Where AI actually shows up in the EBITDA bridge.
"AI will drive margin expansion across the portfolio" is not a number a CFO can budget.
We write AI as a named workstream: a sized delta, an implementation cost, an owner and a date. Not a theme — a line item.
Five questions every AI line must answer.
If a plan cannot answer all five, the AI line is decoration. It will not survive the first board meeting that asks for the variance explanation.
Which named process?
Not "customer service" — which queue, which document type, which decision step, which team. Vague scope produces vague savings.
What does it cost today?
FTE-hours multiplied by loaded cost, or vendor spend already in the P&L. Without a current-state cost there is no saving to state — only a percentage of an unknown number.
What displacement rate, on what evidence?
Observed pilot, comparable deployment at a peer, or reasoned estimate — and we say which. Each carries a different confidence level and a different risk of variance at completion.
What does the replacement cost to build and to run?
One-off implementation plus recurring licence, compute and maintenance. We report net, never gross. A saving that costs 80% of itself to sustain is not the number that belongs in the bridge.
Who owns it, and by which date?
A named role and a dated milestone, or the line does not go in the plan. Ownerless initiatives drift; dateless milestones never arrive.
Inside the deliverables you already use.
RAG assessment, KPI baselines, and phased initiatives.
Runs across Days 1–30, 31–60, and 61–90 alongside the six core functional areas. Sized and owned from Day 1, not added as an afterthought in month three.
A Day-1 owner, a first milestone, and the CP checks.
Data rights, GDPR/nLPD compliance, works council timing — the legal and operational prerequisites that determine whether an AI initiative can actually start on the date the plan assumes.
AI-addressable cost and capacity, sized outside-in.
Run before the books open. Named processes where automation removes cost or releases capacity — with an implementation cost, an owner and a date. Not as a theme; as a finding with a label.
Sometimes the answer is no.
On a business where the AI-addressable cost is small relative to the implementation overhead, we say so. A finding that an AI initiative is not worth the management bandwidth in year one is a useful finding. A firm that only ever finds upside is not doing analysis.
Capacity released is only an EBITDA delta if it is actually banked — either headcount avoided on a growth path, or cost removed. If it is neither, we label it as capacity released and keep it out of the bridge total.
Add the workstream, or run it standalone. Thirty minutes to scope it.
Scope it against one portfolio company or one live deal. We'll tell you whether the addressable cost justifies the workstream — or whether the honest answer is to skip it this year.