AI costing prompts that a cost specialist actually trusts
Controllers, CFOs and finance teams are already asking ChatGPT, Claude, Gemini and Copilot to "build my costing model" or "help me with budgeting and forecasting". Most get a confident answer built on invented numbers. This hub gives you the opposite: prompts we wrote and tested as cost and profitability specialists, each with a worked example, an anti-hallucination guardrail, and an honest note on what the AI can and cannot do with your data.
Tested on current AI assistants. Updated June 2026.
In short
An AI assistant is excellent at giving you the structure of a cost model, explaining a method, and doing the arithmetic once you supply the data. It is dangerous when you let it guess the data, because large language models produce statistically plausible text, not verified facts. The fix is a well-built prompt: give it a role, give it your real figures in a table, ask it to show every formula, and forbid it from inventing numbers. The prompts on this hub do exactly that. They will not replace a model built on your own data and validated by a specialist, but they will get you a clean first draft and stop you trusting a fabricated one.
Why "build my costing model" usually backfires
We ran the test so you do not have to. We gave a leading AI assistant a single line: "Help me build a TDABC model for my company." The answer was articulate and the method was right. The problem was everything it made up along the way: it assumed a department costing 560,000 a year, 28 employees, a rate of 0.08 per minute, and activity times of 8, 44 and 50 minutes. None of that came from any real business. It then ended by asking for the data it should have asked for first, which means a second full round of conversation before you get anything usable.
In casual use that is harmless. In finance it is not. A fabricated rate that looks reasonable is exactly the kind of number that survives into a board pack. The discipline that separates a useful answer from a risky one is not the tool. It is the prompt.
Then we gave the same assistant a structured prompt: a clear role, the department's real figures in a short table, a step-by-step instruction to show each formula, and one rule, do not invent anything, flag every assumption. It returned a complete, auditable Time-Driven Activity-Based Costing model in a single pass, with the maths correct and eight assumptions listed in plain sight. Same tool, same task, a different result, because the prompt did the work.
A word on tokens, honestly
You will see claims that good prompts "save 90% of tokens". That is not what we found, and we will not pretend otherwise. A single vague request and a single structured request consume a similar number of tokens. The real cost of a vague prompt is different and worse:
- It wastes a round. A vague prompt has to come back and ask for your data, so you pay for two conversations instead of one.
- It invents figures. The tokens it spends are spent producing numbers you then have to detect and delete.
- A structured prompt lands first time. One pass, correct arithmetic, assumptions flagged, nothing to unpick.
Rough scale, English text: about 4 characters is 1 token, and 100 tokens is about 75 words (source: OpenAI). Portuguese and Spanish use more tokens per word. Exact counts depend on the model, so treat every token figure on this site as approximate.
The prompt library
Every page below gives you a ready-to-copy prompt, a simple worked example so you can see what it produces, an approximate token note, and an anti-hallucination guardrail. Start with the task you have today.
By task
By role
Read this first
Which AI assistant for costing work
Any of these can run the prompts on this hub. They differ in how much data they can hold at once, how well they handle files, and where your data lives. This is a snapshot for June 2026; models and prices move fast, so check the current version before you commit.
| Platform | Maker | Context window | File upload | Best for finance |
|---|---|---|---|---|
| ChatGPT | OpenAI | ~128K | Yes | Broadest ecosystem, custom GPTs for repeat tasks |
| Claude | Anthropic | 200K to 1M | Yes | Largest practical context, strong on long documents and careful reasoning |
| Gemini | 1M | Yes | Huge context, native to Google Sheets and Docs | |
| Microsoft Copilot | Microsoft | ~128K | Yes | Lives inside Excel and the Office files you already use |
| Perplexity | Perplexity | Varies | Yes | Research with cited sources |
| Mistral (Le Chat) | Mistral | 128K to 256K | Yes | Low cost and EU data residency for GDPR-sensitive work |
| DeepSeek | DeepSeek | ~128K | Yes | Very low API cost for high-volume analysis |
| Grok | xAI | 131K to 256K | Limited | Live market and social signal from X |
| Meta Llama | Meta | ~128K | Via tools | Open weight, self-host with no vendor lock-in |
| Notion AI | Notion | Inherited | Yes | Inside your Notion workspace and databases |
Sources: published 2026 pricing and model trackers. Verify the current model and limits before relying on them. We have no commercial relationship with any of these platforms.
