Finance software splits in two: tools agents can drive, and meters for what those agents cost.

The week in finance

The dominant move in finance this week is agents crossing from feature to substrate. Sequence shipped both halves of an agent-native billing stack in one arc: an MCP endpoint that lets external agents query invoices, schedules, and revenue in natural language, and an AR Agent that reads an inbound billing request, validates the sender, and stages the change for human approval in its Watchtower review gate before anything commits. Two sparks in one window, and the design choice is the story — the agent prepares, a human confirms. Intuit Intelligence is threading the same review-not-enter pattern through the QuickBooks bank feed, where AI now predicts payees on money-in transactions and surfaces color-coded confidence for an accountant to confirm.

The flip side showed up just as clearly: someone has to meter what these agents cost. CloudZero shipped visibility into Claude Code spend, extending cost attribution from cloud infrastructure to AI coding agents themselves — the clearest signal of where finops is heading as engineering teams turn agents loose. When the same week produces both the tools agents drive and the meter for their bill, that is a sector pricing in a new line item in real time.

Leaders

Sequence is the week's standout, pairing its MCP surface with the AR Agent to build a system an agent can both query and act within, without ever letting it commit unreviewed. The conventional integration work underneath — Xero journal post timing, automatic Stripe payment retries, CPQ renewals appended as schedule phases — is the reliable substrate the agent features need to be trusted.

Credit Repair Cloud shipped the one release in its window that changes what the product is for. BuildCredit Rent reports on-time rent as credit history with up to 24 months of backfill, moving the platform onto the building side of the market rather than only the disputing side. Everything else in its feed removed client-churn friction; this adds positive history, a more durable relationship with the end consumer.

CloudZero carried no spark but the most on-theme improvement stream in the sector: Claude Code cost tracking, AI Explorer CSV export into existing reports, and API-key scopes reorganized by use case to tighten governance as its surface widens into AI spend. The direction is a single meter for inference-heavy workloads.

Intuit Intelligence kept threading AI through the bank feed — payee prediction on money-in transactions, a toggle to hide AI explanations, and bank rules that now carry priority ordering and a broken-rule view for accountants managing many client books. The arc is AI-assisted bookkeeping where the human reviews rather than enters.

Wildcards

Payhawk is the off-pattern player: no AI in sight, and instead bolting a travel agency and a payments bank onto an expense platform. Its Five Layers of Payment Fraud Prevention spark landed alongside bulk global FX transfers — insider-fraud controls and real accounts-payable volume arriving together is the tell that Payhawk is building the trust surface finance teams gate before routing money through a platform. Travel depth (allowances in policy, booking on behalf of others) and ERP export hardening round out a spend-lifecycle consolidation bet that owes nothing to the agent story everyone else is chasing.

Themes that compounded

• Human-in-the-loop gating is the shared design choice: Sequence routes agent actions through Watchtower, Intuit Intelligence has the AI propose and the accountant confirm.
• AI cost became a finance line item, with CloudZero extending attribution to Claude Code and per-agent spend.
• MCP surfaced as the integration primitive, Sequence exposing billing data to any MCP-compatible agent over OAuth 2.0.
• AI moved into the daily bookkeeping loop — Intuit Intelligence in the bank feed and Financial Cents taking AI file renaming from beta to live under a new AI Agents settings area.
• Reliability substrate work kept pace with the flashier bets: Sequence on Xero and Stripe mechanics, Payhawk on supplier-sync diagnostics and ERP export consistency.

Watch this week

Watch whether Sequence widens the AR Agent from triaging billing requests toward more billing actions, all still routed through Watchtower review — that is the test of whether human-gated agents can carry real customer-state changes. On the metering side, watch CloudZero push past Claude Code cost into per-agent and per-inference attribution; the AI Explorer CSV export suggests finance teams are already pulling that data into their own reports. And watch Credit Repair Cloud — BuildCredit Rent is sold as an add-on, which points to more alternative-data reporting arriving the same way rather than folded into the base plan.