Xero has more AI built into it than most business owners realise. Between JAX answering financial questions, Syft-powered analytics generating profitability insights, and smart bank reconciliation learning your coding patterns, the platform has shifted significantly in the last twelve months. But some of the most interesting AI work in the Xero ecosystem isn't coming from Xero alone - it's showing up across the tools and platforms that connect to it

This isn't a roundup of every app in the Xero App Store that mentions AI in its marketing copy. There are dozens of those, and most use "AI" to describe basic automation they've been doing for years. This is a practitioner's teardown of the tools that are actually changing how we work with Xero across our client base - how they work technically, what they can and can't do today, and what's on their roadmaps

Xero's built-in AI layer

Xero's AI strategy sits on two pillars. JAX (Just Ask Xero) for automation and conversational queries, and the embedded analytics layer powered by Syft for reporting and insights. They serve different purposes and they're at different stages of maturity

JAX - smart bank reconciliation and natural language queries

JAX's strongest production feature is automated bank reconciliation. It uses four layers of intelligence to suggest how each bank feed transaction should be coded. It draws on your historical coding patterns, supplier name matching, transaction amount patterns, and contextual signals from the broader Xero data set. Xero claims over 97% accuracy on suggested matches. From what we see across client files, that holds for businesses with consistent transaction patterns - a trades business buying from the same suppliers weekly gets accurate suggestions fast. A business with irregular, varied spending takes longer for the model to learn

One detail that matters. JAX builds a per-organisation model. It trains on your specific coding history, not a generic model shared across all Xero users. That's why accuracy improves the longer you've been reconciling consistently in a file. It also means a new Xero file or one with messy historical coding will get poor suggestions until the model has enough clean data to work from

On the conversational side, you can ask plain-English questions about your data. "What were my top expenses last quarter?" returns a chart. "Show me my profit trend for the last 12 months" gives you a visual summary. JAX can also pull in external information via Xero's OpenAI partnership - tax rates, regulatory details from sources like the Australian Taxation Office (ATO) website. What it can't do yet - comparative analysis with adjustments ("How does this quarter compare to last year if I strip out that one-off equipment purchase?") or proactive anomaly detection ("What changed that made margins drop?")

Where's it heading? Xero's Chief Product Officer Diya Jolly has framed 2026 as the year JAX moves from "just ask" to "just done", with a stated target of automating 90% of routine bookkeeping tasks. Automated reconciliation of high-confidence transactions launched in beta in November 2025 and is now live globally - opt-in per bank account, with a new Reconciled page showing full audit trail of what JAX matched and why. The target is auto-reconciling over 80% of bank statement lines in real time. From what we see across client files, it handles the straightforward stuff well - regular supplier payments, recurring subscriptions, simple one-to-one matches. Where it struggles is anything with complexity - split transactions, partial payments, transactions that need tracking categories applied, or spend that doesn't follow a consistent pattern. The 80% target is realistic for a clean file with predictable spending. For a business with messy, varied transactions, expect less. Next on the roadmap - AI-assisted bill payments that handle data capture through to payment and reconciliation in a single workflow. JAX usage has increased 61% in the last three months, which Xero is treating as validation that practitioners trust the outputs enough to let it work

Xero Analytics (embedded Syft)

Xero acquired Syft Analytics in late 2024 for around $70 million. The embedded product went live globally in January 2026 and is available to all business plan subscribers

What it does today. Customisable dashboards covering revenue, expenses, and KPIs. Visualisations for profitability, cashflow, and balance sheet health. A cashflow manager that projects up to 180 days with scenario planning. You can model "what if revenue drops 20%" or "what if we hire two more staff" and see the projected cash impact over six months

Within the analytics, AI generates plain-language summaries of profitability trends. Select a profitability graph and it explains what drove the movement - which expense categories increased, where revenue shifted. It's powered by prompts against your data, not a static rules engine, so the explanations adapt to what actually happened in your file rather than applying generic commentary

Syft continues as a standalone product alongside the embedded version. The standalone goes deeper. Consolidated reporting across multiple entities, industry benchmarking, detailed forecasting, custom report packs, and SOC 2 Type II accreditation. Xero has trailed further updates for the second half of 2026 including anonymous benchmarking against similar businesses and a native budgeting tool

Dext - AI-powered pre-accounting

Dext processes receipts, invoices, and financial documents using OCR combined with machine learning classification. The architecture works in three layers

First, optical character recognition. Dext reads documents in any format - photographed receipts, scanned PDFs, email attachments, even handwritten notes - and extracts structured data - supplier name, date, amount, tax treatment, line items, due dates. The claimed extraction accuracy is 99.9%, and across our client base that holds for standard printed invoices. Handwritten or damaged documents still need human review

