Every practice inherits files. A new client arrives with three years of history built by someone else, a colleague hands over a book mid-year, or you realise you have not looked closely at a long-standing client since the software started doing more of the work. The question is the same each time: how good is this file, really? Not how tidy the dashboard looks, but whether the figures underneath would survive an HMRC enquiry with your name on the filing.
The software will not answer that for you. Reconciled does not mean right. A bank feed proves that money moved; it says nothing about whether the movement was coded correctly, given the right VAT treatment, or supported by a document. What follows is a method you can run on any client book in an hour or two, whatever package it lives in. It is not a statutory audit and does not pretend to be one. It is a structured quality review, and it finds most of the defects worth finding.
Set the scope before you open anything
Pick one period. The most recent complete VAT quarter is usually the right choice, because the coding decisions are fresh and the VAT consequences are still correctable. Then export four things: the trial balance, the full nominal activity for the period, the aged debtors and creditors listings, and the workings behind the last VAT return. Everything below runs off those four reports, and every mainstream package can produce them.
1. Sample-check coding against source documents
Pull 25 to 30 transactions, but do not pull them at random alone. A useful stratified sample is: the ten largest expense postings in the period, ten chosen at random, five from the accounts most prone to abuse (repairs, motor expenses, subsistence, sundry), and anything suspiciously round. For each one, ask three questions in order. Does a source document exist at all? Do the date, supplier, net, VAT and gross on the document match what was posted? And is the nominal code defensible, meaning you could explain it to an inspector without wincing?
Record the hit rate on each question separately. A file can score well on existence and badly on coding, and those are different problems with different fixes. Resist the temptation to check the ledger against the bank feed and call it verification. That only proves the cash happened.
2. Interrogate suspense and sundry
Suspense should be nil at every period end. A balance in suspense is not an error in itself; it is a pile of unfinished decisions, and the useful question is how old they are. An item parked last week is workflow. An item parked two VAT quarters ago has already flowed through a return with the wrong treatment, or none.
The sundry expenses account deserves the same suspicion for the opposite reason. Nothing sits visibly unresolved there; difficult coding decisions simply go there to hide. Look at the trend across periods. A sundry balance that grows faster than the business tells you someone, human or software, was guessing and nobody was checking the guesses. Open the ten largest sundry postings and recode what you find; in my experience most of them belong somewhere specific.
3. Scan for VAT anomalies
Run down the input VAT detail behind the return and look for two patterns.
- Standard-rated where zero or exempt was expected. Scheduled passenger transport in vehicles designed to carry ten or more people is zero-rated, so a train or bus fare showing reclaimed input VAT is wrong on its face; a taxi fare is standard-rated, so there a reclaim can be genuine. Insurance is exempt. Postage stamps carry no VAT to reclaim. Most bank charges are exempt too, though some specific fees are not, so check the document rather than assuming either way.
- Input VAT on blocked categories. VAT on entertaining clients is blocked, however commercially sensible the lunch was; entertaining staff is treated differently, and there is a narrow exception for overseas customers. VAT on buying a car is blocked unless the car is used exclusively for business and is not even available for private use, a test HMRC applies strictly. Leased cars normally carry a 50 per cent block on the rental VAT where there is any private use.
Then flip to outputs and scan for the reverse: standard-rated sales coded as zero-rated or outside the scope. None of these checks needs software; sorting the VAT detail by rate and by account gets you most of the way. If the client is heading into a new MTD obligation, this scan belongs in your onboarding routine as well; there is a fuller treatment in the MTD VAT practice checklist.
4. Measure the evidence attachment rate
Go back to your sample of 25 to 30 transactions and count how many have a source document attached or linked in the file. Express it as a percentage. I know of no published benchmark to compare against, and I would be wary of anyone who offers one; the useful comparison is this file against itself last quarter, and against the other files in your practice.
The legal backdrop is simple: VAT records must normally be kept for at least six years, and a line on a bank statement is not a VAT invoice. A file with a low evidence rate is not necessarily wrong, but it is undefended, and an enquiry will treat undefended and wrong as close cousins.
5. Reconcile control accounts to sub-ledgers
This is the fastest test in the method and the one most often skipped. Compare four pairs: the debtors control account against the aged debtors listing; creditors control against aged creditors; the VAT control account against the return actually filed for that period; and the payroll liability accounts against the payroll reports for the same months.
Differences of pennies are rounding. Differences of round hundreds or thousands are almost always a manual journal posted directly to a control account, which brings us to the next check.
6. Look for manual journals near period ends
Filter the journal listing to the last week of the period and the first week of the next. Accruals, prepayments, depreciation, payroll postings and honest corrections all legitimately live there, and each of them should carry a narrative that explains itself. What you are hunting is the other kind: journals with no narrative, journals that move balances between control accounts, journals that empty suspense into sundry two days before the quarter closes, and journals that exactly reverse a figure that would otherwise have looked bad.
A defensible journal answers three questions: who posted it, why, and on the strength of what evidence. If the file cannot answer those for its period-end journals, note it as a finding even when the numbers happen to be right.
7. Check period-close discipline
Two tests. First, is there a lock date, and does it actually sit at the end of the last filed period? Second, re-run the numbers for a period that has already been filed and compare them with what was submitted. If they no longer match, someone has edited a closed period, and every report you print from that file now disagrees with what HMRC holds. The discipline you want to see is corrections made by reversal in an open period, with the original left intact, never by editing history.
What automated capture gets wrong, and what to sample for
Most files now contain thousands of entries that no human keyed in. Whatever capture tool produced them, the same failure families recur across the industry, so aim part of your sample directly at them.
- Dates misread across formats. A document dated 03/04 can be read as 3 April or 4 March depending on the format the tool assumed. Sample transactions posted within a few days of a period end and check the date on the document itself; a misread date silently moves expenses between VAT quarters.
- Line items confused with totals. Capture can post one line's amount instead of the invoice total, or a subtotal instead of the gross. Check your largest sampled invoices line by line against what was posted, not just heading to heading.
- Duplicates posted twice. The same invoice often reaches a file twice: once as an emailed PDF, once as a photograph of the paper copy. Sort the period's purchases by supplier and amount and scan for pairs; where the client's suppliers issue statements, check the statement against the ledger.
- VAT treatment inferred rather than read. A faded thermal receipt with no visible VAT breakdown does not stop some tools from assuming standard rate and claiming input tax anyway. Any reclaim in your sample should trace to a document that actually shows the VAT.
The common thread is that these are quiet errors. Nothing crashes, nothing flags, the file balances. That is why sampling against source documents beats any amount of dashboard review, and why it matters whether your tools ever say I am not sure. I have written separately about what confidence scores actually mean and about the difference between review-first and silent automation.
Make it a routine, not an event
Turn the findings into a one-page score per file: coding hit rate, evidence rate, suspense age, control account differences, unexplained journals, close discipline. Imagine a 40-client firm reviewing one file a week on this method; the whole client base rotates through a quality check roughly once a year for the cost of a couple of hours weekly, and the score sheet tells you where training, pricing or a difficult client conversation is needed. Fix what you find by reversal, in the open period, with a note. The point is not a perfect file. The point is a file whose imperfections are known, dated and explained.
Where AIONA fits
AIONA was built around these checks rather than bolted onto them: every extracted figure carries a GREEN, AMBER or RED confidence tier, low-confidence items cannot post until a person has reviewed them, and every journal links back to the source document that produced it. Posted entries are immutable, so corrections happen by reversal and the audit trail is tamper-evident. If a client's book lives in Xero, you can connect it read-only and AIONA will grade it for mis-codings, VAT exposure and evidence gaps without writing anything back.