ChatGPT for Finance Teams: Setup and Workflows

A Leeds-based private equity firm managing eight portfolio companies came to us with a familiar problem. Their finance team was producing monthly commentary packs for each portfolio company: four to five pages of variance analysis, KPI commentary, and executive summary per entity. With eight entities, that was 32-40 pages of bespoke narrative every month. Two analysts spent roughly four hours each on commentary alone, before review.

They had ChatGPT Team licences. They were using it occasionally, mostly to rephrase sentences. The setup was wrong. Within three weeks of restructuring their ChatGPT configuration, they cut monthly commentary time to under 45 minutes across all eight entities.

The difference was not the model. It was Projects, a portfolio-specific Custom GPT, and a consistent system prompt that enforced IFRS presentation standards and the firm's house style. This guide covers how to replicate that setup for your finance team.

Which ChatGPT Plan Finance Teams Need

The free and Plus tiers of ChatGPT are not appropriate for real financial data. Both may use your conversations to train future models unless you actively disable this in settings — a setting that resets under some conditions and that new team members are unlikely to configure correctly. For confidential financial data, you need a plan that turns off model training by default and keeps data within defined boundaries.

The PE firm used Team licences for the finance team (eight analysts, two finance managers) with Enterprise reserved for the investment professionals handling price-sensitive deal data. This is a sensible split for most financial services contexts.

Before buying additional licences, check whether your organisation has an existing OpenAI enterprise agreement. Large firms often have enterprise licences already provisioned through IT that finance teams are unaware of.

Setting Up ChatGPT Projects for Finance

ChatGPT Projects creates a persistent workspace that holds files, instructions, and conversation history across sessions. Without Projects, every ChatGPT session starts blank. You paste in context, explain the task, and get a result. The next session, you repeat the same pasting. Projects eliminates that overhead.

What to include in a Finance Project

A well-configured finance Project includes three types of content: standing context, current-cycle data, and instructions.

Project structure for multi-entity finance teams

The PE firm created one Project per portfolio company. Each Project held the entity-specific chart of accounts, prior-year results, and budget assumptions. Monthly, analysts added the current-period actuals to their entity Project and ran the commentary prompts.

An alternative structure is a single Project for the consolidated group, with entity-specific context loaded as separate files. This works for simpler consolidations but becomes unwieldy when entities have significantly different cost structures or reporting requirements.

For most finance teams, one Project per major reporting entity is the right structure. For month-end close workflows specifically, also see our guide to month-end process audit findings for what typically holds teams back before AI can help.

Building Custom GPTs for Finance Tasks

Custom GPTs are pre-configured assistants that any team member can invoke without setting up context from scratch. You build the GPT once, configure its system prompt, upload its knowledge files, and share it with the team. Anyone using it gets a consistent, correctly-configured starting point.

This is the difference between a junior analyst copying a prompt from a shared document (inconsistent, often wrong) and invoking a Custom GPT that has the prompt, the style guide, and the relevant context already embedded.

Finance Custom GPTs worth building

Sharing Custom GPTs with your team

In ChatGPT Team, Custom GPTs can be shared within your workspace so every team member with access can use them. Build the GPT in one account, publish it to the workspace, and team members find it in their GPT list under your organisation.

Governance note: document who owns each Custom GPT and what its knowledge files contain. When accounting policies update or the house style guide changes, the Custom GPT needs to be updated too. Assign ownership to a specific person rather than leaving it unmanaged.

Using ChatGPT Tasks for Scheduled Finance Workflows

ChatGPT Tasks lets you schedule prompts to run at set times without manual triggering. For finance teams, this is most useful for recurring analytical tasks where the structure is fixed but the content changes each period.

What Tasks can and cannot do

Tasks run a configured prompt at a scheduled time and deliver the output to your ChatGPT inbox. They work well for: weekly commentary drafts, daily cash position summaries (where you provide the data manually), and periodic review reminders with structured checklists.

Tasks cannot pull live data from your ERP, accounting system, or bank feeds automatically. The scheduled prompt runs, but it only has access to what you have already provided in the Project or Custom GPT configuration. If your use case requires live data ingestion, you need an integration layer — n8n, Make, or Power Automate connecting your data sources to the ChatGPT API — before the output reaches the analyst.

