What AI should do with audit evidence.
A recent UK audit fine turned on scrutiny and evidence, not arithmetic. That is where AI and audit evidence should meet: each conclusion traced to its source.
We think the useful link between AI and audit evidence is not calculation. It is a link from every conclusion in the audit file back to the document that supports it, so a reviewer can see what each judgement rests on. The partner still weighs that evidence and makes the call.
One recent fine was about scrutiny, not sums
The Financial Reporting Council, an independent accounting watchdog in the UK, has fined Deloitte UK $7.9 million (6.05 million pounds) for flawed audits of Go-Ahead Group, as CFO Dive reports. The audits ran “during a four year period ending in June 2020”.
Subsidiaries of Go-Ahead “retained $39.6 million in erroneous over-payments”, and “the subsidiaries marked them as profits before and after Deloitte was appointed as auditor”. Penrose Foss, who leads investigations and enforcement at the FRC, pointed to “decisions and actions by GAG which were clearly questionable”. He said some of the company’s decisions and actions put very large amounts of taxpayers’ money at risk.
The findings against Deloitte were about what it did with those decisions. It “failed to enquire sufficiently into the actions of LSER and LM”, “failed to apply sufficient professional scepticism” and “failed to evaluate the evidence indicating the existence of fraud risk factors”. It also “failed to appreciate that the company was under a contractual obligation to act in good faith”. Foss described “a highly concerning pattern of failure by Deloitte to apply sufficient scrutiny”.
The FRC “reduced the penalty from $14.5 million”, citing Deloitte’s “exceptional cooperation” and its “admissions and early disposal” of the matter. Deloitte UK said it “regrets that aspects of its audit work did not meet the standards expected”.
This is one case, and one case does not describe audit enforcement in general. But read the list of findings. None of them is a column that did not add up. They are about enquiries not pushed far enough, evidence not weighed and an obligation not recognised.
Judgement fails where the evidence goes quiet
In our experience, an audit file is mostly conclusions. Revenue recognised correctly. Balance supported. No indication of fraud. Each is a sentence somebody wrote and somebody signed.
We think the question worth asking of each one is simple: supported by what. If the answer is a document, the conclusion stands or falls on that document. If the answer is the memory of a senior who has since left, the conclusion is an assertion. It looks fine. It has nothing under it.
In the files we see, the chain from judgement to evidence is where weakness hides, and it tends to be checked by reading, often late. The sums are rarely where the trouble starts.
What a system worth trusting should do
Every conclusion should carry its source. A statement in the file should point to the contract, the bank letter, the board minute or the ledger extract it relies on. The reviewer should be able to read the page itself, not a summary of it.
A gap should be shown, not filled. When no document supports a conclusion, a system worth trusting says so. Ask your vendor to show you what happens when the evidence is missing. We think a confident sentence with no source is the most dangerous thing a file can hold.
Evidence that points the other way should sit next to the conclusion. A letter about a payment received in error, or a clause setting out a duty to act in good faith, belongs beside the judgement it challenges. It should not be filtered out as noise.
The partner should decide. We do not think software can be sceptical on anyone’s behalf. At most it can make contrary evidence hard to miss and an unsupported conclusion easy to see. Weighing it and signing remain the partner’s job.
Scepticism stays with the people who sign
The finding in this case was that the auditor did not apply sufficient scepticism and scrutiny. No tool supplies either. What a well-built tool can remove is the excuse that the contrary evidence sat in a folder nobody opened.
That is the rule Syntexia builds to: every result points back to its source document, and a person always decides. Client data never trains a model. The point is not a faster file. It is a file where each judgement can show its working.
Where to start
Take one closed engagement. Pick a handful of the judgements that mattered most, the ones a reviewer would ask about first. For each, find the document that supports it, note how long the search took, and whether you found it at all.
The answers tell you where the file is strong and where it only reads well. If you would like to talk through what you found, tell us.