AI voice for accountancy practices: the client line

Every accountancy practice knows the shape of the year. The phones are manageable for ten months, then the last fortnight of January arrives and the client line becomes a wall of the same questions: what do you still need from me, have you filed it, what do I owe, and when is it due. In 2026, more than 88% of enterprises use AI in at least one function, yet only around 6% capture material earnings impact from it, according to McKinsey's State of AI published in November 2025. Accountancy is a vertical where the gap between using AI and getting paid-for value is unusually easy to close, because so much of the seasonal call load is repetitive, rules-based and tied to fixed statutory dates.

The scale is not abstract. HM Revenue and Customs reported that 11,489,825 Self Assessment returns were filed by the 31 January 2026 deadline for the 2024 to 2025 tax year, with an estimated one million people still missing it. A practice does not field eleven million calls, but it does field its own concentrated version of that curve, and it fields it in the same fortnight that its qualified staff most need to be doing chargeable work rather than reading out deadline reminders.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

Why does an accountancy practice's phone line break during self-assessment season?

An accountancy practice's phone line breaks in January because demand spikes at the exact moment capacity is scarcest. Clients chase deadlines they left late, request copies of documents, and ask what they owe, all in the fortnight when accountants are buried in filing. HMRC reported that 475,722 people filed on the final day alone, so the practice absorbs the same last-minute rush its clients are living through.

The pressure is compounded by the state of HMRC's own service. The National Audit Office found that the average wait to reach an HMRC adviser rose from 5 minutes in 2018 to 2019 to nearly 23 minutes across 2023 to 2024, and that customers collectively spent the equivalent of 798 years, around 7 million hours, waiting on hold in 2022 to 2023. Gareth Davies, head of the National Audit Office, put it plainly: "HMRC's telephone and correspondence services have been below its target service levels for too long."

When clients cannot get through to HMRC, they call their accountant instead. The practice becomes the front line for a service problem it did not create, and the January peak is when that overflow is heaviest. That is precisely the load that voice AI is suited to, because most of it is answerable from the practice's own records without a qualified person on the line. It is a pattern we see across regulated verticals, tracked in our industries coverage.

What client calls can voice AI actually handle for an accountancy practice?

Voice AI handles the high-volume, rules-based calls that dominate an accountancy practice's inbound queue: deadline reminders, chasing missing records, confirming whether a return has been filed, booking appointments, and answering fee and payment queries. A platform such as Dilr Voice reads the answer from your practice records, states that it is an AI, logs the call, and escalates anything that needs professional judgement to a named person, so routine volume clears without a human picking up.

The single biggest win is the document chase. Every practitioner knows that most of a return is waiting on the client to send a P60, a set of bank statements, a mileage log or a dividend voucher. An outbound voice agent can call through a list of clients with outstanding items, tell each one exactly what is missing, and take the promise-to-send back into the workflow, writing the result back into whichever ledger the practice runs, whether that is Xero, Sage, QuickBooks or IRIS. Telephony sits on infrastructure like Twilio, so the calls place and record the same way your existing lines do.

Two things this is not. It is not the renewal-reminder motion that a membership body runs, which we cover separately in membership renewal reminders, because a tax deadline is a statutory date rather than a marketing renewal. And it is not the transaction-milestone updates a conveyancer sends across a twelve-week purchase, which we treat in conveyancing status updates. The accountancy pattern is anchored to the tax calendar, and that is what makes it distinct.

The same diagnostic logic underpins our AI operating model work, which maps where automation belongs in a practice before anyone commits to a deployment.

How does Making Tax Digital change the call pattern for accountants?

Making Tax Digital spreads the call load across the year instead of concentrating it in January. Under Making Tax Digital for Income Tax Self Assessment, sole traders and landlords with qualifying income keep digital records and send quarterly updates to HMRC rather than one annual return. That turns a single filing event into a recurring cycle, and every new quarterly deadline generates a fresh wave of confused-client calls, which is exactly the repetitive load voice AI absorbs well.

The timeline matters for capacity planning. According to GOV.UK guidance, Making Tax Digital for Income Tax becomes mandatory from 6 April 2026 for those with qualifying income over £50,000, drops to £30,000 from April 2027, and falls again to £20,000 from April 2028. Each threshold pulls a new band of clients into a quarterly rhythm, and each of them will phone the practice the first time they cannot work out what a quarterly update means.

For a practice, the strategic point is that Making Tax Digital makes voice automation more valuable, not less. A one-off January agent solves one peak. An agent wired into a quarterly cycle earns its keep four times a year across a growing client base, and it does so precisely when the practice is trying to absorb more clients onto the same headcount.

Where does a human accountant still need to take the call?

A human accountant takes the call the moment it turns advisory or touches anti-money-laundering duties. Anything that involves judgement, whether a client should incorporate, how to treat a specific expense, what a letter from HMRC means, belongs with a qualified person, and voice AI should route it there without attempting an answer. The same hard line applies to onboarding: a practice cannot begin chargeable work for a new client until it has completed client due diligence.

