CPCD Calculator

THE FINANCIAL METRIC: What does it cost to make one AI decision you can legally, auditably and defensibly stand behind?

Framework Three // When Agents Rule

Cost Per Compliant Decision

Price your governance as unit economics rather than as a cost line. Compare two infrastructure or governance positions, get the trajectory of both, and get the present value cost of waiting a year to act.

The formula, Chapter 10 CpCD = (Compute + Compliance Overhead + Latency Cost + Governance Overhead) divided by Decision Volume

4 cost components 2 scenarios compared 5 year trajectory 6 minutes

It needs no new instrumentation. Every input already exists somewhere in your enterprise. The metric assembles them into the one number that closes the gap between three dashboards nobody looks at together: the cloud bill, the regulatory exposure register and the model performance report.

It is loaded with the book’s own worked example so you can see the shape of an answer before you enter anything. Replace every figure with yours, or clear the form and start blank. Nothing is transmitted or stored, and everything is computed in your browser.

Two traps it checks for. Whether your owned compute figure survives the utilisation you actually achieve, which is the commonest way an on premises case fails its first finance review. And whether you are about to show a board a falling unit cost while the total bill rises, which is what happens when token prices fall more slowly than agentic token consumption grows.

The workload

CpCD is calculated per workload, not across the portfolio. Pick one.

Loaded with the worked example from Chapter 10. Continental Industrial Holdings credit decisioning, one million decisions a month, moving from public cloud to a private on premises stack. Replace every number with your own, or use Clear the form at the bottom of the page.

Workload name Labels the output only.

Compliant decisions per month Decisions that must be auditable and defensible. This is the denominator.

per month

Currency Three letter code. Labels the amounts so a board paper cannot be read in the wrong currency.

Named owner accountable for this number A person, not a function. An unowned unit cost is an analysis. An owned one is a commitment.

Annual technology budget Optional. Used only to express the saving as a share of spend, which is how a board reads materiality.

USDper year

One off transition cost Migration, build, licensing, change. Enter zero if this is a governance comparison rather than a move.

USD

Baseline, what you run today

Monthly spend per component. The tool divides by volume to get the per decision figure.

Compute GPU time, tokens, API fees, or amortised owned hardware. From your FinOps data.

USDper month

Compliance overhead Audit logging, lineage, regulatory documentation, evidence preservation.

USDper month

Latency cost Round trip delay converted to money: abandonment, service level penalties, risk windows.

USDper month

Governance overhead Human review sampling, escalation handling, model monitoring, override documentation.

USDper month

Target, what you are proposing

The same four components under the proposed placement or governance model.

Be careful with compliance overhead. Chapter 10 is explicit that it does not fall with an infrastructure change and may rise slightly, because more of the logging infrastructure becomes yours. A target that shows compliance overhead falling is the first thing a sceptical CFO will challenge, so have a reason ready or leave it unchanged.

Compute Include amortisation over the expected useful life if this is owned capacity.

USDper month

Compliance overhead Usually unchanged. The regulation does not care where the inference runs.

USDper month

Latency cost Recalculated from the target round trip, using the same calibration.

USDper month

Governance overhead Unchanged unless the autonomy level or the review sampling changes.

USDper month

Does the target rely on owned or reserved accelerator capacity?

An amortised figure is only valid at the utilisation you actually achieve. Measured enterprise accelerator utilisation is commonly reported far below nameplate.

No. The target uses metered or on demand capacity only. Yes, and sustained utilisation above 40 percent is measured, not assumed. Yes, and measured utilisation is between 15 and 40 percent. Yes, and utilisation is below 15 percent, or we have not measured it.

How was the latency cost figure produced?

This is the softest of the four inputs. Published sensitivity ranges from roughly one to seven percent conversion loss per hundred milliseconds, a sevenfold spread, so a borrowed benchmark is not evidence.

Calibrated against our own conversion, abandonment or penalty data. Estimated from published industry benchmarks. Not quantified. Both latency figures are set to zero or are nominal.

Trajectory assumptions

Unit cost and total cost move in opposite directions. These four figures decide by how much.

Annual growth in decision volume Your own forecast. This is what makes the do nothing position get worse rather than stay flat.

percent

Annual decline in compute unit price Per token and per inference prices have been falling steadily. Enter the rate you will plan against.

percent

Annual growth in compute consumed per decision Agentic workloads consume several times more tokens per task than chat. This is what cancels the price decline.

percent

Annual growth in compliance overhead per decision Rising evidence requirements. A workload whose compliance overhead is climbing is a signal, not a nuisance.

percent

Discount rate Your corporate hurdle rate. Ask finance rather than choosing one.

percent

Ready.

The ask

The decision, the amount, the owner and the consequence of deferral, in the order a board wants them. Everything below this panel is the evidence for it.

Cost per compliant decision

Baseline 0.000 per compliant decision

Target 0.000 per compliant decision

Change 0%

Where the money is, and what actually moved

Per decision, to four places. The component that does not move is usually the most informative line in the table.

Swipe the table sideways to see every column.

