AI Agents for FP&A Teams: Automate the Board Deck and Spend Tracking

Priya Raman is the entire FP&A function at a 45-person Series B SaaS company doing $9.2M ARR. Every first Tuesday of the month she spends three days, roughly fourteen hours total, pulling numbers out of Stripe, the payroll system, and four disconnected spreadsheets to assemble the board deck. By the time the deck ships, the runway model buried inside it is already wrong: it assumed the two open sales hires from last week's headcount plan, not the three the CEO approved two days ago. Meanwhile $38,000 a month in AWS, OpenAI, and Anthropic invoices sits in her inbox with no attribution to a team, a feature, or a customer segment, so when the board asks where the AI spend is going, the honest answer is "we're not sure yet." This is not a hypothetical. It is the default state of FP&A at almost every company between Series A and Series C: one person doing the job three people used to do, with tools built for a slower reporting cadence than the board now expects every single month.

Why a generalist chatbot and an autocomplete tool both miss the job

Paste last month's numbers into ChatGPT or Claude.ai and it will happily write board-deck narrative around them. What it will not do is remember that the hiring plan changed since last month, hold a persistent model of the P&L across sessions, or notice that the runway assumption baked into March's deck is now stale in April. Every conversation starts from zero. You are the one who has to re-explain the burn rate, the headcount plan, and the variance drivers each time, which means the generalist tool is doing the narrative-writing step, the easiest ten percent of the job, while Priya is still doing the hard ninety percent: pulling the actual numbers, reconciling them across systems, and deciding what the board needs to see.

Cursor and GitHub Copilot solve an adjacent but different problem. FP&A teams increasingly write the SQL that pulls MRR cohorts out of the warehouse or a short Python script that projects burn under different hiring scenarios, and Copilot is genuinely useful for autocompleting those lines. But it has no concept of what a healthy runway number looks like, no memory of the board's last variance question, and no mechanism for turning a script's output into a package a board trusts. It will finish your pandas groupby. It will not flag that your churn assumption is six months stale, and it will not produce the monthly board financials on its own.

The mismatch is structural, not a matter of the tool trying harder. FP&A work is not writing prose or writing code, it is holding a live financial model in your head across cash, headcount, revenue, and burn, and translating it into a package the board can trust on a fixed monthly cadence. That requires an agent built for the finance-ops role specifically, one that starts every engagement with a recon of the actual books rather than whatever numbers you happened to paste in that session.

There is also an institutional-memory problem neither generalist tool solves. A CFO or a solo FP&A hire carries context across months: which department is over plan, which variance was already explained to the board last quarter, which spend line is a one-time cost versus a new baseline. A chatbot session resets that context every time. An autocomplete tool never had it in the first place, it only ever saw the fragment of code you had open. Getting that continuity back with a generalist tool means re-typing the same background brief every single month, which defeats the point of automating the cycle at all.

Meet Mint, with Lens and Budget alongside it

Mint is Tonone's finance engineer for Claude Code: P&L, runway, unit economics, fundraising materials, and board reporting. Where a generalist chatbot treats each finance question as a fresh conversation, Mint starts from the actual state of the books. The mint-recon skill audits the current P&L, burn rate, runway, and unit economics before anything else runs, so the board package and the runway model are both grounded in where the constraint actually is, not in whatever numbers happened to be pasted into the prompt that day.

From that baseline, mint-board produces the monthly or quarterly financial package: P&L, cash position, key metrics, and variance against plan, in the shape a board actually reads. mint-runway calculates current runway from cash and burn rate and, critically, models the levers available to extend it when a hiring plan changes mid-month, which is exactly the failure mode that made March's deck wrong by the time it shipped.

Tonone's Mint runs a financial recon before every board package or runway model, so the numbers reflect the current state of the books, not last month's assumptions.

Lens keeps the metrics dashboard from going stale between board meetings

Mint's board package is only as current as the last time someone rebuilt it by hand. Lens, Tonone's data analytics and BI engineer, closes that gap. lens-dashboard specs and builds the recurring metrics dashboard, defining the question each chart answers, writing the underlying SQL, and setting the refresh cadence, so the ARR, churn, and burn numbers Mint pulls into the board deck are live rather than reconstructed from a spreadsheet someone last touched six weeks ago. For a 45-person company running FP&A with one person, the difference between a dashboard that refreshes automatically and one that gets manually rebuilt every month is the difference between fourteen hours of prep and ninety minutes of review.

