Yes, but a possible omission is not a debt
AI can compare evidence of completed work with invoice records and flag jobs that appear to have slipped through. It cannot establish, from a completion note alone, that a customer owes money. Some completed work is covered by a retainer, included in an earlier invoice, still awaiting acceptance or being corrected under warranty.
The useful product is therefore an exception report for a person to investigate, not an autonomous debt generator. Each suggested omission should show the job reference, customer, completion evidence, candidate invoice matches and the reason it was flagged. If a supplier cannot show those links, the tool may be producing an interesting story rather than a reliable billing control.
This is different from chasing overdue invoices. An overdue-invoice report starts with an invoice that exists. Missed-billing detection asks whether work that may be chargeable has reached the invoicing process at all. Buying another payment reminder tool will not necessarily solve that gap.
The strongest use case is a service business with enough jobs, timesheets or approved extras that checking every completion manually is becoming expensive. The weakest is a business with no dependable record of what was agreed or delivered. AI cannot reconstruct contractual entitlement simply by guessing what a busy employee probably meant.
When would existing reports be the better option?
Before paying for AI, ask your bookkeeper or software administrator whether your existing job-management and accounting systems can report completed but unbilled work. If every job has a unique reference carried through to the invoice, an ordinary comparison may be more accurate, cheaper to maintain and easier to explain.
Xero, QuickBooks and Sage are examples of accounting products worth checking alongside your job system. Do not assume a particular plan includes the exact report or integration you need. Ask the supplier to demonstrate the relevant feature using your record structure, and include existing reporting options in any buying comparison.
A spreadsheet or rules-based match is often sufficient when the inputs are structured. Compare completed job identifiers with invoice-line identifiers, exclude cancelled jobs, and let a colleague review the remainder. No language model is needed just to find an identifier missing from another list.
AI earns its place when the evidence is less tidy: a job described differently in an engineer note, an approved variation hidden in correspondence or several small tasks billed together under a different description. Even then, keep exact matches and arithmetic deterministic where possible. The buying question is whether interpreting messy evidence improves the exception review enough to justify the additional uncertainty and cost.
What information does the system need?
You need a reliable record of the work, the agreed commercial basis and the relevant invoicing history. Depending on the business, that might mean job cards, approved quotes, timesheets, customer purchase orders, completion confirmations, invoice lines and credit notes. Start by checking whether these records can be accessed lawfully and exported usefully.
A shared customer name is not a dependable join key. Two branches may use similar names, a customer may pay through a parent company, or one purchase order may cover several visits. Your supplier should explain how the system handles those relationships and when it refuses to make a match.
Ask how it treats partial completion, staged invoices, deposits, bundled work and warranty visits. A missing final invoice may be a real issue, but an apparent gap may also be a perfectly valid commercial arrangement. The system should preserve that distinction instead of classifying every unmatched job as lost revenue.
Document who can see the underlying information. A billing assistant does not automatically need access to every customer email or attachment. A narrow export or restricted read-only connection may be enough. If the supplier proposes unrestricted access because it is easier to configure, that convenience is a cost to assess, not a reason to skip the question.
What must a person check before an invoice is raised?
The reviewer should confirm that the work was completed, chargeable under the agreement and not already billed elsewhere. They should check the agreed price, purchase-order requirements, any approval for extra work, applicable tax treatment and whether a dispute or correction changes the position. A model confidence score does not replace those checks.
The distinction between documents matters. GOV.UK invoicing guidance says an invoice is not the same as a receipt: the latter acknowledges payment. It also explains the invoice requirement where both seller and customer are VAT registered in a business-to-business transaction. An AI assistant should not treat a receipt, quote and invoice as interchangeable evidence.
HMRC provides a useful numerical example in its VAT charging guidance: £60 excluding standard-rate VAT becomes £72 including VAT. The standard rate shown is 20%. That is not a rule that every job attracts 20%; exemptions, reduced rates, zero rating and other circumstances must be handled correctly.
These examples explain why the assistant should flag and support a decision, while your normal accounting controls produce the invoice. Do not let it invent invoice numbers, amend posted records or send a customer an unexpected charge based solely on a generated interpretation of a job note.
What would make this worth paying for?
Measure confirmed omissions, review effort and errors. Count how many flagged jobs are genuinely billable, how many are already covered and how many cannot be resolved from the evidence. Separately track whether correctly raised invoices are paid. A flagged value is not cash recovered and should not be presented as achieved revenue.
Consider a clearly hypothetical example, not a client result or market benchmark. A review flags twelve jobs. A person confirms that four should be invoiced at £150 each excluding VAT, producing £600 of potential additional billing. The other eight are already covered or not chargeable. The headline is not twelve recovered invoices.
Suppose the review takes two hours at an assumed internal cost of £30 per hour, or £60, and assume another £40 for that period of software and operating costs. The illustrative direct cost is £100. That still does not prove £500 profit: you must consider collection, delivery costs, disputes and whether the same omissions would have been found during the ordinary month-end review.
Ask suppliers to price discovery, data preparation, integration, exception review, maintenance and support separately. Do not accept an invented recovery percentage as a substitute for testing your records. A short evaluation should establish whether there is a repeated, addressable gap before anyone promises an expensive custom build.
What should a supplier prove in a demonstration?
Use a bounded historical sample for which your finance team can establish the correct answer. Include genuinely missed items, work already invoiced under another description, cancelled jobs, credits and non-chargeable return visits. A demonstration made only from obvious omissions does not tell you whether the product will annoy staff or customers.
Require explanations tied to source records. For each flag, a reviewer should be able to inspect the evidence and understand why the proposed invoice match was rejected. Include a category for unresolved cases. Forcing every job into either billed or missed conceals uncertainty and can make reported accuracy look better than the operational reality.
Check the access boundary as carefully as the matching. Provision 2.6 of the UK government's voluntary AI Cyber Security Code of Practice says permissions on connected systems should be limited to what the functionality requires and risk assessed. A read-only omission report normally should not need authority to issue credits or alter bank details.
Agree who maintains the connection when an accounting export changes, who investigates incorrect matches and how you stop the service. Compare the tool with your existing manual or rules-based review on the same sample. Choose the approach that produces useful, explainable findings at an acceptable total cost, not the one with the most confident demo.
When this is NOT right for your business
Do not start here if you have no consistent completion records or agreed prices. Your first investment should be a clearer handover from delivery to billing. Otherwise staff will spend their time debating missing information rather than confirming missed invoices. Improving the ordinary process may remove the need for AI entirely.
It is also a poor choice when job volume is low and a weekly manual check takes only a few minutes. A new subscription, integration and review obligation can add more work than it removes. Where everything already shares a stable job identifier, prefer a simple report unless testing establishes a genuine limitation.
Be particularly cautious with old, disputed or sensitive matters. Contract terms, limitation questions, tax timing and professional obligations can require specialist advice. Discovering an old job note is not permission to issue a charge without checking the circumstances. Keep sensitive customer and employee information out of a trial unless the processing arrangements have been assessed.
If your records are sufficiently complete and the omissions are recurring, AI-assisted review may be a sensible purchase. Keep the initial scope narrow: identify possible gaps, show evidence, obtain human approval and retain the ordinary invoicing process. If you want a second opinion, bring a small anonymised example of the gap and compare the existing-report option before commissioning anything bespoke.
Is This Right for You?
This can fit service businesses with reliable completion records, recurring billing omissions and enough volume to justify review. It is not a substitute for agreed prices, good job records or finance oversight.
Start with your existing accounting and job-management reports. Consider AI only where interpreting inconsistent descriptions creates measurable additional value.
