Background
Xero just gave 4.5 million subscribers access to the Claude model (many of them hardly use it anyway, but thats another story).
Their share price fell 64%.
That’s not a coincidence. It’s a scoreboard.
In March, Xero announced a multi-year partnership with Anthropic.
By May, Claude was live inside Xero for every subscriber on the planet.
The press releases talked about “real-time financial intelligence” and “agentic workflows.” Xero’s Chief Product Officer called it a milestone. They followed it with a Microsoft Copilot deal and announced XeroForce, a no-code AI agent builder.
The market shrugged.
Then it sold.
Xero hit $193.77 in June last year.
It’s trading around $62 today.
Three consecutive sessions of multi-year lows last week. Down 45% calendar year. Down 64% from the peak. The more aggressively Xero leans into AI partnerships, the less the market seems to believe it changes anything.
I think the market is right.
The wrong problem, solved confidently
The entire pitch of Claude inside Xero is that you can ask natural language questions about your finances. “What’s my cash position?” “Which invoices are overdue?” “Can I afford to hire?”
These are fine questions. But they’re the wrong problem to solve, because the people who need those answers already know how to find them. They open Xero. They look at the dashboard. They run a report. It takes 30 seconds.
The hard part of financial management was never accessing data. It was turning data into decisions. And decisions require structure: a model that ties revenue drivers to cash flow, that reconciles the balance sheet to the P&L to the cash flow statement, that lets you toggle scenarios and stress-test assumptions before you commit capital.
Claude inside Xero doesn’t give you any of that. It gives you a paragraph.
Here’s what a paragraph can’t do:
1. Reconcile.
A proper 3-way model proves that the balance sheet balances, the cash flow ties to the bank, and the P&L feeds both. Every line is formula-driven. Every subtotal cross-checks. Claude gives you a chat response. Chat responses don’t reconcile. They summarise, and sometimes they summarise wrong, and you have no way to tell the difference unless you already knew the answer.
2. Survive an audit.
When a board member, a lender, or an auditor asks “where did this number come from?”, the correct answer is a cell reference. B47 on the Output sheet, pulling from the Revenue Calc engine via a SUMIF on the mapped COA, which traces back to the Trial Balance that was pulled live from Xero at 9am on Tuesday. That’s an audit trail. That’s what governance looks like. “Claude told me” is not governance. It’s faith.
3. Persist.
The output of a Claude + Xero conversation exists in one chat window, for one session. You can’t version it. You can’t compare it to last month. You can’t send it to your auditor or attach it to a board paper. Try running your month-end close on that. Try explaining to your accountant why this month’s numbers look different from last month’s when both came from the same chatbot but neither was saved.
4. Handle scenarios.
“Can I afford to hire?” is not a yes/no question. It’s a scenario. It depends on your revenue trajectory, your cash conversion cycle, your payroll tax obligations, your leave accrual assumptions, and whether your biggest customer pays on time. A model lets you toggle Base, Bull, and Bear cases, run a sensitivity table, stress-test the downside. Claude gives you one answer with no assumptions visible and no way to challenge them. That’s not analysis. That’s a coin flip dressed in confidence.
5. Validate itself.
A proper financial model has 20+ automated checks. PASS/FLAG/FAIL on every critical assumption, every reconciliation point, every cross-sheet dependency. If the balance sheet is out by a dollar, a red flag fires before anyone sees the output. Claude + Xero has no validation layer. It has vibes. If the AI hallucinates a number, who catches it? Not the small business owner who asked the question because they didn’t know the answer in the first place.
6. Prove formula transparency.
If Claude tells you your gross margin is 42%, you have no way to verify whether it excluded a one-off credit note, handled FX correctly, included that intercompany elimination, or accidentally double-counted a reversed journal. In a model, you click the cell and trace the formula. In a chat, you trust and pray. For a profession built on verification, that should be disqualifying.
7. Produce a deliverable.
Your bank doesn’t accept a screenshot of a chat. Your board doesn’t accept “Claude says we’re fine.” A lender reviewing your covenant compliance needs a structured schedule, formatted, reconciled, with a sign-off block. Claude + Xero produces conversational text. That’s not a deliverable. That’s a draft of a thought about a deliverable.
