Skip to main content
// FREE TOOL

The Copilot
Cost Calculator

Four legs make up a Copilot bill: seats, Copilot Credits, the Azure tail, and the prerequisites nobody budgeted. Enter your estate and this assembles all four into one monthly figure, then ranks your levers by what they save on your numbers.

Free. No signup, no email, no AI. It runs in your browser and nothing you type is sent anywhere or stored.

What this is

A forecast you can defend line by line, built from Microsoft list rates you can re-derive. It answers the question a rate table cannot: which leg of your bill is actually big, and therefore which conversation to have.

What this is not

A quote, and not a reconciliation. It does not read your tenant, your invoice or your negotiated rates. It is the pre-flight estimate, which is the thing most estates ship without.

First

Two facts that decide most of the output

The metered layer is additive. Agent runs and Cowork bill in Copilot Credits on top of the flat per-seat licence. The per-seat price did not change. If someone in your organisation is saying Copilot moved to consumption pricing, correct them before you design anything, because the seat leg is almost certainly still the largest number on your page.

Interactive turns by licensed users are zero-rated. Under fair use, a user holding the Copilot add-on draws no credits for interactive use, including grounded generative answers. Unlicensed users in Copilot Chat draw credits for all of it, and autonomous runs always meter, licensed or not. That boundary is the difference between a forecast and a scare.

The machine

Your estate, assembled

It opens on a worked example: an invented 1,200-person manufacturer that shipped two agents without a cost review. Load month 2 to see the same estate after two design levers, or clear it and enter your own. Both worked months reproduce the arithmetic in Appendix G of the book, line for line, so you can check the machine before you trust it.

Start from
Leg 1

Seats

The flat layer. Deterministic, invoice-faithful, and on most estates the largest number on the page.

users

The billing term is a separate decision from the commitment term, and it is priced.

Leg 2a

Credits, interactive

Only unlicensed users in Copilot Chat draw credits for interactive turns. A licensed user's interactive turn is zero-rated under fair use, so counting your whole population here is the most common way to overstate the bill.

users

Not headcount. Not licence count. The people without the add-on who actually use an agent.

%

Finite, static questions routed to a knowledge source instead of a grounded generative answer. This is the single biggest knob on the credit leg.

Leg 2b

Credits, autonomous

Autonomous runs always meter, licensed or not. Count runs, not records: a job configured to trigger per record fires once per record.

Check the trigger config, do not assume the design.

Each action bills 5 credits.

Leg 3

The Azure tail

What your agents call outside the credit meter. It bills in dollars on a different invoice, which is exactly why it goes missing.

Partitions x replicas.

A model reached through a custom connector bills in dollars here, not in credits above. Same capability, two different meters.

millions

Records x tokens per record x runs, divided by a million.

millions
users
bots
Leg 4

Prerequisites

Not Copilot costs. The cost of being allowed to turn Copilot on. Both rates below are directional, so treat the output as a magnitude and not a quote.

users

Only if your security team requires auto-labelling or advanced DLP before grounding on tenant data. Leave empty if you are already on E5.

users

SAM is free while at least one user holds a Copilot licence, so this is usually zero.

Rate path

How you pay for credits

Prepaid packs buy a cheaper effective rate and strand whatever you do not use, because pack credits do not roll over. Leave this on automatic to see the cheaper path for your volume.

The assembled bill

Seats500 users x $30.00$15,000.00
Credits161,280 credits, 7 prepaid packs$1,400.00
Azure tailSearch, ranker, model tokens, Power Automate$557.74
Copilot-attributable total$16,957.74
Prerequisites, flaggedDirectional rates. A magnitude, not a quote.$6,000.00
With prerequisites$22,957.74
Base M365 licenceDocument the key, do not re-litigate it every closeout of scope

Every line is rounded to the cent and the rounded lines are summed, so the column adds up the way an invoice does.

Chapter 2, the cost stack; Appendix G

Which conversation to have

Seats, share of bill88.5%
Credits, share of bill8.3%
Azure tail, share of bill3.3%

If a fifth of your seats are unused, that is $3,000.00 a month, which is 2.1x your entire metered layer. Nobody in your building is talking about that number.

Appendix G; Chapter 11, right-sizing seats

Packs against pay-as-you-go

Credits for the month161,280
Pay-as-you-go161,280 x $0.01$1,612.80
Prepaid, 7 packs175,000 credits of capacity, 13,720 unused and non-rolling$1,400.00
Packs wins by$212.80
Overage cutoff, 125% of pack capacityCustom agents stop here unless pay-as-you-go is connected behind the packs218,750
Interactive credits0 classic at 1, 11,340 generative at 12136,080
Autonomous credits25,200

Chapter 3; Appendix A, metered layer

Inside the Azure tail

Azure AI Search, Standard S1Structural. It does not care how many queries you run, and no design lever touches it.$245.00
Semantic ranker11,340 grounded queries, 10,340 chargeable$10.34
Azure OpenAI o1, reasoning, input$151.20
Azure OpenAI o1, reasoning, output$151.20
Azure tail total$557.74

The search unit is 43.9% of the tail. Cutting it is a capacity decision, not a design decision, and it belongs in a different review.

