Your Biggest Decision This Quarter Doesn't Need a CFO. It Needs Three AI Agents.

You already have a finance department. You are just doing all three jobs yourself, at nine at night, with a spreadsheet that argues with itself.

The Decision Every Growing Small Business Eventually Faces

It is nine at night, and Priya is still at the register table in her fifth café, laptop open, three tabs of spreadsheets fighting each other.

She runs five bakery cafés across Manchester and Leeds. A landlord just offered her a sixth site: good foot traffic, six months rent-free for the fit-out, decision needed by Friday.

She knows her revenue. She does not know, not with any real confidence, whether store six pays for itself in eighteen months or drags the other five down with it.

So she does what most owners do. She guesses, prices in a margin of fear, and either signs the lease anxious or walks away from a good site because the maths took too long to trust.

This is not a technology problem. It is a staffing problem. A business Priya's size does not have an FP&A analyst modelling the scenario, or a CFO stress-testing the assumptions before she signs anything. She has herself, a spreadsheet, and a deadline.

What Is a Multi-Agent AI Finance Team?

A multi-agent AI finance team is a set of specialised AI workers, built inside a tool like Claude, that each own one stage of a financial decision and hand their finished work to the next worker in line, the way an actual finance department would.

Not one assistant answering one question. Three workers, each with a defined job, passing a file forward until a decision comes out the other end.

The clearest working example right now comes from Luke Finance, a quant and finance-automation builder who published a full walkthrough titled "I Built a Complete AI Finance Team With Claude" on 27 July 2026. It has passed 27,000 views. He built three workers:

• Finance Analyst: reads the current business (store performance, revenue, margins) and reports on financial health.
• FP&A Strategy Partner: takes that report and models the decision. In his example: three expansion scenarios, five, ten, and fifteen new stores, scored on ROI, NPV, IRR, and payback period.
• CFO Advisor: reviews both outputs against the company's own investment policy and delivers one of three verdicts: proceed, proceed with conditions, or do not proceed, with every claim traced back to a worker's evidence.

A fourth file, the orchestrator, is not a finance role at all. It just runs the other three in the right order, and will not let the next worker start until the previous one has actually delivered its report.

You Are Not Short on Data. You Are Short on a Finance Department.

Priya has more usable data than she thinks. Point-of-sale numbers by store. A spreadsheet of monthly costs. A rough sense of what a good site looks like.

What she is missing is not the data. It is the three-step relay that turns data into a decision: read it, model it, judge it.

Right now she does all three steps herself, in one tired pass, with no one checking her assumptions. That is the actual gap. Not a data gap. A department gap.

Build the department instead of hiring it.

Why This Is Possible Now

Two things changed in the past few weeks that make this a small-business tool instead of an enterprise one.

First, Claude can now hold an entire project, source files, worker instructions, and a running record of what each worker has produced, in one place, and reopen it days later without you re-explaining anything.

Second, the accounting industry itself is racing to build this same pattern at enterprise scale. Billow, an AI-native accounting firm backed by Y Combinator's Summer 2026 batch, is positioning itself publicly as a challenger to Big Four-style bookkeeping and audit work. Its co-founder, Joanathan McIntosh, wrote in early August that "workflows belong to agents, not people": that accounts payable, vendor validation, and month-end checks are moving from living in one employee's head to running as defined, agent-owned processes, with a human doing oversight instead of execution.

That shift is happening at firms with dozens of accountants on staff. There is no reason the same pattern cannot run at a business with five cafés and no accountant at all. Workers with defined jobs, handing off structured output, just need a project folder and an afternoon, not a finance department's budget.

Isn't This Just Claude Answering a Question?

No. That is the mistake worth naming.

Ask Claude in one prompt, "should I open a sixth café," and you get a plausible-sounding paragraph. It has no memory of your investment policy. It has not checked your actual numbers. It cannot tell you why it landed where it landed, because there is no trail.

A worker team is different because it forces a paper trail. The Finance Analyst has to write a report before the FP&A worker is allowed to touch it. The FP&A worker has to model real scenarios against real policy thresholds before the CFO Advisor is allowed to weigh in. Nothing skips a step, because the orchestrator will not let it.

The result is not an opinion. It is a recommendation you could defend in front of a landlord, a bank, or a business partner, because every number in it traces back to a specific worker's output.

How to Build Your Own Three-Worker AI Finance Team

You do not need to code this. You need one Claude project, five folders, and an afternoon.

1. Set up one project folder with five areas. Company Context (your numbers policy: how much return you need, how fast you need to earn it back), Source Data (wherever your numbers already live: a spreadsheet, Airtable, a QuickBooks export, a Notion page, it can stay there, it does not need to move), Workers (the three role files below), Worker Outputs (where each finished report lands), and Final Deliverable (the one-page recommendation).
2. Write the Finance Analyst worker. Its only job: read your current numbers, revenue, costs, margin by location, and write a plain report on how healthy the business is right now. Nothing about the new decision yet.
3. Write the FP&A Strategy Partner worker. It reads the Finance Analyst's report, then models the actual decision in front of you. In Priya's case: one new site versus staying at five, against clear return and payback thresholds she sets in advance.
4. Write the CFO Advisor worker. It reads both prior reports and your investment policy, then commits to one answer, proceed, proceed with conditions, or do not proceed, with every claim linked back to a specific number from the earlier two workers.
5. Write one orchestrator file. Its only job is sequencing: run the Finance Analyst, confirm its report actually exists, then run FP&A, confirm that report exists, then run the CFO Advisor. No worker starts until the one before it has actually delivered.
6. Run it end to end from a blank Claude session. Attach the project, give it one instruction, "read the master file, take over the orchestrator, and run the team from the top," and let it work through discovery, analysis, modelling, and recommendation without you touching any of the three worker files yourself.

In the original build, this produced a twelve-section store analysis, a ten-scenario expansion case, and a board deck that put the answer, "proceed with conditions," on slide two, with the phased rollout and risk mitigations right behind it. Priya's version does not need twelve sections. It needs three: is the business healthy, does site six clear her bar, what is the honest risk. That is a Friday-afternoon build, not a quarter-long project.

If reading your own numbers into a project folder still feels like a leap, our piece on automating the weekly report is the smaller first version of this same idea: one worker instead of three.

What You Have After This

A week from now, you have a Finance Analyst worker that already knows how to read your numbers, on call whenever you need a health check.

A month from now, you have run it against one real decision, a hire, a lease, a loan, and you have a written recommendation to show for it, not a gut feeling.

Six months from now, you are not building a new worker team for every decision. You are rerunning the one you already built, on new numbers, in minutes.

Same owner. Same business. A finance department you built in an afternoon and never had to hire.