Accounts Payable on Autopilot: The AI-Native AP Stack for Outsourced Firms

Jun 18

Part 2 of the Lumiere Strategies series on building an AI-enabled finance function.

For an outsourced accounting firm, AP is a multiplication problem. A single bookkeeper running ten client files isn't processing one hundred invoices a month, they're processing a thousand. Each file has different vendors, different approval chains, different coding rules, and a client who keeps changing their mind about what counts as "office supplies." The work doesn't get harder per invoice. It just gets harder per firm. That multiplication factor is also why AP is the highest-ROI layer of the stack to attack first. Every minute of touch time you remove from a single invoice gets multiplied by every client file you serve.

The good news in 2026 is that the AP tooling category is no longer about routing rules and approval matrices. It's about agents that make first-pass decisions, with audit trails, and a human reviewing exceptions. The bad news is that the marketing language has gotten ahead of the reality, and choosing the right tool for an outsourced firm is meaningfully different from choosing one for a direct-to-business buyer. This post is a practical map: what changed in AP this year, which tools we'd lead with for an outsourced firm and why, the agent-by-agent workflow that's actually doing the work, and a rollout plan that doesn't lose you control of client approvals on day one.

What actually changed

The inflection point is recent enough to date. In October 2025, Ramp launched Agents for AP, a set of four specialized AI agents handling invoice coding, fraud detection, approval summaries with vendor context, and card payment execution. In February 2026, they extended the same model to bookkeeping with the Accounting Agent, claiming 3.5 times more auto-coding than legacy tools with 90%-plus accuracy on items it marks ready to sync. BILL, the long-standing accountant-channel leader, has been layering agentic features onto its existing platform throughout 2025 and into 2026. Stampli's Billy AI copilot has matured. Vic.ai continues to push deeper into the mid-market.

The marketed numbers are vendor-reported and should be treated with appropriate skepticism, but the directional shift is real. Rule-based AP automation, the kind that has been the backbone of practice management for fifteen years, is being displaced by tools that look at an invoice, decide how to code it, decide whether it looks suspicious, decide who should approve it, and execute the payment, all in a single sequence with a human reviewing exceptions rather than every transaction. The job is shifting from clicking "Approve" on screens of pre-coded invoices to reviewing a much smaller queue of agent-flagged exceptions.

The honest market map for outsourced firms

If you're running an outsourced practice with twenty or more client files, four vendors deserve serious evaluation, with a clear position on which to lead with.

BILL. The default for outsourced firms managing many SMB clients on QuickBooks Online or Xero. BILL's entire commercial model was built around the accountant channel: firms manage AP and AR pricing and access through an Accountant Console with an entity model, adding or removing client entities as your book of business changes rather than paying per firm user. That's the pricing structure that actually makes sense for an outsourced firm, and it's a meaningful advantage BILL still holds. The vendor payment network is the largest in the market. The AI features have been catching up but are still less audacious than Ramp's agents. For a firm whose clients are mostly SMBs on QBO, BILL is the default choice and the burden of proof is on whatever wants to displace it.
Ramp. The default for clients that already use Ramp cards, or for tech-forward clients that want a unified spend, expense, and AP platform. The agent suite is genuinely the most advanced in the market right now, and the AP module is free if the client is paying for cards. The catch for outsourced firms is that Ramp was built direct-to-business first and the accountant experience is newer than BILL's. The integration with QBO, Xero, NetSuite, and Sage Intacct is solid and improving fast. For a tech-forward client where you're picking the stack greenfield, Ramp is increasingly the right answer. For a portfolio of established QBO clients, BILL still wins on switching costs.
Stampli. Best when the invoice itself is the workspace and approval routing is genuinely collaborative across multiple stakeholders. The Billy AI copilot is good for surfacing context inside the invoice conversation. Stampli is overbuilt for a single-bookkeeper client and underbuilt for a firm running across thirty client files. There's a real niche, but it's not the outsourced default.
Tipalti and Vic.ai. Both step up when complexity steps up. Tipalti for clients with global vendors, multi-currency, 1099 and 1042 tax requirements, or a meaningful supplier network. Vic.ai for mid-market clients where AP-specific AI depth matters more than a unified platform. Both are overkill for a typical SMB client and warranted for a specific subset. Don't lead with either; reach for them when the client's complexity forces the move.

The position we'd take with a new firm or an unfamiliar book of clients: lead with BILL for most clients, lead with Ramp for tech-forward and card-heavy clients, and avoid the trap of standardizing on a single tool across a portfolio where the clients genuinely don't fit the same model.

