand, the line PM and FDE: who owns what draws.
In an interview
The case round often ends with "who owns this after you leave?", and the deep-dive asks the same of an eval. The interviewer listens for rows that are decisions, columns that are people, one A per row, no R or A in the vendor's column at handover, and the owner costed in hours. Fill the empty cell out loud: "the QA lead, four hours a month, approved by the sponsor".
The mistake that loses points is a RACI with the vendor accountable for anything after go-live. It reads as an engineer planning to be the support contract, and the interviewer will ask who approves the first prompt change in month three.
Check yourself
1. The VP proposes that claims operations and the QA lead be jointly accountable for prompt changes, "so they both have skin in it". Agree?
No. Two As let each wait for the other, and the first prompt change after handover sits until one decides it is theirs. The QA lead is A, because her eval is how a change is approved; claims operations is C, because the floor feels it.
2. The QA lead's manager offers one hour a month for the tool after week 8. Estimate what goes unowned and decide what you do before handover.
She needs about four: an hour for each of two wording issues, one for the drift sample, one for the occasional prompt change. With one, the drift sample goes unread and the slow decline unseen. Take the arithmetic to the VP, who can approve hours; if he will not, automate the check that has no reader: the re-index job refuses an index whose eval falls more than a point below the last reading.
3. The IT on-call lead objects to being consulted on threshold changes: "we run infrastructure, not models". What do you change?
Keep the C and change its form. They need the alert consequence in one line on the change request: "expected misses go from 0.5 to 0.8 a day; the trigger stays at three; false alarms go from about once in 70 working days to about once in 20". The pager is theirs, so the consultation is written in pager terms.