How to scope a cloud and AI cost engagement

Six items in scope, six deferred, and eight acceptance criteria stated as figures on a named denominator. The one most often resisted is reporting the attribution effect separately from the engineering effect, because that is what shows which half of the saving will come back.

A cost proposal is in front of you and it promises a percentage. The percentage is the least useful thing in it.

What decides whether this engagement is worth buying is narrower: whether it produces a bill with owners, and whether anyone will be able to check the result in a year.

What belongs in the scope

What belongs in the scope, and what gets added

The deferred column is where these engagements turn into platform projects. Each item is real work and none of it makes the bill governable sooner.

Ask for the out-of-scope list in writing. A cost proposal with no out-of-scope section usually becomes a migration.

Source: Thinklytics case library, published engagement scopes and outcomes.

Six items, and the first and last are the ones usually missing.

The attributable share today, measured and written down. This is the baseline. Attributed means a named person would recognise the spend as theirs, not that it carries a tag.

The attribution dimension and the shared-cost method, decided. Team, product, customer or use case; and even split, by consumption, or a reported central pool. Tagging implements this decision. Teams that tag first re-tag, and re-tagging an estate costs more than settling the model.

One unit of outcome per product area, with the method stated. An order, a claim, a report, an active customer. Cost per order beats cost per compute hour in every budget conversation.

A named owner per attributed line, and the forum where it is discussed. This is what converts a report into governance. Attach it to a meeting that already happens.

The fast engineering levers, in parallel. Rightsizing, auto-suspend discipline, storage tiering, commitment coverage. They need nobody's consent, so they should not wait.

Separate baselines for the attribution effect and the engineering effect. The section below.

Deferred: a target before a baseline, a tagging sprint before the model, a full unit-economics model across the estate, a new cost dashboard, re-architecting workloads, and a single blended savings number. Ask for the out-of-scope list in writing. A cost proposal with no out-of-scope section usually becomes a migration.

The criterion most often resisted

Reporting the attribution effect separately from the engineering effect.

The two decay differently. An engineering cut can regrow, because the provisioning behaviour that created the spend is untouched and consumption returns over the following quarters, often in different services so the comparison is awkward. A behavioural change from a team seeing its own number does not regrow, provided the owner and the threshold survive the engagement.

A supplier who reports one combined figure is reporting the number that flatters the engagement rather than the one that predicts next year. That is why it gets resisted, and it is why it belongs in the contract rather than in a conversation.

The FinOps Foundation's State of FinOps 2026, covering 1,192 practitioners representing more than $83 billion in annual cloud spend, found 49% of organisations tracking unit economics, up 9 points year over year, and 63% with a dedicated FinOps team. Half are still managing a total rather than a rate, which is the condition a one-off cut leaves intact.

Acceptance criteria

Acceptance criteria, written before the work

Most of these are stated as a figure on a stated denominator, because a percentage saving with no baseline is not checkable.

• Attributable share at or above a stated percentage, on the agreed dimension. The criterion that makes everything else possible. State the target and the dimension in the statement of work.
• Cost per unit of outcome reported, with the method recomputable by someone else. Only 49% of organisations track unit economics at all, so this is the criterion that separates a durable result from a one-off cut.
• A named owner per attributed line, confirmed by that owner. Confirmed, not assigned. An owner who learns about it from a report will dispute it.
• The unattributable pool at or below a stated share, with a named home. And a date for shrinking it further, or it becomes permanent.
• Annual run-rate reduction stated against the recorded baseline. $4.7M to $1.1M, $1.9M to $420K, $1.3M to $310K in three published engagements. A before and after, not a percentage.
• The attribution effect and the engineering effect reported separately. Conflating them hides which half will decay, and the decaying half is the one that returns next year.
• A monthly showback live in an existing forum, with a threshold. Attached to a meeting that already happens. A new quarterly review is where this dies.
• The residual annual cost of running the governance, stated. Tagging maintenance, the reporting, and the owner's time. A proposal with no ongoing cost is hiding one.

The sixth is the one most often resisted and the most useful. A supplier who reports one blended saving is reporting the number that flatters the engagement rather than the one that predicts next year.

Source: FinOps Foundation, State of FinOps 2026, n=1,192; Thinklytics case library, published delivery outcomes and acceptance measures.

Eight, written before the work starts and tested against the recorded baseline. Three are worth expanding.

Attributable share at or above a stated percentage, on the agreed dimension. Name both the figure and the dimension in the statement of work, because a high share on the wrong dimension produces reports nobody can act on.

Annual run-rate reduction stated against the recorded baseline. A before and a after, not a percentage. Three published engagements landed the residual near a quarter of the prior cost: $4.7M to $1.1M, $1.9M to $420K, $1.3M to $310K. Three data points rather than a benchmark, and enough to test a claim against.

The residual annual cost of running the governance. Tagging maintenance, the reporting, and the owner's time. A proposal with no ongoing cost is hiding one, and here the ongoing cost is what keeps the saving from reverting.

Six questions before signing

Questions to ask before signing

Six, and the first two establish whether the firm is selling governance or a one-off reduction exercise.

• What share of our spend do you expect to be attributable at the end, and what will the remainder be?. A firm that will not name a share has not measured an estate like yours.
• Will you report cost per unit of outcome, or only cost per unit of consumption?. The second is easier and does not survive a budget review.
• What reduction do you expect from attribution alone, before any engineering?. A firm that has done this will have a view, because it is usually a material share of the total.
• Will the attribution effect and the engineering effect be reported separately?. If the answer is one blended number, you will not know which half decays.
• Who owns the tagging standard after you leave, and what stops it decaying?. Tag coverage falls the month the engagement ends unless somebody owns it and something enforces it.
• What happens if our owners cannot change anything without a change board?. Then the showback produces awareness and no reduction. A firm that has met this before will say so rather than promise savings.

