Plenty of fractional CFOs run their practice on ChatGPT. It was the first assistant many of them used, their clients use it, and it is very good at the work that fills a fractional week: summarising a messy export, drafting a founder update, explaining a variance in plain language, turning a trial balance into a chart before a call.
This guide is for that CFO. It covers what ChatGPT does well across several clients, how to set it up so each client stays separate, and where a structured model layer adds what a chat cannot hold. It is the companion to Claude for Fractional CFOs; the principle behind both is the same, and we call it the CFO harness.
A note on names: OpenAI renamed "connectors" to "apps" in December 2025, and the vocabulary keeps moving. The features below are described by what they do, so check the current label in your plan.
What Does ChatGPT Do Well for a Fractional CFO?
Four jobs, each useful on its own.
Projects: one workspace per client
A Project groups chats, reference files and instructions. Project instructions apply only inside that Project and take precedence over your account-wide custom instructions, so each client can have its own fiscal year, currency and tone without changing your global settings. Projects can also run with project-only memory, drawing on that Project's chats and files rather than your other conversations. For a fractional CFO, that last setting is the one that matters: client A's context should never surface in client B's chat.
Custom GPTs: procedures you repeat at every client
A custom GPT is a saved assistant with fixed instructions and reference files. Build one per procedure, not per client: a "variance commentary" GPT that always ranks by materiality, writes for a founder who reads one page and flags what needs your decision; a "board pack narrative" GPT that follows your section order. The procedure stays constant, and the client-specific inputs come from the Project or the file you give it.
Data analysis: code on your exports
When you upload a CSV or an Excel export, ChatGPT can write and run Python on it. That changes the arithmetic question: the totals come from code, not from the model predicting digits. For one-off work (reconciling two exports, finding duplicate invoices, aging receivables, building a quick chart before a call), it is fast and checkable, because you can ask to see the code.
Apps: where client context lives
Apps connect ChatGPT to documents and tools such as cloud drives, so it can search the engagement letter, last quarter's board deck or the founder's hiring plan without you uploading each file. For the reading-heavy part of the job, this saves real time.
Builders also have Codex, OpenAI's coding agent, for scripting imports and checks. It is useful, not required.
How Should a Fractional CFO Set Up ChatGPT Across Clients?
One Project per client, one custom GPT per procedure, and a written rule about data. In order:
- Decide whose account each client runs on. If the client has a ChatGPT Business or Enterprise workspace, work inside it. If you work in your own account, get written consent on which data goes in. What those clauses usually cover is in Your Client's Contract Says No AI; data terms per plan are in AI and Confidential Financial Data.
- Create one Project per client, with project-only memory. Keep the instructions short: who the client is, fiscal year, currency, materiality, where the model lives.
- Build your procedure GPTs once. Variance commentary, close acceptance checklist, founder update, board narrative.
- Connect documents, not ledgers, by default. Apps for the drive where client documents live. For the books, export reports (trial balance, general ledger) and upload them, unless you have a read-only connection you trust.
- Decide where the numbers live between sessions. This is the step most setups skip, and the subject of the next section.
Where Does ChatGPT Alone Reach Its Limits?
On state, not on intelligence. A fractional practice is not a series of questions. It is four or five models that evolve every month, and a chat is built around conversations, not models.
Uploads are snapshots. Each month you upload a new export. ChatGPT reads it fresh, and the structure (which line is a driver, how the cash flow closes into the balance sheet, why one cost centre is excluded) has to be re-understood or re-explained.
The analysis is rebuilt each time. Data analysis runs real code, which is the right instinct. But the code is written for that conversation. Next month, a new conversation writes new code, slightly differently. Each version is defensible; the series is not consistent. In finance that is called drift, and it is what makes month-on-month comparisons unreliable.
The output is a new file. A generated spreadsheet is a fresh workbook, not an update to the model your client already has. After three months, you have three files and no single history of what changed.
Memory holds conversations, not models. Project memory helps with context ("we agreed to exclude the one-off legal fees"). It does not hold a living model with formulas, timelines and links between statements. We listed the six things that must persist for the next session to start where the last stopped in Persistent Memory for a Financial Model.
There is no gate before the pack. Nothing checks that the balance sheet balances, or that every figure in the commentary exists in the model, before you read the draft. You are the check, every time.
