A CFO harness is everything around an AI model that makes its finance work signable. It has six parts: the structured financial model the AI works on, the conventions it reads before touching anything (written in a file such as FINANCE.md), the permissions that bound what it can read and change, a deterministic engine that does the arithmetic, a version history that records every change, and the review gates a person passes before a number leaves the building. The model drafts. The harness decides what counts.
That is the definition we use. The rest of this page explains why finance needs its own, what goes in each part, and how a fractional CFO assembles one across several clients.
Where Does the Word "Harness" Come From?
From software. Engineers who build coding agents call the code around the model its harness: the loop that feeds it context, the tools it can call, the files it reads at the start of a session, the tests that tell it whether it broke something. Anthropic's engineering team described one in November 2025, in Effective harnesses for long-running agents. The core problem, they wrote, was "finding a way for agents to quickly understand the state of work when starting with a fresh context window." Their answer was not a better prompt. It was a progress file, a feature list, and git history the agent reads at every start.
In 2026 the word reached finance, through three different doors:
- Governance. PYMNTS wrote in April that CFOs are turning to AI harnesses: "The harness is not the model itself, but the system that governs how models act in the real world." For a CFO, the article adds, the useful way to see it is "as an extension of financial controls."
- FP&A practice. An AFP FP&A on-demand session sponsored by Ramp, presented by Glenn Hopper, sets out to "define an 'agent harness'" and show how it makes probabilistic agents reliable enough for finance, using a 13-week cash flow forecast as the worked example.
- Risk. Cobo's deep dive, Why Finance Needs a Different AI Harness (May 2026), starts from a formula, "Agent = Model + Harness", and argues that finance flips the priorities of coding.
Three sources, three angles, no shared definition. That is normal for a word in its first year. It is also why a definition by components is worth writing down: it works whichever assistant you use, and it gives you a checklist rather than a slogan.
Why Does Finance Need a Different Harness Than Coding?
Because a coding harness is built on the assumption that mistakes are cheap to undo, and in finance they are not. Cobo puts it in one line about financial systems: "There is no safe retry, no rollback to a previous state." Four differences follow from that.
1. There is no compiler. Code that is wrong usually fails loudly: a test goes red, the build breaks. A financial model that is wrong looks exactly like one that is right. The formula computes, the total foots, the chart renders. The error is in an assumption, a sign, a unit, or a line that silently dropped out. Nothing turns red.
2. The output leaves the building. A bad commit can be reverted before anyone outside the team sees it. A forecast sent to a lender, a board pack, a covenant certificate: once those are out, the correction is a conversation, not a revert.
3. State lives in the wrong format. A coding agent's work lives in text files with a history (git) that shows what changed, line by line. A financial model usually lives in a spreadsheet, a binary file where nothing records which cell moved, when, or why. The agent cannot "read the git log" of a workbook. That piece of the harness has to be built.
4. Someone signs. In software, the team owns the code. In finance, a named person owns the number. The harness has to produce something that person can stand behind: each figure traced to a source and a formula, each change attributed.
The most common failure in practice is not dramatic. A survey by The Harris Poll for Zuora, reported by CFO.com in April 2026 (321 US finance decision-makers), found that the top-cited gap between AI's promise and reality was "difficulty integrating AI outputs into finance workflows", at 41%. Another 35% pointed to AI insights that do not match the data in core finance systems. Both are harness problems. The model produced something; nothing around it gave that output a place to land or a way to be checked.
What Are the Components of a CFO Harness?
Six, and none of them is a prompt. Each one answers a question you would otherwise answer by hand, every session.
1. A persistent, structured model
Question it answers: what are we working on, and where did we leave it?
The AI works on a model that exists outside the chat: named variables, formulas that reference each other, timelines, the links between the P&L, the cash flow and the balance sheet. It reads that model at the start of every session instead of reconstructing it from a pasted export. Chat memory remembers a conversation about the model; the harness needs the model itself. We wrote up what has to persist in Persistent Memory for a Financial Model.