A cloud provider's earlier consolidation attempts had stalled on business unit resistance rather than on technology. The engagement that worked gave each unit ownership of its own data and published through a shared catalogue, and recorded zero escalations.

Source: Thinklytics case library, published engagement approaches and outcomes.

The first two establish whether you are buying governance or a reduction exercise.

What share of our spend do you expect to be attributable at the end, and what will the remainder be made of. Will you report cost per unit of outcome, or only cost per unit of consumption. What reduction do you expect from attribution alone, before any engineering. Will the attribution effect and the engineering effect be reported separately. Who owns the tagging standard after you leave, and what stops it decaying. And what happens if our owners cannot change anything without a change board.

The last one is the one that saves the most. If an owner cannot resize, reschedule or decommission anything without a change board, the showback produces awareness and no reduction. A firm that has met that situation will say so. A firm that has not will promise a percentage.

Timeline, and what drives it

The consolidation and cost engagements in our case library ran 16 to 30 weeks, and the duration tracked the number of systems or business units rather than the size of the spend.

Thirty weeks for 14 business unit data warehouses, taking annual cost from $4.7M to $1.1M with zero business unit escalations after earlier centralising attempts had stalled. Twenty-six for nine reporting systems at a national telecom, $3.1M a year, with the monthly close falling from 18 days to three. Twenty-four for 11 campus warehouses, $2.3M a year. Twenty for four core banking systems after three acquisitions, $1.3M down to $310K. Sixteen for seven warehouse systems, and 16 for 220 store reports cut to 14 certified dashboards at $890K a year.

Attribution and showback on their own, without changing the estate, is a much shorter piece of work, and it is where to start if the question is who is spending this rather than whether the estate is right. Whether the estate itself needs consolidating is a separate decision covered in data stack consolidation.

What we would do first

Record the baseline before reading the proposal again. Three figures for last month: total spend in scope, the share attributable to a named owner, and cost per unit of outcome for one product area.

Then check the proposal's promises against those three numbers rather than against a percentage. Most cost proposals are written against a percentage because a percentage needs no denominator, and a denominator is exactly what makes the result checkable.

The cost attribution and showback plan is the worksheet we fill in with clients to produce those three figures, and it is deliberately usable without us. Why the order matters is in cut the bill or fix attribution first.

Delivery sits in cloud and AI cost optimization for the attribution and the levers, system consolidation where the estate itself is the cost, and data foundation for the platform layer the spend is buying. The full set of work in this area sits under the platform bill keeps climbing.

Frequently asked questions

What belongs in the scope of a cost engagement?

Six items. The attributable share today, measured and written down as the baseline. The attribution dimension and the shared-cost method, decided before any tagging. One unit of outcome per product area with the calculation stated. A named owner per attributed line and the forum where the number is discussed. The fast engineering levers, run in parallel because they need nobody's consent. And separate baselines for the attribution effect and the engineering effect, so the durable half stays visible.

What should be deferred?

A target before a baseline. A tagging sprint before the attribution model is decided, because teams that tag first re-tag. A full unit-economics model across the whole estate when one unit per product area would do. A new cost dashboard, which is where this work goes to be ignored. Re-architecting workloads, which is a platform project wearing a cost label. And a single blended savings number, which hides which half will decay.

What acceptance criteria should go in the contract?

Eight, most stated as a figure on a named denominator. Attributable share at or above a stated percentage on the agreed dimension. Cost per unit of outcome reported with a method someone else can recompute. A named owner per attributed line, confirmed by that owner. The unattributable pool at or below a stated share with a named home. Annual run-rate reduction against the recorded baseline. The attribution effect and the engineering effect reported separately. A monthly showback live in an existing forum with a threshold. And the residual annual cost of running the governance.

Why report the attribution effect separately?

Because the two halves decay differently and a blended number hides which is which. An engineering cut can regrow, since the provisioning behaviour that created the spend is untouched. A behavioural change from a team seeing its own number does not, provided the owner and the threshold survive. A supplier who reports one combined figure is reporting the number that flatters the engagement rather than the one that predicts next year, and it is the criterion most often resisted.

What should the baseline contain?

Three figures for a stated period, recorded before any optimisation starts: total spend in scope, the share attributable to a named owner, and cost per unit of outcome. Without them no later claim can be verified, and the second year's budget request will not be believed. A percentage saving with no denominator is not a measurement, and reduction work gets credited to whoever is in the room when the bill falls.

Who owns the tagging standard afterwards?

Ask, because tag coverage falls the month an engagement ends unless somebody owns it and something enforces it. The answer should be a named person plus a mechanism: a policy that blocks untagged provisioning, or a report that surfaces untagged resources to the owner weekly. A standard with an owner and no enforcement decays more slowly than one with neither, and both decay.

What if our owners cannot change anything?

Then the showback produces awareness and no reduction, and it is better to establish that before savings are promised. Ask what an owner is empowered to resize, reschedule or decommission without a change board. If the answer is nothing, the first piece of work is the authority rather than the reporting. A firm that has met this before will say so; one that has not will promise a percentage.

How long should it take?

The consolidation and cost engagements in our case library ran 16 to 30 weeks, and the duration tracked the number of systems or business units rather than the spend. Thirty weeks for 14 business unit warehouses, which took annual cost from $4.7M to $1.1M with zero escalations. Twenty-six for nine reporting systems. Twenty-four for 11 campus warehouses. Attribution and showback alone, without an estate change, is a much shorter piece of work.

The work behind this

Nine engagements in the case library carry cost optimization, running 16 to 30 weeks, and seven published an annual before and after figure against a recorded baseline rather than a percentage.

Cost optimization, 9 engagements.