Owning the narrative: what AI really means for your practice

If what you’re hearing about AI makes me you feel either uneasy or excited, that’s normal. And do you know why both reactions are completely fine? 

Because we’ve done this before. The move to the cloud wasn’t instant either. 

Many practices took time to adjust, but those who invested the effort saw results. Now, we’re doing it again, with the cloud and AI together.

As an accountant, I’ve spent months learning from Xero’s AI technologists. Last week, my Xerocon keynote aimed to share these practical insights around the question, what does AI mean for your practice?

Two answers to the same question

Recently, two AI platforms gave me different travel suggestions for a holiday in Wales. That’s fine for a holiday, but inconsistent answers for a tax return would clearly be a disaster.

This tension highlights the difference between deterministic and probabilistic reasoning: and that is  key to knowing where AI fits in your practice.

Show a room full of people an invoice for £20 and ask for the total, and everyone agrees it’s £20. A clear answer, with no context or judgement needed. That’s deterministic reasoning.

Now hand a set of identical bank statements to people in that room and ask them to reconcile it. You’d get slightly different results. Not because anyone’s wrong, but because the correct answer depends on things only you know: how a client codes things, what a payment was really for, how your practice handles certain suppliers. That’s probabilistic reasoning, and it’s exactly the territory Auto Bank Reconciliation is built for.

Think of your statement lines as a layer cake. At the bottom of the cake are the obvious matches, a clean tie to an invoice or a rule you’ve already written. Those run on deterministic logic, the same kind of models that have powered bank reconciliation for years, and they get confirmed automatically. Further up, it gets more nuanced: lines that need context about how a client usually behaves. That’s where machine learning sits, spotting the pattern and only acting when it’s genuinely confident. At the very top, a small slice uses generative AI to match a client’s own naming conventions. Anything that doesn’t clear the confidence bar comes straight back to you for review.

Put it together and Auto Bank Rec currently reconciles 65% of transaction lines at 96% accuracy. The way to think about it in your practice is as a new starter. You train it, review closely for the first few months, and then use your judgement about where to step back and where the material items still need your eye. It learns from every correction you make, specifically for that client. Over time, the accuracy climbs and the review time shrinks.

The principle that falls out of all this: most financials are deterministic. Your workflows can be probabilistic. 

Anything that touches the actual numbers, tax calculations, think of the figures your auditors would rely on, has to produce the same answer from the same input every single time. The process around those numbers, the investigating, drafting, summarising and advising, is where probabilistic AI can genuinely earn its keep, because that’s where your time is going.

Your numbers must stay safe. Your process should change.

Context and the harness

So if the AI underneath is probabilistic, how can Xero guarantee the right answer where it matters? By using probabilistic technology, but building a structure around it that enforces a deterministic outcome. There’s a name for that structure: the harness.

Think about the best junior you ever hired. Plenty of potential, but you didn’t hand them the client list on day one. You gave them procedures, sat them next to experienced colleagues, checked their work before it reached a client: not to hold them back, but to make their potential safe and useful. That structure is the harness.

It’s worth separating the harness from something it often gets confused with: context. Context is what you feed the model so it answers your specific question, things like a client’s history, coding preferences and the business situation. More context means a better, more relevant answer.

The harness is different. It’s the system that sits around the model altogether: the security controls, the permissions, the connection to your actual Xero data, the audit trail, and the hard rule that certain calculations have exactly one right answer the AI cannot deviate from. That’s only possible because of twenty years of accounting logic and tax rules already built into Xero. The harness is what makes the answer safe, governed and repeatable, whatever the underlying model tries to do.

Context makes the answer relevant. The harness makes the answer trustworthy.

A public chatbot can use context. It doesn’t come with a harness built for your work. When you use AI through Xero, you’re working inside our controls, our permissions, with your client data governed exactly as it’s always been governed. 

So with all of this in mind, the question worth asking isn’t “should I use AI”, it’s “how do we adopt it intentionally?”

Building the modern practice

And this is less of an AI question than it sounds. It’s a strategy question. Which clients you want to serve, what you offer them, your competitive edge, and how you configure your people, processes and platforms to deliver on it. 

AI changes what’s possible, and it’s changing fast, so this is the moment to ask those questions again through a new lens.

A great metaphor for the modern practice is like a ship, built in three layers from the bottom up.

The engine room is where client data lives and moves: document capture, bank feeds, auto-categorisation, the continuous ledger. This is exactly where AI should be doing the heavy lifting rather than your team, whether that’s Smart Document Capture pulling data straight from invoices and receipts, or Auto Bank Rec handling the routine and surfacing only the exceptions. 

Shaky data here means shaky insight everywhere above it, so this is where to start. 

Get the engine room right and you get time back, and that time is what unlocks everything else.

The bridge is where a clean data foundation lets your practice look forwards instead of backwards: forecasts, anomaly detection, benchmarks running continuously rather than only at year end. This layer is a genuine mix of machine and human, and it’s where context matters most. 

Picture two clients with identical profit and loss statements, same revenue, same costs, same margin, purely by coincidence. One is caring for elderly parents and saving for a child’s university fees. The other wants to sell the business in two years. 

The AI might surface the same observations for both, but the right advice is completely different, resilience and a buffer for one, positioning the business for a buyer for the other. That judgement is yours, and no model can replicate it. 

Every rule you write and every note you add feeds this layer, building a body of context that becomes one of the most valuable assets your practice owns.

The deck is where your client actually experiences the journey with you. This is where the strategy question becomes real: a sole trader might need efficient, reliable compliance and someone to call when things get complicated; a growing business might need cash flow forecasting and a regular sounding board; others might want a fully outsourced finance function, the kind of support that used to require an internal team of ten. 

Compliance isn’t going away, and it doesn’t need to. 

In the modern practice, the deck is where the client boards, and delivered efficiently, is used as the launch pad for everything that follows. A question about a tax bill can become a conversation about pensions, then retirement, then eventually selling the business. That’s the voyage, and you’re the one they trust to navigate it.

Why this matters

Nearly 800 small businesses close in the UK every day. Two in five don’t know whether they were profitable last month. That’s not a failure of capability, it’s because those owners are busy running restaurants, building sites and workshops, not reconciling bank statements. 

They need someone brilliant with the numbers, and our research consistently shows that the small businesses that thrive are the ones working closely with an accountant or bookkeeper. Not marginally better outcomes. Significantly better ones.

That’s why the harness matters beyond safety. It’s the foundation of trust that makes the whole relationship between your firm, your clients and AI possible, all built on one system, one audit trail, from the same data.

The bit that doesn’t change

AI won’t replace accountants; 94% of UK firms see it as a way to free up capacity. As technology improves, the need for human judgement and trusted relationships only grows.

Your role is to lead your team through this transition. AI is simply a new instrument; how you play it, using your experience, is something no model can replicate. You define what AI means for your practice.