From an annual cost model to a living one: what changes and what stays hard

In brief

Usage-based cloud and AI spend outruns a TBM model rebuilt each year. What a living cost model means, what stays hard, and where AI agents help and don't.

Steve Elliott

Steve Elliott · · 7 min read


On this page
  1. Why does a close-cycle cost model fall behind?
  2. What does a living cost model mean in practice?
  3. What stays hard?
  4. Where can AI agents help, and where can't they?
  5. How do you move from an annual model to a living one?
  6. Sources

A common pattern is a TBM cost model rebuilt once a year for the budget and refreshed after each month-end close. In our view that cadence no longer fits: usage-based cloud and AI spend, standing product teams and quicker decisions need a model that updates as data arrives and answers questions when a decision is made, not weeks later.

A living model doesn't replace the close. The ledger still decides what the books say. What changes is that the model carries a labeled, provisional view between closes, shows how complete each number is, and reconciles to the ledger when the close lands. Parts of that are hard, and AI agents help with some of them, not all.

Why does a close-cycle cost model fall behind?

Because its clock is the general ledger, and the costs that matter most now move faster than the ledger does. Much current guidance sets its review cadences against that clock. The TBM Council's page on model fallout says to "treat fallout detection as part of monthly financial operations, not an ad hoc task." It suggests revalidating cost pool mappings quarterly, or when accounts are added, repurposed or retired, and reassessing allocation drivers at least annually.

In our view, that rhythm suits staffing, depreciation and data center costs, which change slowly. Three things strain it.

Usage-based spend moves inside the month. Bain found the cost of tokens fell by half from December 2024 to December 2025, while tokens consumed grew about 4.5 times. Its authors write: "Annual planning cycles don't accommodate weekly capability shifts." They also note that token spend doesn't fit an existing line item. Take an illustrative AI service budgeted at $500,000 for the year on last year's run rate. If its usage followed that trend, by year-end it would be running at 2.25 times the budgeted rate (half the unit price at 4.5 times the volume), an annualized pace of about $1.1 million. A model tied to the budget and the close sees it only as actuals land, after the spend is committed.

Funding follows products, not just projects. TBM Taxonomy 5.0 renamed the Business layer the Consumer layer and added value streams, products, portfolios and digital services, so costs can be attributed to Agile portfolios. A product owner deciding whether to add a team this quarter needs this quarter's unit cost, not the one set in last year's budget.

Some decisions are starting to be made by software. In November 2026 the TBM Council repositioned TBM as the business operating system for AI. One of its five strategic moves positions the TBM Framework and Taxonomy as "the contextual and semantic layers for Agentic TBM," a standard for autonomous AI workflows. The Council doesn't say cost models must update continuously. That is our inference: if agents act on cost context, stale context gets acted on faster.

What does a living cost model mean in practice?

In our view, five changes separate a living model from a close-cycle one. The left column describes common practice, not TBM Council guidance.

Close-cycle modelLiving model
Data feedsMonthly ledger extract, plus spreadsheetsAutomated feeds from the ledger, cloud and AI billing, HR and asset data, each on its own schedule
MappingReviewed or rebuilt each cycleVersioned rules; new accounts, vendors and cloud accounts mapped as they appear
FalloutCounted at closeMonitored continuously against thresholds
DriversRefreshed at close or once a yearUsage drivers refreshed with usage; slower drivers on a set cadence
OutputsMonthly reports and the annual budgetAlso scenario and forecast views, tied to the portfolio, funding and renewal decisions they inform

Some of this already sits in standards guidance, for fallout monitoring and mapping governance. The same TBM Council page recommends dashboards that monitor unmapped spend, exception counts and reconciliation variances continuously, reports that surface ledger lines without a cost pool in near real time, and mapping logic stored in version control. The FinOps Framework's principles say to "process and share cost data as soon as it becomes available," and call for real-time financial forecasting and planning.

The piece that makes provisional numbers usable is status. FOCUS 1.3, released December 11, 2025, requires conforming providers to timestamp datasets and flag completeness status, so you can tell whether billing data may still be revised. We think every number in a living model should carry a similar label: provisional (usage-based, before close), closed (reconciled to the ledger) or forecast. Without that label, a fast model is just a confusing one.

The last row matters most. A model that updates hourly but feeds no decision is a faster report. The test from our post on stalled transparency programs still applies: list the decisions the model informed last quarter.

What stays hard?

Speed creates problems that a once-a-year rebuild hides.