8. Support your advisor.
Here’s the angle nobody talks about. Xero has 4.5 million subscribers, and most of them have an accountant or bookkeeper. Those advisors are trying to scale from compliance into advisory. They need structured reporting they can repeat across 50 clients, not a chatbot that gives a slightly different answer each time they ask the same question. Advisory at scale requires templates, models, and repeatable processes. Claude + Xero gives you a one-off response. That’s the opposite of scalable.
9. Protect the small business owner.
This is the one that genuinely concerns me. Xero is pitching Claude to millions of small business owners who, by definition, are not financial experts. These are the people most likely to take a confident-sounding AI answer at face value. When Claude says “your cash position looks healthy” and the owner skips the invoice chase, or takes on a lease, or hires ahead of revenue, based on a response that wasn’t validated, wasn’t stress-tested, and can’t be audited, the consequences land on them. Not on Xero. Not on Anthropic. On the business owner who trusted the machine. AI fluency is not the same as AI accuracy, and the gap between the two is where businesses get hurt.
10. Replace the need for a model.
This is the core delusion. The pitch assumes that if you can ask a question in natural language, you don’t need a structured model. But a model isn’t just a way to get answers. It’s a way to organise thinking. It forces you to state your assumptions explicitly, to connect causes to effects, to see how a change in one variable cascades through the whole business. A conversation with Claude doesn’t do that. It gives you an endpoint with no visible logic chain. That might feel like insight, but it’s actually the opposite. It’s a conclusion without a workable path to challenge it.
Every one of these gaps is a governance failure waiting to happen.
Retrieval is not architecture
The deeper issue is that Xero + Claude conflates two fundamentally different things. Answering a question about your cash balance is retrieval. Building a 3-way model with a tax engine, working capital profiles, scenario toggles, and a DCF valuation block is architecture. AI is spectacularly good at the first and demonstrably unreliable at the second.
I’ve tested this extensively. The FMI benchmarked it in February. AI can produce a 5-year model from a single prompt, but the output isn’t deployable. Hardcoded values where formulas should be. Confident numbers that don’t reconcile. Structures that look right until you pull one thread and the whole thing unravels. I call it the fluency trap: AI that appears capable without being deployable.
The real unlock isn’t bolting a chatbot onto a ledger. It’s encoding human judgement into reusable AI skills, so the modeller’s value shifts upstream. You stop doing the structural work and start defining it. The AI handles the repetitive delivery layer: refreshing commentary, building the board pack, running the variance analysis, generating the HTML dashboard. But the architecture, the thing your board is relying on, that stays human-built and audit-grade.
What actually works?
That’s what I built EXL Cloud to do.
Live Xero data pulled into Excel in 8 clicks.
3-way driver-based financial models with full reconciliation. Scenario managers. Validation engines with PASS/FLAG/FAIL on every check. Power BI integration. HTML board packs. And 19 AI-powered skills built on Claude that don’t just answer questions but build actual deliverables.
The skills don’t replace the modeller.
They encode judgement so the modeller can scale.
An accountant with EXL Cloud and Claude skills isn’t being replaced by AI.
They’re being promoted.
Their value shifts from doing structural work to defining it, from building the model to governing the output, from producing one board pack a quarter to producing one a month without doubling their hours.
That’s what 6th Generation Modelling looks like. Human-built architecture. AI-orchestrated delivery.
Structured, auditable, repeatable.
Excel isn’t the problem. 750 million people use it because it’s the most flexible analytical tool ever built.
The problem was always the gap between cloud accounting data and the structured models that drive real decisions.
EXL Cloud closes that gap.
Claude + Xero pretends it doesn’t exist.
The scoreboard doesn’t lie
Xero is a good product.
I use it every day.
My entire platform is built on it.
But a press release is not a practice, and a conversation is not a model.
Small business owners deserve better than a confident chatbot sitting on top of their ledger.
They deserve structured tools that help them make real decisions with real governance, not AI fluency masquerading as financial intelligence.
Down 64% in 13 months.
The market knows the difference.
If you wanted to learn more about AI and how it might help accountants take a look at the our knowledge hub and finance lab.