Useful scheduled tasks for finance teams

The PE firm used a weekly Task to generate a one-page summary of each portfolio company's current trading position. Analysts added the weekly management accounts export to their entity Project on Friday afternoon. The Task ran on Monday morning and produced a structured summary the investment team read over coffee. Review of eight entities took 20 minutes per week instead of two hours.

If your team is also evaluating Claude, it handles equivalent scheduled workflows through Cowork — its background agent mode — paired with a scheduling layer in n8n or Make. The setup requires slightly more configuration than ChatGPT Tasks but supports more complex multi-step automations. See our ChatGPT to Claude migration guide for a full comparison of how Tasks and Cowork differ in practice.

System Prompt Structure for Finance

A system prompt is the instruction set that shapes how the GPT behaves. Whether you are configuring a Custom GPT or a Project's instructions, the system prompt structure determines whether outputs are consistent and useful.

Poor system prompts say things like: "You are a finance assistant. Help with financial analysis." This produces generic outputs. Good system prompts define role, scope, constraints, and output format with enough specificity that a new team member reading the output could not tell it from a senior analyst draft.

ROLE
You are a senior financial analyst for [Company/Entity name].
You write variance commentary for management and board reporting.
You follow IFRS presentation standards and [Company] house style.

CONTEXT
Reporting currency: GBP. Materiality threshold: £[X]k.
Reporting structure: [cost centres / P&L categories].
Audience: [Finance Director / CFO / Board].
Format: [Management accounts commentary / Board executive summary].

OUTPUT REQUIREMENTS
- One paragraph per major P&L section (Revenue, COGS, Gross Margin, Opex, EBITDA)
- Each paragraph: 3-4 sentences maximum
- Lead with the variance (amount and % vs budget)
- Explain the primary driver in commercial terms
- Note any forward-looking implication if material
- Flag any items requiring management attention as [FLAG]
- Do not use em dashes. Use commas or full stops.
- Do not speculate on causes not evidenced in the data provided.

CONSTRAINTS
- Never invent figures not present in the data
- If a figure is unclear or missing, state this explicitly
- Do not use jargon the Board would not recognise
- Present in narrative paragraph form, not bullet points

The PE firm's system prompt added three portfolio-specific rules: reference EBITDA multiples context when commenting on major opex lines (relevant for the investment team readers), flag any items with covenant implications separately, and always reference the prior quarter as well as prior year for YoY-plus-QoQ context.

These specifics are what turn a generic finance GPT into something the team relies on. Build the first version with the standard structure, then update it each time a reviewer makes the same correction twice. If the Finance Director edits every draft to add prior quarter comparison, add that to the system prompt. Within two cycles, the edits should be minimal.

PE Firm Implementation: 4 Hours to 45 Minutes

The Leeds private equity firm had the right tools and the wrong setup. ChatGPT Team licences, eight portfolio companies, two senior analysts producing monthly commentary packs. Each entity required five to six pages of narrative commentary monthly. Total monthly analyst time on commentary: eight hours combined.

We restructured their ChatGPT setup over three weeks. Eight entity-specific Projects were configured with standing context: each entity's chart of accounts, prior-year P&L for YoY context, and the firm's commentary style guide. One Portfolio Commentary Custom GPT was built with the system prompt structure above, including IFRS presentation rules and the firm's covenant-flagging requirement.

Monthly workflow for each entity became: export actuals vs budget from the portfolio company's accounting system (CSV), upload to the entity Project, run the Commentary Custom GPT with a one-line prompt naming the period. Total prompt time: under five minutes per entity. Review and edit: ten to fifteen minutes per entity. Total across eight entities: under 45 minutes.

The investment team noticed the commentary improved. Not because ChatGPT writes better than a senior analyst, but because the system prompt enforced consistency that manual drafting never achieved: every entity commentary had the same structure, the same materiality threshold applied, and no analyst forgot to include prior-quarter comparison.

The time saving was approximately seven hours per month across the team. That time went to deeper variance investigation on the two or three entities with material exceptions, and to the quarterly LP reporting pack that had previously been rushed. For a full comparison of ChatGPT against Claude and Copilot for finance tasks, see our tool comparison guide.

Frequently Asked Questions

Related guides