Accountancy onboarding: why intake quality matters before AI automation

The client you onboarded twice

In March a small Leeds practice signs a new limited-company client. The manager runs the identity check, asks for last year's accounts, the bank feed, the VAT history and a list of directors, and gets most of it back across nine emails, two WhatsApp photos of a passport, and one phone call where the client mentions — in passing — that there is a second company nobody asked about. The engagement letter goes out. The work begins. Good onboarding. Then in September a different team member preparing the year-end advisory note cannot find the source-of-funds answer, re-asks for the bank statements the client already sent, and re-discovers the second company as if for the first time. Nothing was missed in March. It was just never structured — so the practice paid for the same onboarding twice.

That is what happens to every practice that does onboarding well and stores it badly. The fix has a name: the intake spine — the structured backbone of everything you learn about a client before you advise them, captured as named, dated, sourced fields rather than scattered across an inbox. And the trap is the exact moment people reach for AI: "automate the onboarding." This briefing argues they are reaching one step too early. Here is the bet it is willing to lose on — a practice that fixes its intake structure first will get more out of onboarding AI within six months than one that buys an onboarding bot first. If the structured firm does not pull ahead, the frame below is wrong. We do not think it is.

“Nothing was missed in March. It was just never structured — so the practice paid for the same onboarding twice.”

So what is the intake spine, exactly?

Most practices treat intake as the boring bit you get through to reach the real work. That framing is the mistake. The intake spine is the structured backbone of everything you learn about a client before you advise them, captured as named, dated, sourced fields rather than scattered across an inbox. It has four vertebrae: identity (who they are, verified), scope (what they actually need, agreed in writing), evidence (the documents and source-of-funds facts that back the relationship), and risk (what about this client needs a closer look). Get those four standing upright and connected and you have a spine. Leave them as loose emails and attachments and you have a pile.

The distinction matters because of what sits on top. ICAEW guidance points to a new client receiving an engagement letter that sets out the agreed scope before any work begins, and to running client-acceptance checks at the start [1] — checks that in practice weigh up things like identity, the nature of the work, the firm's own competence and capacity, and any conflict. That is the spine being built by hand today. The thesis here is simple: AI is brilliant at putting that spine to work — drafting the letter, triaging the early queries, prepping the advisory note — and useless, or worse, at standing it up from a pile. Structure first, then automate. Automate a pile and you have automated a pile.

“Structure first, then automate. Automate a pile and you have automated a pile.”

The vertebra you can't skip: due diligence is law, not admin

One vertebra is not optional, and it is the one practices most often treat as paperwork. Customer due diligence is a legal obligation under the Money Laundering Regulations, supervised for accountants by ICAEW or HMRC depending on the firm [2][3]. The practice must verify the client's identity, understand the purpose and intended nature of the business relationship, identify beneficial owners, apply a risk-based approach, and keep the relationship under ongoing monitoring [2]. HMRC's own guidance for the sector is blunt about why the early questions matter: requesting supporting information helps you consider whether it is consistent with what you know about the purpose and nature of the relationship [3].

Read the Leeds story again with that in mind. The undisclosed second company is not a filing inconvenience — it is exactly the kind of inconsistency CDD exists to surface, and a pile-not-a-spine is how it stays buried until September. This is the part of onboarding the law already requires you to capture and keep, which makes it the highest-value vertebra to structure and the one where a clean, dated, sourced record is not a nice-to-have but the evidence you would put in front of a supervisor. Structure the obligation you already have, and you have built most of the spine for free.

Why now? The tools arrived; the order didn't

The intake-and-onboarding software market is real, named and maturing. Ignition sends proposals, engagement letters and payment terms and triggers onboarding when a client signs; Karbon chases outstanding client requests automatically and threads the team's tasks and emails around each client; Dext captures receipts and invoices, extracts the data and posts it; Glasscubes runs a client portal that collects documents and answers in a structured request rather than an email thread. A named UK practice makes it concrete: Hannah Barnes Accountancy Solutions cut the time it spends on engagement letters and AML compliance by managing them through BrightManager, the practice-management tool, per the vendor's published case study [8]. None of this is speculative. Practices are buying it now.