None of this is specific to ChatGPT. Any assistant used as a chat on uploaded files has the same limits. The fix is the same too: put the model, the conventions and the history outside the chat.
How Does ChatGPT Fit Into a CFO Harness?
As the drafting layer. A CFO harness has six parts around the assistant: a persistent structured model, written conventions, scoped permissions, deterministic computation, version history, and review gates. ChatGPT is very good at what sits on top: reading, summarising, drafting, explaining. The harness holds what has to survive the conversation.
Split every deliverable the same way we do in AI for Fractional CFOs:
| Layer | Who does it | ChatGPT's role |
|---|---|---|
| Drafted | the assistant | commentary, summaries, founder emails, first-pass mappings |
| Computed | a structured model with formulas | proposes changes, explains results, never the source of a signed figure |
| Signed | you | none |
Conventions belong with the model, not only in a Project. A FINANCE.md file (currency, units, sign convention, fiscal year, glossary, how the statements connect) travels with the model, so ChatGPT, another assistant or a colleague reads the same rules. Paste it into the Project instructions if you like; keep the master with the model.
Can ChatGPT Connect to a Financial Model Over MCP?
Yes, with conditions. MCP (Model Context Protocol) is the open standard assistants use to call external tools. ChatGPT supports custom remote MCP servers through developer mode, which OpenAI launched as a beta in September 2025 with support for write actions, not only search. Which plans have it, and whether write actions are allowed on each, has changed since; OpenAI's help center currently describes it for Business, Enterprise and Edu workspaces, where an admin enables it. OpenAI also warns that developer mode is meant for people who understand how to configure and test it safely. Check the current page before building a client workflow on it.
When it is available to you, the workflow changes: ChatGPT reads the model and its conventions at the start of a session, loads actuals into the model rather than into a chat, and leaves it in a state the next session finds. When it is not, the fallback still works: export the model to Excel, upload it, and use ChatGPT for analysis and drafting on that copy.
What Should You Keep Out of ChatGPT?
The same five things as in any assistant:
- One client's data in another client's Project.
- Write access to ledgers you do not own.
- Figures you sign that came from a chat, rather than from a model with formulas.
- Credentials pasted into a conversation.
- Client data in your personal account when the contract says otherwise.
Where Layerz Fits
Layerz is a structured spreadsheet, ready for AI. It holds the computed layer: each client's model persists between sessions, with its structure separate from its data, its own FINANCE.md, a deterministic engine for every figure, and a version history with diff and restore.
Layerz's MCP server is a remote server at https://layerz.cc/mcp, so it works with assistants that support remote MCP. In ChatGPT, that means developer mode, where your plan offers it. There is no Layerz app in ChatGPT's directory today. Claude connects natively as a custom connector, which is the setup we document in Claude for Fractional CFOs. Whatever assistant you use, every model exports to clean Excel with formulas intact, which ChatGPT's data analysis reads well.
It is free for one person working alone, with unlimited models.
FAQ
Is ChatGPT good for fractional CFOs? Yes, for the language work around the numbers: summaries, commentary, founder updates, reading inherited files, one-off analysis on exports. It is less suited to holding a living model across months on its own, which is where a structured model layer helps.
Should I use one ChatGPT Project per client? Yes, with project-only memory, so one client's context never surfaces in another's chats. Keep the instructions short and the conventions with the model.
Can ChatGPT's data analysis replace a financial model? For one-off questions, often. For a model you update monthly and sign, no: the code is rewritten each conversation and the output is a new file, so the series drifts and there is no history of changes.
Can ChatGPT connect to QuickBooks or Xero? Through apps or MCP, depending on what your plan offers and what the ledger publishes. On a client's books, use read-only access and let the bookkeeper post.
ChatGPT or Claude for a fractional CFO? Many practitioners use both. The harness matters more than the assistant: a persistent model, written conventions, scoped permissions and a history work with either.
The Takeaway
ChatGPT is a strong drafting partner for a fractional CFO: Projects keep clients apart, custom GPTs keep procedures consistent, data analysis computes with code, apps find the documents. What it does not hold on its own is the model itself, month after month, with its rules and its history.
Use ChatGPT for what it does well. Put the numbers on a structure that persists. Keep the signature. And judge the setup the usual way: is the next client cheaper than the last?