How much capacity to reserve for unplanned work

In the end-to-end piece on capacity planning I wrote that if you have no data, twenty per cent is a defensible opening claim for the interruption reserve. This essay is about the words "if you have no data". You do have it. It is in the last two quarters, and getting it out takes about an hour per team.

Why a borrowed percentage fails

Any figure you adopt from someone else is an average over teams that are not yours. A platform team that carries the on-call rota and a product team two layers away from production do not have the same quarter, and a single percentage applied to both is wrong for each: too thin for the first, so its commitments slip, and too generous for the second, so capacity sits behind a number nobody can justify.

A borrowed figure is also impossible to defend. When someone asks why a team of six is only promising fourteen FTE-months, "the usual rule of thumb" invites the obvious reply, that this quarter will be calmer. "This team lost four FTE-months to unplanned work in each of the last two quarters, and here is the list" does not.

Measure your own interrupt rate

Do this per team, with the team, for the two most recent complete quarters.

1. List what was done that was never on the plan. Incidents and their follow-up. Customer escalations. Urgent asks that were accepted mid-quarter. Help given to other teams: reviews, migrations, borrowed people. Sources are the incident log, the calendar and the team's memory, in that order of reliability.
2. Size each item coarsely. In FTE-months, to the nearest half. "The March outage and its clean-up: one and a half." Anything under a quarter of an FTE-month goes into a single "small stuff" row the team sizes as a whole.
3. Divide by gross capacity. Headcount times three months. The result is the team's interrupt rate for that quarter.

An example with invented numbers, for a platform team of five:

For the coming quarter I would reserve four and a half, the higher of the two, and not the average. Two quarters is thin evidence, and the costs of being wrong are lopsided: a reserve that turns out too large hands back capacity you can put to use, while one that turns out too small breaks a promise somebody was planning against.

One boundary matters here. This is a coarse reconstruction per team, done once a quarter, and it is not time tracking. Nobody logs hours, and nothing is attributed to a person. The moment it becomes per-person it turns into consumption data, and people start managing the number instead of reporting it.

What counts as unplanned, and what does not

Two things get mixed into the reserve that do not belong there, and both make it meaningless.

Known recurring work is not unplanned. The support rotation, booked holidays, the upgrade nobody may skip: these are predictable, and each gets its own line, subtracted before the reserve. If the rotation hides inside the reserve, a quiet quarter for incidents looks like slack that is not there.

Overrun on committed work is not unplanned work. If a four FTE-month commitment took six, the extra two are estimation error and belong to that commitment's close. Hide them in the reserve and you inflate next quarter's reserve to cover an estimating problem, which guarantees the estimating problem never gets looked at.

The reserve is for work that arrives from outside the plan. Nothing else.

A named line, not hidden padding

There are two places to keep a buffer. One is inside each estimate: the team believes the work is three FTE-months and says four. The other is a single line in the team's capacity for the quarter, labelled as what it is.

Padding fails for reasons that have nothing to do with honesty. It is invisible, so the first person to press on an estimate negotiates it away without knowing what they removed. It is spread across every commitment, so it cannot be moved to the one that turns out to need it. It cannot be measured afterwards, because nobody recorded how much was buffer. And it destroys the estimates as evidence: next quarter you cannot tell whether four meant four.

A named line is defended once, at planning, with the table above. It sits in the capacity plan next to the support rotation and the holidays (the one-sheet capacity template has a column for exactly this), and what is left after it is the team's capacity line. Estimates can then be what the team actually believes, with a confidence level attached, because the uncertainty about the outside world is held elsewhere.

When the reserve is exhausted

Keep a running list through the quarter: each unplanned item, sized when it lands, subtracted from the reserve. It is the same list you will need at quarter end anyway.

When it reaches zero in week seven, three things follow.

Say so, that day. "The platform reserve is spent" is a status worth reporting on its own, before anything has slipped. It turns the next interruption from a surprise into a decision.

The next unplanned item is an edit to a named commitment. There is nothing left to absorb it. So it displaces something specific, and the person who was promised that something hears about it from you. The mid-quarter ask covers how to make that edit in the open.

Do not borrow from everything a little. The tempting move is to let all commitments run slightly late. That is over-commitment by another route, and the honest options are the same three as whenever a team is over-committed: remove something, move a date, or cut named scope.

The opposite case deserves a sentence. If the reserve is half unused in week ten, give it first to a commitment that is running over, visibly. Pull new work in only if it is small and already shaped. An unspent reserve is not a failure of planning. It is one data point, and a single calm quarter is not a reason to cut the line.

Close the loop at quarter end

At the close, the running list gives you a third quarter of evidence. Compare reserve held with reserve used, per team, and adjust slowly: move the line towards the evidence, not all the way to the latest figure. Look at composition as well as size. If two-thirds of the unplanned work was urgent asks from one direction, the fix is at intake, not in a bigger reserve. If it was incidents in one system, you have just assembled the evidence that system's maintenance work needs to win a place in the plan.

Where DeliverySheet sits

DeliverySheet will not measure your interrupt rate. It has no time tracking and no integrations, so nothing in it counts incidents or reads a calendar; the reconstruction above is yours to do. What it records is the other side of the subtraction: capacity estimated per team in FTE-months on each demand, so what a team has promised is visible in one place. And when the reserve runs out and a commitment has to move, re-baselining it requires a reason, which is kept on the commitment ledger.