And which model within each platform
The table above compares the platforms. Inside the two we use most, the choice does not end there: each offers several models, and the one you pick changes the quality of the reasoning, the speed, and what the run costs you in tokens. These prompts are model-agnostic and run on any capable assistant. We build and test our client models mostly on Claude and ChatGPT, so those are the two we can speak to with confidence. The others work; we simply have less mileage on them for costing.
Claude Anthropic · preferred, most tested
Claude reads a long financial export in one pass and is careful with arithmetic and assumptions, which is what a costing model needs. Pick the tier by how hard the task is.
- Opus 4.8 - the top tier. Deepest reasoning, best for a full multi-department TDABC model, a whale curve across thousands of rows, or a P&L rebuilt across several dimensions. Slower and higher token cost, so reach for it when the model is complex and the answer has to hold up in front of a board.
- Sonnet 5 - the balanced default. Strong quality at good speed and cost. Right for most single-department models and day-to-day analysis. Start here unless the task is unusually heavy.
- Haiku 4.5 - fast and inexpensive. Best for the light, iterative turns: reformatting a table, checking one formula, a quick sensitivity pass. Iterate here, then confirm the final model on Sonnet or Opus.
- Fable 5 - specialised; not the tool for building a cost model.
ChatGPT OpenAI · preferred, most tested
Broadest ecosystem, and it runs the calculation steps directly on an uploaded file, which is convenient for one-pass costing. It offers the same shape of choice as Claude.
- Top reasoning model - for the heavy, multi-step models where the arithmetic and the logic have to be exactly right. Use it when you build the full model or the multidimensional P&L.
- Faster, cheaper model - for iteration and lighter passes, when you are refining tables or testing a single assumption.
- Upload your data and let it run the steps on the file rather than pasting numbers by hand.
Gemini Google
- Very large context and native to Google Sheets and Docs, so it suits teams already in Google Workspace.
- Point it at a Sheet with the same columns the prompt expects.
- We have tested it less for costing than Claude and ChatGPT, so treat its first pass as a draft and check the arithmetic.
Microsoft Copilot Microsoft
- Its advantage is location: it lives inside Excel and the Office files where your figures already sit.
- Good for running a prompt against a workbook without moving data out of your environment.
- Verify the maths as you would with any assistant.
The rest Perplexity · Mistral / Le Chat · DeepSeek · Grok · Llama
These also work, and each has a reason to exist: cited research, EU data residency, low API cost at volume, live signals, self-hosting. For structured costing work they are the "also ran the prompt" tier rather than our tested default. If you use one, keep the guardrail line in the prompt and reconcile the output to your ledger before you trust it.
Why these prompts are different
We are not an AI company. We are cost and profitability specialists who build Time-Driven Activity-Based Costing models for a living, and have done since 2010. Every prompt here is written in the language of the method, tested on a real assistant, and built around the one rule that matters in finance: the model is only as good as the data, and the data has to be yours, not the machine's guess. The prompts get you a credible draft. When you need a model your board, your auditor or your buyer will trust, that is built on your real figures and validated by a person.
From a prompt to a model you can defend
A prompt gives you a structure and a first pass. It cannot see your general ledger, reconcile to your accounts, or stand behind a number in a dispute. That is the work we do. We take the draft an assistant helped you sketch, replace every assumption with your real data, build the model in CostCtrl so it stays alive after the project ends, and hand you a result you can put in front of anyone. If you have used one of these prompts and want the real thing, that is the conversation to have.
Proof
A distributor in New Zealand. €1.335M of cost-to-serve made visible, then halved, and 830 loss-making customers brought down to 295.
Read the case study →Who you would be talking to
Miguel Guimarães, Founding Partner
Cost and profitability practitioner for 25+ years. Presented the Damco cost-to-serve case at Managing for Profit (Amsterdam RAI, December 2009), on the same programme as Robert S. Kaplan.
Call +351 910 313 731