  • The ledger runs on its own clock. Invoices arrive late, and usage data and the ledger won't agree within the month. The FinOps Foundation notes that missing rows may not surface until after month-end close. A living model should show the gap between provisional and closed figures, not paper over it.
  • Accruals and the close belong to the controller. A provisional usage view is a management estimate, not an accrual. If the two are allowed to blur, the model becomes a shadow ledger that we think finance would rightly reject. How usage-based costs are accrued is an accounting question; confirm treatment with your controller and auditors.
  • Allocations need to be stable to be useful. If every cost driver refreshes daily, a team's allocated cost can move because another team used more of a shared service. In our view, driver quantities can refresh with usage while driver methods change only on a governed cadence, and chargeback rates stay fixed for a set period.
  • Auditability gets harder with every change. Each figure must be reproducible as of a date: which data, which mapping version, which drivers. The Council recommends a changelog of mapping updates to preserve historical comparability. A model that changes continuously multiplies versions, so this has to be automatic.
  • Trust is easy to lose. Numbers that change every time someone looks lose credibility. In our view, a stable closed view, published at the same time each month, is what lets people trust the provisional one.

Where can AI agents help, and where can't they?

We think agents are useful for the repetitive work that keeps a model current, and should not be trusted with the judgments that make it credible.

Where they can help, with a person approving:

  • Suggesting mappings. The Council already recommends keyword lookups against account descriptions to suggest mappings, with the source and confidence documented. A language model can propose mappings for new accounts, vendors and cloud subscriptions in the same way, on the same terms.
  • Triaging fallout. Grouping exceptions and pointing to a likely cause, such as a new cloud account or a reused ledger account, can shorten the queue.
  • Drafting variance explanations. A first draft of why a service cost moved, with links to the lines behind it, can save analyst time.
  • Answering questions in context. If the model is current and well mapped, an agent can answer "what does this service cost per transaction this quarter?" when the question comes up.

Where they can't:

  • Create data that doesn't exist. If time records don't tag AI work, an agent can only guess at AI labor, which the Council's Beyond Tokenomics paper puts at 50–70% of AI project cost, without citing a source. Our post on AI initiative costs covers that split.
  • Own policy. Allocation methods, capitalization and accrual estimates are judgments someone in finance is accountable for.
  • Make a number auditable by themselves. Language model output can be wrong and can vary between runs. Every change an agent proposes needs a logged approval and a trace to the data and rule versions it used.

The Council's Beyond Tokenomics paper argues TBM lets leaders "see the fully burdened costs, both variable and fixed" of AI. In our view an agent reasoning from the taxonomy can only be as good as the model under it. That is the strongest practical argument for keeping the model current.

How do you move from an annual model to a living one?

Not by rebuilding everything at once. A sequence we think works:

  1. Name the decisions and their dates. Quarterly product funding reviews, major renewals and AI capacity requests are good first candidates.
  2. Split the model by clock. Keep ledger-driven pools such as staffing and depreciation on the close cycle. Put usage-driven lines, such as cloud, AI and SaaS, on feed cadence.
  3. Automate feeds and label status. Use FOCUS recency and completeness data where providers supply it, and mark every figure as provisional, closed or forecast.
  4. Monitor fallout continuously. The Council notes that dashboards often use bands such as under 1% fallout green, 1–3% yellow and over 3% red.
  5. Govern drivers. Let quantities float and change methods on a published schedule, with every version kept.
  6. Add forecast and scenario views. Feed usage-driven lines into a rolling forecast, and let product and portfolio owners test a funding change before they commit it.
  7. Reconcile at close. Publish the closed view as the record, and report the gap from the provisional view so the provisional numbers earn trust over time.

Cloud cost data in particular sits at the line between TBM and FinOps; our comparison of the two covers who should own which allocation.

This is what an operating model is for: cost, funding and the work should sit in one connected model, so the numbers are current when the choice is made rather than explained afterward.

Sources

Frequently asked questions

What is a living IT cost model?

It is a TBM cost model that updates as data arrives rather than once a year or only at month-end close. Feeds and mappings are automated, fallout is monitored continuously, usage drivers refresh with usage, and each figure is labeled provisional, closed or forecast. It reconciles to the general ledger at close and feeds the funding decisions it was built for.

Does a living cost model replace the month-end close?

No. The general ledger and the close still determine the books, and accruals remain the controller's responsibility. A living model adds a labeled provisional view between closes, mostly for usage-based cloud, AI and SaaS spend, then reconciles to the closed figures. Showing the gap between the two is part of how the model earns trust.

How often should TBM allocation drivers be updated?

The TBM Council's fallout guidance says to reassess allocation drivers at least annually and to revalidate cost pool mappings quarterly or when accounts change. In our view driver quantities, such as compute hours, can refresh with usage, while driver methods should change only on a governed schedule so allocated costs don't swing for reasons a team didn't cause.

Can AI agents maintain a TBM cost model?

They can help with repetitive work: suggesting mappings for new accounts, triaging fallout, drafting variance explanations and answering questions from a current model. They cannot create missing data, own allocation or accrual policy, or make figures auditable alone. Every change an agent proposes should have a logged human approval and trace to the data and rules it used.

Free playbooks

In-depth guides for product and operations leaders.

Browse the playbooks
Steve Elliott

Steve Elliott

Published November 22, 2026

More on TBM & IT Financial Management

Cost pools, towers, showback, unit cost.