Chapter 5, the bill before you start

Your two biggest levers

  1. 1
    Route the repeatable 70% of turns to classic answers$807.94

    A week of transcripts usually shows most traffic is a handful of finite questions a static knowledge source already answers. Those intents do not need live tenant data, and the 10-credit grounding surcharge is what you stop paying.

    87,318 credits a month. Chapter 9, lever 2; Chapter 3

  2. 2
    Move the batch job to a small model$299.38

    Extraction and summarisation rarely need a reasoning model. On list rates o1 output is 100 times GPT-4o mini output, which makes this one dropdown the largest single multiplier in the tail.

    Chapter 5; Appendix A, Azure tail

  3. 3
    Batch the autonomous run to once per working day$200.00

    A run configured to trigger per record fires as many times as there are records. Batching changes the schedule, not the output: finance gets the same file in the morning.

    24,780 credits a month. Chapter 9, lever 3

Ranked by what each is worth on your numbers, not by how good it sounds in a deck. Every figure is the whole model re-run with that one input changed.

Chapter 9, design discipline as cost control

Plain text, with the levers and the caveats attached, so the number does not travel without them.

Accuracy

Three cells decide whether the output is worth anything

Every other input moves the total a little. These three move it by multiples, and all three are facts about your configuration rather than facts about Microsoft’s price list. Get them wrong and the arithmetic is flawless and useless.

Unlicensed users on agents

The usual mistake: Entering headcount, or licence count, instead of the people without the add-on who actually use an agent.

This is the cell that overstates a bill by several times and gets a project cancelled before it ships. A licensed user’s interactive turn is zero-rated. Count the others.

Autonomous runs per month

The usual mistake: Entering the number of runs the job was designed as, rather than the number the trigger config actually fires.

A nightly job drawn as one batched pass and configured to trigger per record fires once per record. Same code, same output, sixty times the meter. Open the trigger and read it.

Grounding, and the classic share

The usual mistake: Assuming every question needs live tenant data.

The grounding surcharge is five times the answer it attaches to. Most traffic is a handful of finite questions a static knowledge source already answers, and routing those is usually the largest single saving on the page.

Honesty

What this does not count

A calculator that pretends to cover everything is worse than one that names its edges. Four things sit outside this model, and two of them are bigger than the meter everyone is worried about.

The base licence

M365 E3, E5 or Business Premium underneath the add-on. Out of scope by policy, not because it is small. Write the key down so nobody re-litigates it at every close.

Soft costs

Training, change management, and the internal hours to run any of this. A Microsoft-commissioned Forrester model puts them near two thirds of three-year TCO, which is directional and vendor-funded, and this calculator counts none of it.

Adoption growth

Every figure here is one flat month. Real adoption curves climb, and a governed agent still costs more each month than the month before. Design levers change the slope, not the direction.

Your negotiated rates

These are Microsoft list prices. If you are on an enterprise agreement your rates differ. Swap them in. The method does not change, which is the only reason a calculator is worth more than a rate table.

Provenance

Where the rates come from

Every rate in this calculator is Microsoft list price as of May 2026, and every one of them lives in a single dated file. That is a design decision, not a filing convention: when Microsoft moves a price, one number changes and the whole tool is current again.

Two rates are flagged on screen because they are directional rather than confirmed first-party: the per-user cost of E5 Compliance for Purview, and standalone SharePoint Advanced Management. Treat anything flagged as a magnitude, not a quote from your reseller.

Verify against Microsoft’s current pricing before you commit a budget. That instruction is in the book too, and it is not boilerplate: the metered layer is the youngest part of this stack and the most likely to move.

This tool gives you the number. The book gives you the argument.

Every panel above cites a chapter, because the calculator is the arithmetic and the chapters are the reasoning: why the grounding surcharge dominates, where the meter hides on a second invoice, what a prepaid pack strands, and how to say all of it to a CFO who has already read a vendor deck.

The Real Cost of Copilot is 14 chapters and seven appendices, including the worksheet this tool is built from and the worked estate it opens on. One purchase, no subscription, no login.

Get The Real Cost of Copilot, $29

You do not need it to use this calculator. That was the point of publishing the calculator.

These rates will change

AI at Work goes out every other Tuesday: what changed, what it costs, and what to do about it. No vendor sponsorship.

Kesslernity. Independent practitioner, no vendor sponsorship. Rates are Microsoft list as of May 2026, access date 2026-05-31, and are reproduced from Appendix A of The Real Cost of Copilot. The worked example is an invented organisation, labelled as invented wherever it appears. Output is a forecast for planning, not a quote, and not tax, accounting or legal advice. Verify against Microsoft’s current pricing before you commit a budget.

Browse the full catalogue · The Copilot Chargeback Playbook, $49