Component	Baseline	Target	Change	Monthly effect

Cash effect

At today’s volume, before any trajectory effect.

monthly saving at current volume

annualised at current volume

Five year present value

Savings discounted at your hurdle rate, with volume growth and unit price movement applied.

present value of the saving, net of transition cost

present value cost of waiting twelve months

Confidence in this case

Graded on the four things that decide whether a number survives challenge, not on how good the answer looks.

Evidenced

The case a sceptic would rebuild

Every unproven input replaced with the conservative version, then recalculated. This is the number to present, not the optimistic one.

challenged present value, net of transition cost

0.0000

challenged target cost per compliant decision

How wrong can we be before this fails

The case is linear in target compute cost, so the breakeven is exact rather than estimated. This is the answer to the question a CFO asks second.

0.0000

target compute cost per decision at which five year present value reaches zero

headroom: how far target compute can overrun before the case fails

The trajectory a board needs to see

Unit cost and total cost move in opposite directions. Showing only the first is how a cost surprise reaches a board.

Swipe the table sideways to see every column.

Year	Decisions	Baseline CpCD	Target CpCD	Baseline total	Target total	Saving

Assumption checks

The three things a finance function will test before it accepts the number.

The same number, read by three executives

CpCD reaches three people who otherwise look at three different dashboards.

The CFO reads a cost figure

The Chief Risk Officer reads exposure per decision

The CIO reads efficiency adjusted for performance

Board paper

Written to be pasted into a board or investment committee paper without editing. Replace the bracketed fields before circulating.

Take it further

CpCD is the executive number. These are the instruments that produce the inputs and the case behind it.

The full investment case

The Business Case and CpCD workbook carries the five year model this page summarises: sourced assumptions, hard cost reduction kept separate from expected loss avoided, a two way sensitivity grid, and a board summary that states plainly whether the case stands on cost reduction alone.

Get the workbook

Why governance overhead is what it is

Governance overhead scales with autonomy and reversibility. The L0 to L5 Agent Classifier derives the level, the tier and the oversight mode the decision velocity actually permits, which is where the governance component of this formula comes from.

Classify the system

The reasoning

Chapter 10 introduces CpCD and works the example this page loads with. Chapter 7 covers the PLACE Framework that makes placement decisions legible in these terms. Chapter 11 applies the same discipline to quantum migration economics.

Get the book

Calculating this by hand

Interactive calculation is unavailable in your browser. The method is unchanged and takes a few minutes with a calculator.

Pick one workload. Establish the number of decisions a month that must be auditable and defensible. That is your denominator.
For your current position, total the monthly spend on each of the four components: compute, compliance overhead, latency cost and governance overhead.
Divide each by monthly decision volume. Add the four results. That is your baseline cost per compliant decision.
Repeat for the position you are proposing. Hold compliance overhead flat unless you can explain why the regulation cares where the inference runs.
The reduction is baseline minus target, divided by baseline. The monthly saving is that difference multiplied by monthly volume.
Then do the part most people skip. Grow decision volume by your own forecast and recompute total spend for each of the next five years. If volume grows faster than unit cost falls, the total bill rises while the unit cost improves. Report both.

The worked example from Chapter 10. One million decisions a month. Cloud baseline: compute 32,000, compliance 19,000, latency 8,000, governance 4,000, giving 0.063 per compliant decision. Private on premises target: compute 4,000, compliance 19,000 unchanged, latency 1,000, governance 4,000 unchanged, giving 0.028. A saving of 35,000 a month and a 55 percent reduction in cost per compliant decision.

Two checks before you present it. If the target relies on owned accelerator capacity, confirm the amortised figure holds at the utilisation you actually achieve rather than at nameplate. And confirm the latency figure is calibrated against your own conversion or penalty data, because published sensitivity spans roughly one to seven percent conversion loss per hundred milliseconds.

Governance funded is governance that happens

If you cannot present Cost Per Compliant Decision today, you are managing the AI portfolio without its unit economics.

The book page Free toolkit Get the book Advisory services When Clouds Fail

About this calculator. Cost Per Compliant Decision is the headline financial metric of When Agents Rule by Steven Oppenheim, introduced and developed in Chapter 10. This tool applies the formula as the chapter defines it and loads with the chapter’s worked example. It is a modelling instrument, not an audit, a valuation or a benchmark, and it reflects the figures you enter: a generous assumption produces a flattering answer and no useful information. No peer benchmark is claimed and no external dataset underlies it. Trajectory results depend on forecasts of volume, unit price and evidence requirements that nobody can make with confidence, which is why the tool asks you to supply them rather than supplying them for you. Nothing you enter leaves your browser: there is no account, no transmission and no storage. This is not legal, technical, accounting or financial advice, and it is not a substitute for your finance function.

ip traffic

Opey IT

CPCD Calculator

Cost Per Compliant Decision

The workload

Baseline, what you run today

Target, what you are proposing

Trajectory assumptions

The ask

Where the money is, and what actually moved

Cash effect

Five year present value

Confidence in this case

The case a sceptic would rebuild

How wrong can we be before this fails

The trajectory a board needs to see

Assumption checks

The same number, read by three executives

The CFO reads a cost figure

The Chief Risk Officer reads exposure per decision

The CIO reads efficiency adjusted for performance

Board paper

Take it further

The full investment case

Why governance overhead is what it is

The reasoning

Governance funded is governance that happens

About Opey IT