Xero's AI vs AI-Native Ledgers vs Your Own AI Agent

By Jude Lee · · Comparison

Two accountants reviewing general ledger reconciliation figures on a monitor in a modern firm office

Why the category suddenly looks crowded

Two forces are pushing in opposite directions at once. Incumbent platforms are layering AI onto the ledger you already use — Xero, for example, used its Xerocon US stage to announce new AI capabilities and Microsoft 365 integrations. Before you plan around any of it, read Xero’s own newsroom and product documentation rather than conference coverage: announcements and general availability rarely arrive on the same date, and regional availability differs.

Pulling the other way, venture-backed AI-native accounting platforms — Rillet is the one most often named — have been reported at valuations in the billion-dollar range. If you want to cite a specific figure internally, pull it from the outlet’s original article yourself; a number repeated third-hand is not a number you should put in a board deck. What matters for planning isn’t the valuation anyway. It’s the thesis behind it: that the general ledger itself gets rebuilt around automation rather than decorated with it.

You will also see survey headlines about the gap between firm leaders who say the future is automated and those with an actual transformation plan. Check the methodology before quoting any of them. Our position doesn’t depend on the statistic, and it is offered as opinion rather than measurement: enthusiasm is cheap, sequencing is hard. This piece is about sequencing.

The three layers, and what each one is actually good at

Embedded AI in the software you already run
Lives inside QuickBooks Online, Xero, your practice-management or document system. Sees that vendor’s data natively, ships with the subscription or a modest add-on, and requires no engineering. Strong at in-app jobs: suggesting categorisations, matching bank feeds, drafting an invoice reminder, summarising a report. Weak wherever the answer requires data the vendor doesn’t hold — your engagement letters, your workpaper standards, your billing realisation by client.
Your own AI agent connected via MCP
You connect an AI assistant (Claude and comparable tools) to your systems through MCP — an open protocol for giving an assistant governed access to data and tools — and define the procedure yourself. Strong at cross-system, firm-specific work: pull the trial balance, compare to prior period, check the PBC tracker, draft the client email. Weak on cost and discipline: someone owns the permissions, the logging, and the prompt-to-procedure translation.

The third layer — the AI-native ledger — doesn’t fit that comparison because it isn’t a feature you switch on. It’s a general-ledger migration. The pitch is that if the system is designed around automated close from the start, categorisation and reconciliation stop being tasks. That may well be true for the right client profile. It’s also a conversion project with data history, integrations, and staff retraining attached, and it changes what your team knows how to support. Treat it as a platform decision, not an AI decision — and price in the structural risk that attaches to any young category: early-stage vendors get acquired, repriced, or repositioned, and your migration plan should have an answer for what you do if yours is.

How firms are actually putting AI to work

Strip out the marketing and the real uses cluster into four buckets:

Where these deployments tend to fail, in our view: unattended posting to the ledger, judgment calls on materiality, and anything requiring a defensible audit trail the tool doesn’t itself produce. That’s why we argue for review gates rather than autopilot.

Concrete examples, from plain rules to agents

Not everything labelled automation needs a model. A useful ladder:

  1. Deterministic rules — bank rules in Xero or QuickBooks Online, recurring journals, approval routing. Cheap, auditable, boring. If a rule solves it, use the rule and stop there.
  2. Embedded AI suggestions — the vendor proposes, a human accepts.
  3. A skill — a packaged instruction set that makes an assistant perform one job the same way every time (e.g. “prepare the fixed-asset rollforward workpaper to firm standard”).
  4. An agent — multi-step, takes actions across systems, stops at defined checkpoints. This is where an agentic month-end close lives.

There is no single best tool — pick by where the data lives

Any list that hands you one winner is ranking for a search query, not answering your question. The honest decision rule: the best tool is the one that already holds the data for the job you’re automating.

  1. Name the job, not the category

    Write the actual task: “reconcile 40 client bank accounts monthly” beats “automate bookkeeping.” Categories don’t have owners; jobs do.

  2. Check whether one system holds all the inputs

    If yes, the embedded AI in that system is your first stop — it’s the cheapest test and the shortest security review.

  3. Count the systems if the answer is no

    Cross-system work (GL + practice management + email + document store) is exactly the gap MCP fills. Start with a read-only connection; our guide to connecting an assistant to QuickBooks or Xero via MCP covers least-privilege setup.

  4. Only then ask whether to build

    A custom MCP server earns its cost when the data lives in a system with no AI story and the workflow is high-frequency. Otherwise you’re maintaining plumbing for its own sake.

  5. Set the review gate before go-live

    Decide what the agent may do unattended, what needs sign-off, and where the log lives. Do this first — retrofitting governance after a bad month-end is expensive in trust.

Related reading if you’re weighing the build question directly: rules, AI agents, or neither.

Model the payback yourself — don’t borrow anyone’s number

Skip vendor ROI headlines. Use your own inputs:

Annual recovered capacity = (minutes saved per instance ÷ 60) × instances per year × blended hourly cost. Then subtract licence cost, build/config hours, and — the line most models omit — ongoing review time, because a human still checks the output.

Your own numbers
Minutes per instance × instances per year × blended rate
Worked example — substitute firm data
Subtract review time
Human sign-off doesn't disappear; it changes shape
Worked example
Then ask: billable or not?
Recovered hours only convert to revenue if they're redeployed
Worked example

That last line matters more than the first two. Recovered hours are not revenue until they’re sold or reallocated. Firms asking whether practices stay profitable under automation usually find the answer turns on pricing and capacity planning, not on the tooling.

Automation converts time into capacity. Only pricing converts capacity into profit.

What this does to the work — and the jobs

The honest read: AI is compressing preparation, not judgment. Categorisation, tie-outs, first-draft narrative, and document chasing are compressible. Attest opinions, tax positions, materiality, and client conversations are not — and the licensure, independence, and quality-management expectations set by the AICPA and state boards of accountancy attach to a person, not a model. “Will AI replace accountants” is the wrong frame; “which parts of the day survive contact with a competent agent” is the right one.

Stated plainly as our opinion rather than a measured finding: rule-based automation still wins on predictable, structured tasks, and AI earns its keep on messy, unstructured, cross-system work. A firm that runs everything through a model because a model is available will pay more for less auditability. New roles follow that split — an automation lead who can read a trial balance and write a clear procedure spec is, we think, one of the more defensible profiles in the profession right now.

Not sure where to start?

Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.

Get a free automation audit

Related guides