All insights
Case study  /  21 August 2026

An AI agent for our invoices, that never touches the ledger

Quorum Agent

By Quorum Agent

The admin nobody enjoys

Supplier invoices are a small, relentless tax on every business. An email arrives with a PDF attached. Someone opens it, reads it, keys the supplier and the amounts into the accounts system, picks the right expense codes, splits the cost across clients where it needs splitting, attaches the document, and checks it is not a duplicate of one already entered. None of it is hard. All of it is repetitive, easy to get slightly wrong, and always a little behind.

In our own business it was quietly costing around six hours a week. One person put the invoice entry at roughly three hours. Others chipped in. Because it was being done inconsistently by hand, more time went on finding and fixing what had been keyed wrong, and reconciling it all added at least an hour on top. Six hours a week, the better part of a working day, every week, on a job nobody should be doing manually.

This is the "messy middle" that old-fashioned automation never cracked: too much judgement for a rigid rule, too dull for a person to enjoy. We had it in our own business. So rather than talk about agentic AI, we pointed it at our own books first.

What we built

An AI agent, built on Claude, wired to our Microsoft 365 invoice inbox, to Xero, and to our Halo PSA. The workflow is deliberately simple:

  • A supplier invoice email is dropped into an Invoice Inbox.

  • The agent reads the actual invoice, the PDF itself, not just the email text.

  • It matches the supplier against our records and proposes the coding, breaking the bill into line items and allocating them to the right client where appropriate.

  • It checks the invoice is not already in Xero, so nothing gets entered twice.

  • It creates a draft bill in Xero and attaches the original PDF.

  • The email files itself away as processed.

What used to be five minutes of careful typing becomes a few seconds of the agent doing the legwork and a person giving it a nod.

What we were actually trying to fix

Speed was the obvious prize, but it was not the only one. We set the agent four jobs:

  • Better-quality records. Consistent supplier matching and coding, instead of small human variations that make the accounts harder to trust later.

  • The source document, always attached. Every bill in Xero carries its original PDF: a clean audit trail, no hunting through inboxes when a question comes up.

  • Proper per-client breakdown. Where a supplier bill covers work for several of our clients, it is split into line items allocated to each, so costs land against the right client instead of in one undifferentiated lump.

  • Less manual data entry. The tedious keying, the part that is pure cost and no judgement, largely disappears.

Real invoices are messy, and that is the point

Tidy invoices are the easy case. The reason rigid, rules-based automation never solved this is that real supplier invoices are anything but tidy, and that mess is precisely where an agent with judgement earns its keep:

  • Poor-quality invoices, badly laid out, inconsistent, missing detail. The agent reads them the way a person would, instead of choking on a template it does not recognise.

  • Reverse-charge VAT, recognised and recorded correctly, rather than quietly mis-keyed.

  • No attachment. When the invoice is in the body of the email rather than a PDF, it captures the email itself as the document, so there is still a proper record in Xero.

  • Online-only invoices. When a supplier sends only a link, it is flagged so the document is retrieved, not lost.

  • Misallocation. Costs landing against the wrong client or code get caught, not baked in.

And when something genuinely cannot be handled cleanly, it is not forced through. It is set aside for a person to remediate. Knowing when not to act is as important as the automation itself.

Not just entry, a check on the books

The agent also works as a quiet quality layer over the accounts, not merely an input tool. It audits the invoice inbox and reviews existing entries, surfacing misallocations, spotting how something was recorded incorrectly, and flagging the invoices that need a human's attention. The effect compounds: the records get cleaner over time, not just faster to produce.

A quiet set of financial controls

Because something diligent now reads every invoice before it reaches the books, a set of useful controls falls out for free:

  • Bank-detail checks. It checks the bank details on invoices against what is on record, a simple guard against the "the supplier's bank details have changed" scam that catches out busy finance teams every year.

  • Unknown suppliers get flagged, not guessed. If a supplier is not one it recognises, it does not quietly create a new payee. It flags it for the finance team to check. Onboarding a new supplier stays a human decision.

For a small business, that is a layer of protection that would otherwise depend on someone being careful on a busy day.

The part that pays for itself: buy vs sell

The most valuable thing the agent does is not data entry at all. It reaches into two systems that do not normally talk to each other, Xero for what we buy and our Management System for what we bill clients, reconciles the two, and flags where they do not line up:

  • Where we are being overbilled by a supplier, an overcharge that would otherwise be paid without a second glance.

  • Where we have underbilled a client, a cost incurred on their behalf but never passed on. That is revenue quietly left on the table.

  • Discrepancies between buying and selling, per client, the margin leaks that are almost impossible to catch by eye across the ~150 invoices a month we handle, some of them running to over 180 line items each. Nobody reconciles that by hand, and that is exactly where money quietly leaks.

In our own books, that reconciliation has been surfacing around £2,400 a month, close to £29,000 a year, in under-billed costs, supplier overcharges and buy-vs-sell mismatches we would otherwise have missed. For any business that rebills costs, that is the difference between hoping the numbers add up and knowing they do: an afternoon of tedious admin turned into money found, and protected.

It runs to a rhythm, and reports back

Nobody has to babysit it. The agent runs on a daily schedule. It works through the invoice inbox on its own, then sends the finance team a report to review: the bills it drafted, and the problem invoices it has set aside for remediation. The team's day shifts from keying invoices to reading a short summary and approving the drafts. The machine keeps to a rhythm. The people stay in the loop by exception, where their judgement actually counts.

The line we will not cross: drafts only

Here is the design decision that matters most. The agent never posts to the ledger. Every bill it creates lands in Xero as a draft, waiting for a human to review and approve.

That is not a missing feature. It is the whole philosophy, the one we call Quorum Intelligence. The agent does the tedious, reversible ninety per cent: reading, matching, coding, checking, filing. A person makes the one decision that commits money. You get the speed of automation without ever handing over control of your finances, and if the agent gets something wrong, it is caught at a draft, not discovered in the ledger a month later.

Why we did it on our own books first

We will not ask a client to trust a system we have not run ourselves. Using it on Jera's own invoices means we feel every rough edge and fix it before it is ever pointed at someone else's accounts. It is the same reason this website runs on the very pipeline we build for clients: if it is good enough for our business, we can stand behind it for yours.

The path from here is well understood: the same agent and the same Xero logic, scaled up to a shared finance mailbox and batches of invoices, so a finance team can lean on it daily. The hard part, the judgement and the guardrails, is already done.

If your team keys invoices by hand

This is close to a textbook first AI agent: a high-volume, repetitive task with a clear right answer most of the time, made safe because a human still approves every result. It is the same start-small, keep-a-human-in-command approach we bring to every engagement, and invoices are rarely the only place it fits.

If invoice entry (or something like it) quietly eats your team's week, book a short discovery call and we will give you an honest view of whether an agent would earn its place. Not sure where to start? Our two-minute AI-readiness check will point you to a sensible first move.