That boundary is a legal one, not a preference. Accountants, tax advisers and bookkeepers are relevant persons under the Money Laundering Regulations 2017, which require customer due diligence before an engagement begins. Regulation 28 defines verifying a client's identity as being done on the basis of information "obtained from a reliable source which is independent of the person whose identity is being verified." A caller asserting who they are on the phone is not, by definition, an independent source, so the AI can gather intake details but the regulated firm must complete the check and keep the record for five years.

This is where accountancy differs from the professional-services intake pattern we describe for client intake for law firms. A law firm's overlay is the SRA and privilege; a practice's overlay is HMRC and the money-laundering regime. Getting the hand-off right is the whole game: the AI should be confident on the rules-based majority and quick to escalate the regulated minority, which is the routing logic our AI execution office builds and monitors in production.

How does voice AI fit the document chase and your practice software?

Voice AI fits the document chase by turning an outbound call into a structured data event. Instead of a staff member leaving voicemails, the agent calls each client with an outstanding item, names exactly what is missing, and records what they commit to send. Because Dilr Voice writes the outcome back to the ledger the practice already uses, the chase becomes visible where the return is prepared, not trapped in someone's call notes.

Integration is the part that decides whether a deployment sticks. The agent needs read access to know what is outstanding and write access to log the outcome, connecting to practice platforms like Xero, Sage or IRIS and to the telephony layer, commonly Twilio, that carries the calls. It also needs to disclose itself. The EU AI Act requires that people are told when they are interacting with an AI system, and a practice that greets clients with a clear disclosure avoids both a compliance problem and the trust problem of a client feeling deceived.

Done well, the document chase stops being the bottleneck that pushes returns into the final week. The practice enters January with fewer outstanding items because the chase has been running quietly since November, and the voice AI agents handling it never tire of asking the same client for the same missing bank statement a third time. For the wider view of how this pattern repeats across regulated sectors, see our guide to voice AI automation by industry, and the closely related recall pattern in dental practices.

What does voice AI cost an accountancy practice, and where is the payback?

The payback for an accountancy practice comes from protecting qualified time during the peak, not from cutting reception headcount. If a voice agent clears even half of the routine deadline, document and status calls in January, it hands days of senior time back to chargeable work when it is scarcest. That is the ROI logic to model: value returned equals recovered fee-earning hours plus faster returns, measured against a predictable per-seat or per-minute platform cost.

The honest framing is that the number depends on the practice. A single-partner firm with 300 clients has a different curve from a twenty-partner group running Making Tax Digital across thousands. The right way to size it is to look at your own call logs from last January, count how many calls were rules-based versus advisory, and price the automatable share. A fixed-fee AI operating model review does exactly this before any build commitment, and the DATS methodology exists to make sure the money lands where the practice actually feels the peak. You can read more about our approach to placing AI inside a working practice, and about Dilr.ai and how we work, rather than bolting a tool on.

What is the best voice AI for an accountancy practice in 2026?

The best voice AI for an accountancy practice in 2026 depends on how regulated and how seasonal the workload is. For a simple, one-off reminder campaign, a lightweight self-serve builder such as Vapi, Retell AI or Synthflow can be stood up quickly and cheaply, and for a straightforward single-season blast that is genuinely the pragmatic choice. Those tools win when the job is narrow and the compliance surface is thin.

They win less clearly once the anti-money-laundering hand-off, Making Tax Digital's quarterly cadence and write-back into a practice ledger all have to work together. That is the case Dilr Voice and the DATS delivery model are built for: a regulated deployment where the routing between AI-resolvable and human-only calls is the hard part, and where a mistake at the AML boundary is a real risk rather than a cosmetic one. PolyAI is a credible enterprise alternative in that bracket, so the right test is to score vendors on integration, disclosure and escalation, not on demo polish. Ask any provider how their agent behaves at the exact moment a caller says something advisory or cannot prove who they are, which is why we build our voice platform around the escalation rather than the script.

Can voice AI complete AML identity checks over the phone?

No. Voice AI can collect the information a practice needs to start onboarding, but it cannot satisfy the Money Laundering Regulations 2017 alone, because verification must come from a reliable source independent of the client and a caller vouching for themselves is not that. The agent gathers intake details and books the human step; the regulated firm completes client due diligence and keeps the record. Treating the AI as the checkpoint would be a compliance failure.

Will clients accept an AI voice from their accountant?

Most clients accept an AI voice when it is disclosed, useful and quick, particularly for the routine calls they dread making anyway, like confirming what is still outstanding. Acceptance drops fast if the agent hides that it is an AI or blocks access to a person on a sensitive question. The practices that get this right lead with a clear disclosure, keep the escalation path one sentence away, and reserve the AI for the rules-based majority of the client line.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read about our approach to placing AI inside professional practices.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.