The second property is separation. Structure (logic, formulas) and data (actuals, scenarios) are kept apart, so the AI can load new numbers without touching the logic, and you can reuse the logic without carrying the numbers.
2. Written conventions
Question it answers: how do I read these numbers?
Currency, units, sign convention, fiscal year, what "gross margin" means in this company, how the statements connect. An agent with no conventions does not hesitate; it assumes, consistently, and next session it may assume differently. The conventions go in a file that lives with the model and is read before anything else. We published FINANCE.md as an open format for that file: machine-readable front matter, human-readable body, MIT license.
3. Permissions
Question it answers: what is this agent allowed to touch?
Read-only by default on anything you do not own, such as a client's ledger. Write access only where the agent is supposed to work, and scoped to that place: one model, one client. PYMNTS describes this part of the harness as "the mechanism that enforces permissions, logs actions and ensures accountability." In practice it is a short list per client: which connections exist, which are read-only, whose account they run on.
4. Deterministic computation
Question it answers: who did the arithmetic?
Not the language model. A model asked the same question twice can answer twice differently; a formula cannot. In a CFO harness, totals, balances, ratios and forecasts come from an engine that evaluates formulas the same way every time. The AI proposes the formula, explains the result, drafts the commentary. It does not compute the number you sign. This is the part most often skipped, because a chat that returns a confident figure feels like it did the work.
5. Version history
Question it answers: what changed, when, and who changed it?
Every change to the model is recorded with its author (you, a colleague, or the agent) and a reason. You can compare two versions, see the semantic difference ("revenue growth assumption changed from 8% to 12%", not "cell D14 changed"), and restore an earlier state. This is the finance equivalent of the git log in Anthropic's coding harness, and it is what turns "the AI changed something" from a worry into a line you can read.
6. Review gates
Question it answers: what must be true before a person reads this, and before it leaves?
Two gates. Before you read: automated checks (the balance sheet balances, cash closes into the balance sheet, no figure in the commentary that is absent from the model, no hardcoded value where a formula belongs). Before it leaves: your sign-off, on a version you can name. Changes the agent proposes can be previewed before they apply. AI drafts can go anywhere; only a reviewed version goes to the client.
| Component | Without it, you... | With it |
|---|---|---|
| Persistent model | re-explain the model every session | the session starts where the last one stopped |
| Conventions | correct the same sign and unit errors | the agent reads the rules before it acts |
| Permissions | hope nothing writes where it should not | access is scoped and visible per client |
| Deterministic computation | re-check every figure the chat returned | figures come from formulas you can inspect |
| Version history | ask "what changed?" and get a guess | you read the diff |
| Review gates | review everything, every time | you review exceptions and sign a named version |
Is a CFO Harness the Same Thing as a Good Prompt?
No. A prompt is an instruction for one session; a harness is what persists between sessions and holds regardless of the prompt. You can write an excellent prompt explaining your conventions, and many people try that first. It works until the session ends, the prompt is edited, or someone else picks up the model. The harness keeps the conventions, the structure and the history outside the conversation, where a new session, a different assistant or a colleague finds them in the same state.
Is a CFO Harness an "AI CFO" Product?
No. AI CFO products are applications that produce reporting or forecasts for a company, often a good fit for founders who need numbers before they need a finance lead. A CFO harness is the environment a finance professional builds around whichever assistant they already use, so that its work meets their own standard. You can assemble one from several tools. The question is not which product, but whether each of the six components exists somewhere and whether you can show it.
How Does a Fractional CFO Assemble a Harness Across Clients?
By splitting it in two: the half you carry between clients, and the half that stays with each client. Build the first half once, as a template. Instantiate the second half per engagement.