Second, machine learning classification. Connect a client to Dext and it syncs the Xero chart of accounts, then starts suggesting category codes for each document. Those suggestions train on the specific client's coding history. After a few months of processing, Dext learns that invoices from Bunnings go to "Materials", Officeworks goes to "Office Supplies", fuel receipts go to "Motor Vehicle Expenses". It builds supplier rules automatically. Each correctly accepted suggestion improves the next one

Third, workflow automation. Approval chains, automatic publishing rules (anything under a threshold goes straight to Xero as a draft bill without human touch), multi-currency handling, and duplicate detection that flags when the same document has been submitted twice

The practical impact across our client base is clear. Businesses that adopt Dext properly - photographing receipts as they happen rather than batching them monthly - have cleaner books and faster BAS turnarounds. Our team spends less time on data entry and more time reviewing what's been coded. That's the shift. The AI doesn't replace the bookkeeper. It changes what the bookkeeper spends their time on

We covered Dext alongside other add-ons in our Xero add-ons guide, but the AI layer is worth examining separately because the learning capability is what separates it from basic receipt scanning

The 2026 roadmap is substantial. Dext launched Payments in February 2026 (UK first, other markets following) - closing the loop from invoice capture through approval to settlement inside a single platform. That's significant because manual re-keying into bank portals is where most payment errors and fraudulent diversions occur. Coming next is payroll payments (direct settlement of net pay from uploaded payroll files), spend cards for capturing expenses at point of purchase, and invoice payment links. The more interesting roadmap item is what Dext calls an "AI Agent" - a feature that analyses the actions taken inside client files to identify which tasks should be automated next. It's moving from document processor to pre-accounting intelligence layer

Fathom - reporting intelligence and anomaly detection

Fathom connects to Xero via direct API integration and turns raw accounting data into management reports, KPI dashboards, three-way cashflow forecasts, and consolidated financial statements across multiple entities

For practitioners, the features that matter are anomaly detection and exception-based reporting. Set thresholds on any metric - gross margin, debtor days, operating expense ratio - and Fathom flags when results fall outside those bounds. For firms managing dozens of clients, this changes the advisory workflow. Instead of reviewing every client's numbers each month, you focus on the ones where something has actually moved

Forecasting runs three-way models across P&L, Balance Sheet, and Cash Flow projections up to five years. You build scenarios using operational drivers (headcount, pricing, volume) rather than just trend extrapolation. That produces more realistic forecasts for businesses going through structural change - adding a team, changing pricing, expanding into a new service line

Fathom won Best Management Reporting App in CFO Techstack's State of the Stack 2026 report. 99,000 companies use it. It handles intercompany eliminations, multi-currency translation, and group benchmarking - features that the embedded Xero Analytics layer doesn't touch. Recent feature releases include Forecast Snapshots, which let you compare actuals against budgets and prior forecasts side by side

Fathom doesn't publish a public product roadmap. They ran a "Roadmap Reveal" webinar in late 2025 previewing 2026 plans, but specific features haven't been announced publicly beyond the Forecast Snapshots release. What we do know is the competitive pressure from embedded Xero Analytics is real, and Fathom's response will likely push harder into the advisory and multi-entity territory that Xero's native layer can't reach

Automation platforms connected to AI models

This is the part nobody else in the bookkeeping space is writing about, because most firms aren't doing it yet

Tools like Make and N8N are automation platforms that connect to APIs - including Xero's. They pull data from Xero, pass it to a large language model like Claude or GPT, get an analysis back, and either present the results or post structured data back into Xero. The automation handles the data pipeline. The AI handles the interpretation

In practice it looks like this

We use Make connected to Xero's API and an AI model to do file reviews at scale. When we onboard a new client or run a periodic review, the automation pulls structured data from their Xero file - transaction history, audit trail entries, notes, user activity patterns - through controlled API calls with strict guardrails on what data crosses into the AI layer. The AI analyses that data for patterns. When people are working in the file, how they're working in it, what the workflow looks like across the client base, where the risks sit

For a single file, you could do this manually. Open Xero, run the history report, scroll through the audit log, spot the patterns yourself. Takes about an hour. But when you're managing over a hundred client files and you want to assess consistency of process, identify where reconciliation is falling behind, or flag where transaction patterns have changed in ways that need attention - doing that manually is a full-time job. Doing it through an automation layer connected to an AI that can read patterns at scale takes minutes

Guardrails are the critical design decision. We're not sending entire client ledgers to an AI model. The automation pulls specific, structured data points - metadata about process and workflow without exposing sensitive financial details. The difference between reading the timestamps on someone's calendar and reading their diary. The timestamps tell you about patterns without the content

