The pattern that actually works: filter-and-flag, human-in-the-loop
the single most useful thing AI does in an accounting workflow is narrow the set of things a human needs to look at. not produce the finished output — narrow the review. that distinction matters more than it sounds.
a bookkeeper reviewing 400 bank transactions for anomalies will catch most of them. they will miss some — not because they’re careless, but because sustained attention on repetitive data degrades. an AI model reviewing the same 400 transactions and flagging 18 for human review changes the cognitive task entirely. the human is now reviewing 18 items with full attention, not 400 items with declining attention.
this is different from “AI does the bookkeeping.” the AI is not categorising, reconciling, or producing the finished output. it is pre-filtering so the human does the high-value part of the job better. that’s filter-and-flag. it’s less exciting than “AI replaces your accountant” and considerably more useful. that pre-filter sits on top of our core bookkeeping process — same accountants, same standards, just better aimed.
“AI is better at narrowing what you review than at producing finished work. the value is in the pre-filter, not the output.”
the other pattern that holds up is structured input, structured output — where AI is used to draft content that has a defined format and low tolerance for factual invention. accounts commentary drafts, exception summaries, client-facing variance explanations — these have structure, they have constraints, and a human is reviewing before anything is sent. where AI fails reliably is in tasks that require factual precision (Companies House filing data, IRS form values, specific compliance calculations) or professional judgement (engagement letters, complex tax positions). those require humans. they always will.
Where we use AI in our offshore workflows
these are live applications — not pilots, not things we’re testing, not things we’ve read about. each one has a clear scope and a clear human review step before output leaves our team.
Where we explicitly don’t use AI
the failure modes of AI in accounting are specific and predictable. knowing them is more useful than pretending they don’t exist.
Engagement letters and client-facing compliance documents
an engagement letter with an incorrect date, wrong fee, or missing scope clause is a professional liability issue. AI generates plausible text — it does not guarantee accurate specifics. we write these from templates maintained by humans, reviewed by humans.
Tax form values and compliance calculations
AI is not used to produce IRS form values, state tax calculations, or compliance-critical numbers that your CPA will sign. these are calculated by accountants, reviewed by senior accountants, and cross-checked against source data. the audit trail is human throughout.
Anything requiring professional judgement
complex tax positions, ambiguous categorisation decisions, the treatment of related-party transactions — these are flagged for human judgement, not resolved by AI. a model that answers with confidence on ambiguous professional questions is the most dangerous kind. we don’t use them for this.
Client communication without review
AI-drafted client communications are reviewed and approved by a human before sending. full stop. no AI-generated text reaches a client directly.
narrow the brief, specify the output, review the result. if any of those three steps isn’t happening for a given AI use case, it’s not in our workflow yet.
The tools we work with — and what each actually delivers
one short assessment per tool. no marketing language, no vendor pitch, just what we’ve found in production. a fuller view of every platform we operate in lives on our software we work in page.
What this means for CPA firms working with us
the practical implication for your CPA firm is threefold.
Faster turnaround without additional headcount.AI-assisted workflows compress the time it takes to process document-heavy clients and high-volume bank feeds. the efficiency gain goes directly to turnaround time — clients whose books used to take 3 days to close now close in 1.5. that’s capacity you didn’t have to hire for. the rest of how we partner with CPA firms sits in the same logic — extending your capacity without adding to your overhead.
Higher-quality exception flagging than pure manual review.your clients’ books are reviewed by a certified accountant assisted by anomaly-detection tools. the combination catches more than either alone. the errors and fraud signals that exhausted-human review misses at 11pm get caught by the AI pre-filter earlier in the process.
Your clients still see human accountability throughout.the work that leaves our team and reaches your clients has been produced by a certified accountant and reviewed before delivery. the AI is an internal process tool, not a client-facing system. your clients don’t see the AI layer — they see accurate, timely, well-presented work from a team you trust.
if the question behind your question is “will AI make this cheaper?” — the honest answer is that AI makes the same-quality work faster. it doesn’t replace the accountant. it makes the accountant more effective. the cost structure improves because more gets done per accountant hour — but the human is still there, and still accountable.
Where AI is going in accounting — and where it isn’t
the next 2–3 years in AI for accounting will produce better document extraction, better anomaly detection, and better first-draft generation. those are genuine improvements on capabilities that already exist. the tools will get more accurate and less expensive.
what won’t happen in 2–3 years: AI autonomously managing client books without human oversight. the professional liability structure of accounting doesn’t permit it — a CPA firm cannot hand their clients’ financial statements to a black box and put their signature on the output. the regulatory and liability framework changes far more slowly than the AI tools do.
the firms that will use AI most effectively in the next 3 years are the ones that invest now in understanding its real failure modes — not the ones that adopt it fastest. a tool adopted wrong is worse than no tool at all. our approach is to add each AI application only when we understand both what it does and where it fails, when the failure mode has a human backstop, and when the output quality can be verified by an accountant before it matters.
AI and automation address workflow efficiency. if the underlying problem is that your client’s data lives in six systems that don’t talk to each other, that’s a different problem — and the solution is a different conversation. see our ERP advisory page for that one.