The four agents that do the work

Strip away the marketing and the modern AP workflow is doing five jobs, each with a corresponding agent. Naming them explicitly is useful because it tells you what to evaluate during a demo and what to monitor once you've deployed.

The extraction agent reads the invoice. It pulls header data and line items from any format, PDF, scan, or electronic feed, without you setting up templates per vendor. The bar to beat is roughly 99% OCR accuracy on a clean invoice and visibly graceful degradation on a messy one. Test it on your messiest invoices before you sign anything.
The coding agent assigns GL accounts, classes, locations, and any other dimensions. It learns from historical patterns in that specific client file, which is why it gets better over weeks, not days. The metric that matters is the auto-code rate at high confidence and the override rate when humans intervene. A target worth aiming at is 80% auto-coding at high confidence by month three on a given client file.
The fraud agent checks for duplicates, vendor anomalies, amount outliers, and policy violations before payment, not after. This is where the audit-trail discipline lives. Every action the agent takes should be inspectable.
The approval agent routes the invoice. For an outsourced firm, this is the most nuanced piece, because the approval chain typically lives on the client side, not the firm side. The firm preps and codes, the client approves. The agent should respect each client's approval policy, draft approval summaries with vendor context, and chase the client when something is sitting too long.
The payment agent executes. ACH, card, wire, or check. It schedules payments to optimize for early-pay discounts and cash flow, then writes back to the GL.

The whole sequence runs in a few minutes per invoice when nothing is unusual, and pauses for human review when something is. That's the shape of modern AP.

A rollout that doesn't lose you control

The single most common failure mode for firms installing AI AP tooling is enabling auto-approve and auto-pay in week one and losing visibility before the team has built trust in the agent. The right rollout is sequential, with each phase earning the right to enable the next.

Weeks 1 and 2: historical extraction. Run the extraction and coding agents against the last six months of invoices for a pilot client file. Don't enable any actions. Just see how the agents code the work you've already coded yourself. The override rate tells you how much trust to extend and how much training the agent still needs.
Weeks 3 and 4: shadow mode. Turn the agents on for live invoices, with the agents suggesting and humans confirming every action. Track touch time per invoice and exception rate. The goal of this phase is to get touch time below three minutes for exceptions and below thirty seconds for high-confidence items.
Weeks 5 and 6: auto-code at high confidence. Begin auto-coding invoices the agent marks high confidence, with humans reviewing only exceptions. Watch the auto-code rate, the exception rate, and the late-payment rate. If late payments tick up, the approval agent is not routing correctly.
Week 7 onward: expand to fraud and approval. Turn on fraud detection and approval routing. Keep payment execution under human review for one more cycle, then enable it once you trust the rest of the chain.

Across the rollout, the metrics worth watching weekly are touch time per invoice, auto-code rate, exception rate, days payable outstanding stability, and the number of invoices touched by more than one person. The last metric is the truest measure of how much friction you've actually removed.

For the firm owner, and for the operator

If you're a firm owner deciding what to buy, the math is simple. AP is the highest-volume work in the practice and therefore the highest-ROI place to install AI first. Expect a 60% to 80% reduction in touch time per invoice on a given client file within ninety days of a well-scoped rollout. That capacity, multiplied across your book, is what funds the move up the stack toward advisory. The 2023 CAS Benchmark research from CPA.com and BILL on spend management adoption pointed in this direction even before the agentic shift. The 2025 and 2026 wave of agentic tooling has accelerated it.

If you're the controller or operator inside the firm running the rollout, the work is judgment more than configuration. Pick one client file, ideally a mid-volume QBO file where you know the historical coding well, and run the four-phase rollout above. Bring one detailed write-up back to the partner at the end of the eight weeks. That write-up becomes the playbook for the next client, then the next, then the firm-wide deployment. Don't try to roll out to thirty client files at once. The agents need training data per file, and your team needs muscle memory.

The throughline matches the rest of the series. AP isn't a tool you buy. It's a workflow you redesign, agent by agent, with the audit trail and the human-in-the-loop discipline preserved, until the work that used to consume your team's mornings runs quietly in the background.

Resources

The Lumiere AI Stack Map for Accounting Firms - Current outlook of solutions for Client Transactions, Close & Review, Practice Operations, and Advisory
Two-week Time Study Template for Fina nce Teams - Template to track resource efforts by service layer to determine areas of opportunity or concentration.