Technically, a Make scenario triggers on schedule or manual run, calls Xero API endpoints (audit history, bank transaction metadata, user activity), transforms the response into a structured prompt, sends it to the AI model via API, receives the analysis, and routes the output wherever it needs to go - a dashboard, a Slack notification, a client review document. N8N works the same way with a self-hosted option for firms that want the entire data pipeline on their own infrastructure

We're still building and refining these workflows internally. But this is where AI in the Xero ecosystem is genuinely heading - not smarter data entry, but the ability to analyse how a business's financial operations actually work at a level that wasn't possible twelve months ago

SavvyWise - AI in tax and compliance research

AI isn't only changing how data gets into Xero. It's changing how the professionals working with that data do their research

SavvyWise is a Perth-built AI platform for Australian tax professionals. It indexes Australian tax legislation, ATO rulings, case law, state taxes and duties, Fair Work legislation, and auditing standards. Ask a research question in plain English and it returns relevant legislation, rulings, and analysis within seconds - work that traditionally takes 30 to 60 minutes per query sifting through the ATO website, tax databases, and technical resources

What separates it from a general-purpose AI like ChatGPT is the training data. SavvyWise is built on verified Australian legislative sources, not the open web. General AI models will confidently cite tax rulings that don't exist or misinterpret legislation because they're pattern-matching against internet text rather than verified legal sources. SavvyWise constrains its answers to actual source material with references you can check

It completed an oversubscribed seed round at a $10 million valuation. The founders come from Munro's (a top Perth accounting firm) and software engineering, which shows in the product - it's built for how accountants actually work, not as a generic chatbot with a tax skin. The roadmap extends beyond tax research into HR compliance, Fair Work interpretation, and broader advisory research

SavvyWise represents a different category from every other tool in this article. Dext, Fathom, and Xero Analytics operate on the transaction and reporting layer. SavvyWise operates on the interpretive layer that sits above the numbers - the layer where a professional turns data into advice. Together, they sketch what the full AI-assisted accounting workflow looks like within a few years. AI handling extraction, coding, reconciliation, and reporting at the bottom, with AI-assisted research and interpretation at the top. The human sits in the middle, making the decisions that matter

Where each tool sits today

Production-grade

  • JAX smart bank reconciliation - Per-organisation ML model, 97%+ accuracy on consistent files, improves with clean historical data. Production since late 2025
  • Dext OCR and ML classification - 99.9% extraction accuracy on standard documents, learned supplier rules, bi-directional Xero sync. 700K+ businesses, 320M+ documents annually
  • Fathom reporting and forecasting - Three-way cashflow models, KPI anomaly detection, multi-entity consolidation. 99,000 companies, 13+ years in market
  • Xero Analytics (Syft embedded) - Dashboards, 180-day cashflow projections, scenario planning. Live globally since January 2026

Functional but maturing

  • JAX conversational queries - Answers basic data questions. Can't yet do contextual analysis or cross-period comparisons with adjustments. External data via OpenAI partnership rolling out through 2026
  • Syft AI profitability insights - Plain-language explanations of financial movements. Currently limited to profitability graphs. Expanding to more report types through 2026
  • Dext Payments - Invoice-to-settlement workflow launched Feb 2026 (UK). Payroll payments, spend cards, and invoice payment links on the 2026 roadmap
  • SavvyWise tax research - Australian legislation and rulings indexed and queryable. HR compliance and broader advisory on the roadmap

Early-stage, requires technical capability

  • Make/N8N + AI model pipelines - Powerful for firms with the technical skills to build them. Not a product you install. Requires API knowledge, prompt engineering, and careful data governance. The payoff scales with the number of client files under management

Still mostly marketing

  • "AI-powered" labels on basic automation - Auto-matching invoices by amount and date has existed since 2018. Relabelling it doesn't change what it does
  • Generic chatbot wrappers - Several Xero add-ons offer "ask about your finances" features that are thin interfaces over general language models with no accounting-specific training or business context

Xero's stated goal is 90% automation of routine bookkeeping. Dext is moving from document processor to pre-accounting intelligence layer with its AI Agent feature. Fathom and Syft are converging on AI-powered advisory insights. And the automation-platform-plus-AI-model approach is creating capabilities that didn't exist twelve months ago

None of it works without clean data underneath, though. AI tools pointed at a Xero file with a messy chart of accounts, uncoded bank transactions, or the wrong GST reporting basis will produce perfectly confident wrong answers. Get the foundations right. Then build up

Current as at March 2026