What you carry (your method):
- the model structure you start from (P&L, cash flow, balance sheet and how they connect)
- a FINANCE.md template with your defaults, adjusted per client
- your written procedures: the close acceptance, the variance commentary, the board pack (as Claude Skills, custom GPT instructions, or plain checklists)
- your check list: what must be true before you read a draft
What stays with each client (their instance):
- the data, in that client's own model
- the connections to their ledger and bank, read-only, on their accounts where possible
- their FINANCE.md, with their fiscal year, their definitions, their materiality
- the history of their model, which they keep when you leave, along with a clean Excel export
The cost of the harness is front-loaded: the first client takes the longest. If the second client takes as long as the first, the method is not portable yet. If the fourth is cheaper than the second, it is working. That is the same test we apply in AI for Fractional CFOs, and it is the reason a harness matters more to a fractional CFO than to anyone: you rebuild context four times a month, and the harness is what makes the fourth time cheap.
For the assistant layer itself, we wrote two practical setups: Claude for fractional CFOs and ChatGPT for fractional CFOs. The harness is the same in both. The assistant is the replaceable part, which is the point.
A Self-Check: Do You Have a Harness Yet?
Open a fresh session with your assistant, on a client you have not touched in two weeks, and ask:
- What is this model for, and what currency and sign convention does it use?
- What feeds gross margin?
- What changed in the model since last month, and who changed it?
- If I change the hiring plan, which figures in last month's pack would move?
- What is the agent allowed to change in this client's ledger?
If the answers come from the harness, you have one. If they come from you, re-explaining, you have a prompt.
Where Layerz Fits
Layerz is a structured spreadsheet, ready for AI, and it covers four of the six components in one place. The model is persistent and structured, with structure kept separate from each client's data. Each model carries its FINANCE.md, which the agent reads before acting. Every value comes from a deterministic engine, never from the language model. Every change is versioned with its author, with a semantic diff and restore; on the free plan the history covers 24 hours plus one pinned version, on Team it covers 30 days.
It also helps with the other two. Permissions: API keys can be scoped to one model and set to read-only, and the agent writes only through tools that declare what they change. Review gates: built-in checks (a balance sheet that must balance, flagged when it does not), a validation report, and a dry run that previews a change before it applies.
What Layerz does not cover: your client's ledger (it reads from Pennylane, Qonto and Stripe connectors, or from exports), and your signature. You bring your own agent: Claude drives Layerz over MCP, and any client that supports remote MCP servers can connect. Every model exports to clean Excel with formulas intact, so the exit stays open. It is free for one person working alone, with unlimited models.
FAQ
What is a CFO harness? The layer around an AI model that makes its finance work signable: a persistent structured model, written conventions, scoped permissions, deterministic computation, version history, and review gates. The model drafts; the harness decides what counts.
What is an AI harness for finance? The same idea as a coding agent's harness (the code that feeds an agent context, tools and tests), adapted to a field where errors are not reversible, there is no compiler, and a named person signs the output.
Do I need to code to build a CFO harness? No. The components can come from tools you configure rather than code you write: a project per client in your assistant, a conventions file, read-only connections, a model on a structured layer with history. Coding helps for custom procedures, but it is not the entry ticket.
Which AI assistant works best with a CFO harness? The harness is meant to make the assistant replaceable. Claude, ChatGPT and others can all sit on top of it, through MCP where the client supports remote servers, or through exports where it does not.
Is "CFO harness" a standard term? Not yet. "Harness" appeared in finance press and training in 2026 without a shared definition. We define it by its six components so that it can be checked, not just said.
The Takeaway
The interesting part of an AI-native finance practice is not the model. Models change every few months, and the one you use today will be replaced. What stays is what you built around it: the structure, the conventions, the permissions, the engine, the history, the gates.
Coding agents got reliable when engineers stopped polishing prompts and built harnesses. Finance is at the same point, with higher stakes and no compiler. Let the model draft. Let the harness decide what counts. Keep the signature.