Your Accounting Software Just Became an AI Agent — Here's What to Turn On First

Layered gears turning at different speeds, representing tiers of AI agent trust and autonomy in a finance system

Your Accounting Software Just Became an AI Agent — Here's What to Turn On First

Xero, Expensify, Digits, Ramp and Intuit have all shipped AI agents this year. A practical order of adoption for finance teams that don't want to hand over the keys on day one.

If you run a finance function on Xero, QuickBooks, Expensify or Ramp, your accounting software has spent the first half of 2026 quietly turning itself into something closer to a coworker than a ledger. Xero's agentic platform JAX picked up three new capabilities at Xerocon London in early July. Expensify shipped a Model Context Protocol (MCP) layer in June. Digits and Ramp did the same earlier in the year. Intuit partnered directly with Anthropic in February to build custom AI agents into QuickBooks, TurboTax and Credit Karma. It's worth being clear on what these actually are: MCP is a connectivity standard — it lets an AI assistant query a system's data through a consistent interface, rather than a person exporting a CSV. It's not, on its own, an agent that acts. What each vendor has built on top of that connectivity is where the real range sits, from Digits' read-only query layer through to Xero JAX's action-taking features — which is exactly the distinction the tiers below are built around. None of this requires you to buy anything new — for most NFP, NDIS and SME finance teams, it's already sitting inside software you're paying for. The question isn't whether to use it. It's what to switch on first.

What's Actually Landed So Far This Year

5
Major finance platforms — Xero, Expensify, Digits, Ramp and Intuit — that shipped AI agents or MCP connectors in the first half of 2026.
Read-only
Digits deliberately built its MCP server so an agent can analyse the books but never alter them — a useful benchmark for how cautiously to start.

Intuit and Anthropic announced their partnership in February, giving Claude users MCP-based access to QuickBooks, TurboTax, Credit Karma and Mailchimp data. Digits shipped a read-only MCP server on 21 April, letting an AI agent analyse the books without being able to touch them. Ramp rolled out MCP and CLI connectors alongside a set of procurement agents that same month. Expensify launched its own MCP layer on 8 June, letting AI assistants query expense data through plain-language requests instead of manual exports. Then at Xerocon London on 8–9 July, Xero added three concrete features to its agentic platform JAX: Payment Follow-ups, which automatically chases overdue invoices; Bill Protection, which inspects incoming bills and flags unusual amounts or unfamiliar suppliers before they're paid; and Cash Flow Actions, which builds preventative plans to protect payroll and suggests delaying non-critical bills when cash is tight — all reviewable and adjustable on the same ledger, without leaving Xero.

A Practical Order to Turn Things On

Not all of these agents carry the same risk, and they shouldn't be adopted in the order your vendor happens to release them. A more sensible sequence works in three tiers.

Tier 1 — start here: read-only, ask-a-question agents. Tools like Digits' MCP server, or Expensify's natural-language expense queries, let you or your team ask questions of the books without any risk of the answer changing the books. This is the lowest-risk entry point and the best place to build internal confidence in what these tools actually get right and wrong.

Tier 2 — flagging and alerting, human decides. Xero's Bill Protection sits here: the agent inspects and flags, a person still approves. This is where most of the near-term value is for a small finance team — catching the unusual invoice or duplicate supplier before it's paid, without removing the final decision from a human.

Tier 3 — proceed carefully: agents that act. Payment Follow-ups (chasing customers automatically) and Cash Flow Actions (executing a delay-payment plan) sit at the top of the risk ladder because they take action on your behalf, with real counterparties, using your organisation's name. These are worth adopting last, and only once your team has a clear sense — from Tiers 1 and 2 — of how reliable the underlying data and flagging actually are in your own books.

Before connecting any of these tools to live financial data, confirm whether the vendor trains its models on your inputs, and whether that's something you can opt out of. This matters more for organisations handling payroll data, NDIS participant records, or other sensitive client information — the safer starting posture is a tool where your data isn't retained for training at all, and where you can see exactly what the agent has access to before you turn it on.

The One Question to Ask Every Vendor

Beyond the training-data question, ask specifically what the agent can act on without a human in the loop, and whether that scope is configurable. Digits made a deliberate design choice to stay read-only — that's a statement about where they think trust currently sits in the market, not a technical limitation. Xero, by contrast, has built agents that can genuinely take action, which makes the approval settings and audit trail around those actions the thing to interrogate before rollout, not an afterthought once something's already gone wrong.

In practical terms, that means asking your vendor rep three things before enabling any action-taking feature: can every action the agent takes be traced back to a specific decision with a timestamp and reason; can the approval threshold be set to "always ask" rather than "ask only above $X"; and can the feature be switched off instantly if something looks wrong, without needing a support ticket. If a vendor can't answer all three clearly, that's a reasonable basis to stay at Tier 1 or 2 with that platform for now, regardless of what the marketing promises.

A Simple Rollout Checklist

  • Start with a read-only or query-only agent and run it alongside your existing process for at least a full reporting cycle before trusting its output on its own.
  • Confirm exactly which accounts, ledgers or data sources the agent can see — not just what the vendor's marketing describes.
  • Check whether any action-taking feature (payment chasing, bill payment delays) has a configurable approval step, and set it to require sign-off until you trust the flagging.
  • Document who in your team is accountable for reviewing what the agent flags — an agent that raises an alert nobody reads is no safer than having no agent at all.

None of this needs to slow adoption down. It's the opposite — starting at Tier 1 and moving up as trust is earned is usually a faster path to real value than switching everything on at once and rolling it back after the first mistake.

Why This Matters More for Smaller Finance Teams

A large corporate finance function can afford to run a pilot for six months with a dedicated project team watching it closely. Most NFP, NDIS and SME finance teams don't have that luxury — often it's one or two people covering reconciliation, reporting and payroll alongside everything else. That's exactly why the tiered approach matters more here, not less. A read-only query agent gives a stretched team real leverage — faster answers to "what did we spend on X last quarter" — without adding a new risk to monitor. An agent that can pay bills or chase customers on its own, adopted before anyone's had time to build trust in its judgment, is a much bigger problem for a two-person team to catch and unwind than it is for a finance department of fifty.

There's also a genuine budget angle for FY2027 planning. Several of these tools — Digits' MCP layer, Expensify's natural-language queries — are bundled into subscriptions you're likely already paying for, which makes the Tier 1 starting point close to free to trial. The features worth budgeting for separately are the Tier 3 action-taking agents, where vendor pricing increasingly reflects the value of automation rather than just software access. Treat that as a deliberate line item to evaluate on its own merits next budget cycle, rather than something that arrives bundled in a renewal and gets switched on by default.

Not sure which of these tools are safe to turn on for your organisation?

PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including helping teams work out what to adopt first and what to hold off on.

Talk to PFL →
Timothy, CPA is Managing Director of Professional Financelink (PFL), providing senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. 20+ years in finance leadership across NFP, NDIS and SME.

Comments

Popular posts from this blog

Google Gemma 4 Just Launched — And It Might Solve Finance's Biggest AI Privacy Problem

Claude vs Gemini for Australian Finance: An Honest Comparison After 12 Months of Using Both

Why NFP Boards Are Finally Talking About AI — And What the Finance Team Should Do Before They Ask