Budget attributes the AI and cloud spend the board keeps asking about

The $38,000 a month sitting unattributed across AWS, OpenAI, and Anthropic invoices is Budget's problem to solve. Tonone's AI cost engineer, Budget, runs budget-recon to map the cost topology: billing attribution by team, spend forecast versus actuals, and where the alerting gaps are so a spend spike doesn't surface for the first time in a board meeting. budget-audit goes further, breaking the spend down per model and per feature, identifying the top consumers, and surfacing waste, a support-bot integration calling GPT-4-class pricing for a task that a smaller model handles just as well, for instance. That per-model, per-team breakdown is what turns "we're not sure yet" into a specific line item Priya can put directly into the board deck's spend section.

A worked example: the Tuesday board deck that used to take three days

Back to Priya's company. The board meets the first Tuesday of every month. Historically, prep started the prior Thursday: pull Stripe MRR by hand, reconcile it against the payroll system for headcount cost, rebuild the runway tab in a spreadsheet that already had last month's hiring assumptions wrong, and manually total the AWS, OpenAI, and Anthropic invoices into a single "cloud and AI spend" line with no further breakdown. Fourteen hours across three days, and the resulting deck was stale on the runway line before it was even presented.

With Mint, Lens, and Budget running the same cycle, the sequence looks different. Priya runs mint-recon Thursday morning; it pulls the current P&L, confirms burn rate against the general ledger, and flags that headcount assumptions changed since the last model run. mint-runway recalculates runway under the current plan and under a second scenario where the third sales hire slips a quarter, giving the board two numbers instead of one stale one. mint-board assembles the package itself, P&L, cash position, key metrics, and variance against the January plan. Lens's dashboard, refreshed nightly rather than rebuilt manually, feeds the ARR and churn figures directly into that package. And Budget's budget-audit output slots into the spend section as an actual breakdown instead of a lump total.

Mint, Board Package: April 2026
Recon: Burn $412k/mo, cash $2.8M, 3 sales hires approved (was 2 in
March model). Runway model rebuilt against current headcount plan.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. P&L summary
   ARR $9.2M (+$310k MoM). Gross margin 78%. Net burn $412k.

2. Runway
   Scenario A (plan as approved, 3 hires): 16.4 months
   Scenario B (3rd hire slips one quarter): 19.1 months
   Recommendation: proceed with plan as approved, runway still
   clears 15-month board threshold under both scenarios.

3. Key metrics (fed live from Lens dashboard, refreshed nightly)
   Net revenue retention: 118%
   Logo churn: 1.4% monthly
   CAC payback: 14 months

4. Cloud and AI spend (from Budget audit, this cycle: $38.4k)
   AWS infra:        $19.1k  (flat MoM)
   Anthropic API:    $11.6k  (+$3.2k MoM, support-bot volume)
   OpenAI API:        $7.7k  (embeddings pipeline, migration to
                      smaller model in progress, ~40% reduction
                      expected next cycle)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Prep time this cycle: ~90 minutes review, down from ~14 hours.

The board still gets the same deck format it expects. What changed is where Priya's fourteen hours went. Instead of reconciling four spreadsheets by hand, she spent ninety minutes reviewing output that was already grounded in a recon pass, already reflecting the hiring plan that changed two days prior, and already breaking the AI spend line into something specific enough to answer the board's question before it's asked. The runway scenario comparison alone answered a question that used to take a follow-up email and a week's delay.

Mint vs the alternatives for FP&A work

None of this is a knock on ChatGPT or Cursor as tools, they are excellent at what they were built for. The comparison below is specific to the recurring, cadence-driven work FP&A actually owns: the monthly board package, runway modeling under multiple scenarios, and spend attribution that has to survive a board's follow-up questions.

Tonone's Mint, Lens, and Budget turn a three-day, four-spreadsheet board prep cycle into a ninety-minute review by starting from a financial recon instead of a blank prompt.

Install and try

Tonone is free and MIT-licensed. Install it once and Mint, Lens, Budget, and the rest of the roster are available in your Claude Code session. You pay only for the Claude Code token usage during the work itself, the same spend Budget will help you track.