But notice what every one of those tools assumes: that you know what to ask, in what structure, with what evidence attached. The software automates the movement of the spine — the chasing, the routing, the drafting — superbly. It does not decide what the spine should be. That is the order most practices get backwards: they buy the tool that moves the structure before they have agreed the structure, and then automate the inconsistency. The window is open because the movement layer is genuinely good now. The trap is open for the same reason.

Where does AI help, and where must it stop?

On a clean intake spine, generative AI earns its place in three specific jobs. It drafts the engagement letter and the standard onboarding requests from the agreed scope. It triages the early client queries — sorting the "where do I send my receipts" from the "should I be a limited company", routing each to self-serve or to a person. And it prepares advisory groundwork — pulling the structured numbers into a first-draft note for the accountant to interrogate. ICAEW's own generative-AI guidance frames the boundary well: AI supports the work but the professional remains responsible and must be able to understand and explain the output [5]. Draft, triage, prepare. Not decide.

The line gets sharp where a decision becomes significant. Under UK data-protection law, where a decision is solely automated and has a legal or similarly significant effect on someone, a person has the right not to be subject to it and to obtain human intervention to contest it [6]. A practice will rarely make such a decision in routine onboarding, but the principle scopes the whole spine: AI prepares and proposes; a competent accountant accepts the client, signs the letter and owns the CDD risk call. Clients will trust a machine to assemble. They will not trust it to admit them — and they are right not to.

“Clients will trust a machine to assemble. They will not trust it to admit them — and they are right not to.”

What does it cost to skip the spine?

Start with the scale. The UK and Ireland profession doing this work is large — the regulators' own figures put it at over 408,000 members across the professional bodies in 2024 [7]. That is the non-vendor anchor: a hard, audited count of who carries this drag, and MikaHari's read is that much of it falls on small and mid-sized practices still running onboarding by hand. On duration, the vendor literature offers a working figure: Karbon's published guide describes onboarding running for weeks when handled by hand, which structured intake tooling compresses materially [4]. Treat that as a vendor figure, not a regulated statistic; the point holds whatever the exact number, because the same weeks are spent either way. The cost that does not show up on any invoice is the re-collection: when intake is a pile, every fact already gathered in week one is liable to be re-asked, re-sent and re-keyed months later. That is onboarding paid for twice — and re-collection is the most expensive line nobody costs, because it never reaches a timesheet as its own activity.

Each practice that automates a pile inherits three problems at once. The onboarding bot drafts confidently from incomplete intake, so the gaps look filled when they are not. The CDD record stays scattered, so the supervisory evidence is a reconstruction job rather than a file — and HMRC's sector guidance is explicit that the early supporting information exists precisely to test consistency against what you know about the relationship, which a pile cannot do [3]. And the advisory prep inherits whatever the intake missed, so the partner is checking the machine's homework against a pile, which is slower than doing it cleanly the first time.

Structure first inverts all three. A named, dated, sourced spine means the AI drafts from complete inputs, the CDD evidence is a file not a hunt, and the advisory note starts from numbers the firm trusts. The unglamorous truth is that the highest-return AI move in an accountancy practice this year is not buying intelligence — it is agreeing what you capture, in what shape, with what evidence, before you let anything automate it. The spine is cheap to stand up and expensive to skip.

MikaHari's view: agree the spine before you automate it

So, plainly: do not start by automating onboarding. Start by structuring intake. Agree the four vertebrae — identity, scope, evidence, risk — as named, dated, sourced fields, with the CDD obligation you already carry as the load-bearing one [2][3]. Then, and only then, let the named tools move that spine: Ignition to issue the letter, Karbon or Glasscubes to collect and chase, AI to draft and triage and prep [4]. Keep the accountant accountable for acceptance, the signature and the risk call, because both your professional body and the law expect exactly that [1][5][6].

The frame is one line you can carry into a partners' meeting: structure the spine, then automate the movement — never automate the pile. Keep the adviser. Automate the drag. Evidence the relationship. The Leeds practice did the hard part in March; it just stored it as a pile and paid for it again in September. The intake was never the boring part of onboarding. It was the part holding everything else up.