AI in Finance Meetup: Issue 01

Issue 01 · June 2026 · From Gav and Phil · 7 minute read

Welcome to Issue 01, and to the small group of finance teams putting AI to work before everyone else. Here's the one that matters: AI just crossed from chat into real finance work, month-end close through to a first-cut DCF, while pricing flipped to a cheap seat plus a meter on every token.

The power is real, and so is the bill. Which makes the old job the urgent one: knowing who's spending what, on which workflows, before the first surprise invoice lands. Inside: what shipped this half, six CFO reads, what members built, a prompt to pressure-test any call, and the token count where quality quietly slips.

01 · From the meetup

The first CFO Labs session came back to two shifts every finance leader should be tracking. Key learnings: the gains are worth having, and so is some structure around how your team uses these tools and what they cost.

1. The tools now do real work. The leading models can now carry out a sequence of steps on their own (open a file, run a calculation, write the result back), not just answer a question. Anthropic, OpenAI and Microsoft have each shipped finance-specific releases in the last six months or so, from Claude and ChatGPT living inside Excel to packaged plugins for month-end close, variance analysis and valuations.

2. The bill now scales with use. The flat per-seat licence is giving way to a cheaper seat plus metered token usage. A heavy user and a light user now cost very different amounts to run, so unchecked use is a genuine budget risk. Members showed what this looks like in practice, with skills (reusable instruction sets that make the AI do a fixed job the same way each time) and live visuals built on their own data (see section 04).

02 · The timeline

The named releases from the first half of 2026.

03 · The CFO lens

The headlines, translated into the decision sitting in front of finance.

Token spend is becoming its own cost-out exercise. As vendors move from a flat seat to metered usage, spend now tracks how heavily each person uses AI, and it is easy to let it run because nothing limits you the way typing speed once did. One company reportedly spent US$500 million on Claude in a single month after giving staff licences with no usage cap, an extreme case, but the same meter is running in your stack. The cost-out discipline Finance already knows applies: get visibility of usage by team and workflow, set a budget, and hunt for the standard processes you can make cheaper to run. Decide now who owns the forecast for this line, then watch it like any other variable cost.

Claude and ChatGPT now run inside Excel. Both Anthropic and OpenAI now put their models inside the spreadsheet, so you can build, audit and document a model where the work already lives rather than copying outputs around. Treat the output as a draft to check, not a finished answer: with that caveat, one example in the room took an information memorandum to a first-cut DCF in minutes, then used the Excel version to audit it and restyle it to a house layout.

Your accounting system can talk to AI directly. Xero now has a connector that lets AI read and work with your live ledger, so you can ask for a forecast off the last three years, or pull cash-flow and tax views, without exporting anything.

One attendee runs it across several entities for about ten dollars an entity each month in connector fees, on top of the Xero subscription and the build-and-maintain time (the AI model usage is billed separately). The connector runs locally, but where it is paired with a cloud model the ledger data it reads is sent to that model's provider; it only stays fully on-device with a self-hosted model. Reconciliations still need a human, but for analytics and forecasting it strips out a lot of the manual data wrangling.

The big firms are now partnering to get this into the enterprise. The model makers have realised that getting AI to actually run inside a business is harder than shipping the model, so they are pairing with the big accounting firms to embed it (KPMG and Anthropic, recently). For a mining or energy CFO the lesson is that the cost is not the licence, it is the integration: plugins install in minutes but need real customisation and testing before they fit your processes. Budget the implementation, not just the subscription.

You can keep sensitive data in-house and still use AI. NVIDIA showed an enterprise agent toolkit at GTC that runs on your own hardware with self-hosted models, alongside new chips (with Microsoft and PC makers) aimed at running AI locally. For anyone weighing confidentiality against capability, it points to a split: keep sensitive work on-premises and route only the rest to the top cloud models. It is early and far from plug-and-play, so work out where you genuinely need it first. The governance question to settle before any of this: who signs off that a model’s output is right, and where is that recorded?

Newer is not always better, so test before you switch. A model upgrade can break a workflow that was running fine: one presenter found a new release introduced odd mistakes in tasks that had worked before, only fixed in the version after. In Finance, where a quiet error carries real consequences, validate a new model against your own numbers before you roll it out: keep a small set of figures you already trust, re-run them on any new version, and switch only when they still tie out.

04 · Show and tell

Working examples from the session, shared so you can borrow the idea. Builders are not named here; if you would like to be credited or connected, reply and we will introduce you.

A skill that fixes board papers after everyone has touched them. One member built a skill for the job no one enjoys: when several people have worked on a board paper, the fonts drift, the tables break and the numbers stop reconciling. The skill runs once at the end, aligns the formatting to a set house standard, then reviews the document section by section, without changing a word. The gain is time: it takes a tedious manual pass off the critical path before papers go out.

Live tax and cash-flow visuals on top of the Xero connector. Another member designed a set of live financial visuals for a client whose group spans two Australian trusts, a New Zealand company and personal returns, all run through separate Xero instances. The connector does the plumbing; on top of it they built tax bridges and cash-flow diagrams that read live from the ledgers. The connector and the visuals run on the client's own computer, though where a hosted model is used the ledger data it processes does travel to the model provider. They reckon the connector covers roughly 80 per cent of what you would do in Xero directly, reconciliations aside.

Coding the skeleton of a large cost model with Claude. A member coded, with Claude’s help, the groundwork for a big mining cost model: prices, equipment and the various splits across the operation, all worked into one structure of dimensions. It maps the different paths so you can see what is active, what is dormant and what has been removed. A useful pattern for anyone wrangling a model whose dimensions have grown hard to keep straight.

A searchable library that answers from your own documents, cheaply. Another member built a knowledge database over hundreds of PDFs. It chunks each document by chapter and page into markdown first, then, when asked a question, searches for the most relevant chunks and answers from just those, returning the chapter and page so you can check the source. What it saves is token cost: you reason across the few passages that matter rather than loading the whole library into the context window.

And on a lighter note, the Model Answer team showed MOLYCON Rumble, a mining-themed arcade game, proof the tools are good for more than spreadsheets.

05 · Try this

When you have a thesis you are not yet sure of, ask the AI to stand up a panel of experts and have them debate it, dropping any position that cannot be defended until a conclusion settles out. Give it the full context first, your idea and where you are leaning, then tell it to prove you wrong. What it improves is decision quality: it surfaces the holes a subject-matter expert would find.

One caution from the session: it is excellent when you are the domain expert and can interrogate the argument, and risky when you are not, because a confident case can talk you down the wrong path. Bring your own view first, then use it to stress-test, not to decide for you.

Prompt to copy and paste:

Swap the closing line for your actual question, and paste your context above it.

06 · The number that matters

Resources members shared during the session. New to this? Start with Xero’s AI toolkit, the last one below.

The complete guide to building skills for Claude. Anthropic’s guide (around 30 pages), light on text, including how to bring MCPs into a skill. There is also a built-in skill that helps you create skills.

Xero MCP server. The official open-source connector that lets Claude read and write Xero data (accounts, invoices, reports, payroll). Free and MIT-licensed, on top of the Xero subscription and per-connector cost.

[Start here] Xero’s AI toolkit. The easiest place to begin: Xero’s developer hub for AI builds, with the MCP server, a prompt library, and a step-by-step video for setting it up with Claude.

Here’s a copy of the Meetup Presentation in case you weren’t able to make it. We’ve included recent news, as well as some tips and tricks.