A 14-person UK accountancy practice received 340 inbound client calls in the six weeks before the January self-assessment deadline. Their two admin staff took every one manually. We pulled the call log: 71% fell into four categories — submission status, payment reference, missing reminder letters, and partner connection requests. The first three have a deterministic answer from Xero Practice Manager. That is 241 calls that required no human judgement whatsoever, each one pulling an admin away from billable support work at the worst possible time of year.
The fix is not a FAQ page. Clients will not read it at 5pm on 28 January. The fix is a voice triage agent that answers the call, queries the practice management system live, and gives the client a specific answer about their return — not a generic one. Here is how to build it.
Inbound call triage for UK accountancy firms: the eight query types that do not need a human
The four categories in that call log were not unusual. Running a broader analysis across three practices with combined call volumes of over 1,200 calls across two years, we identified eight query types that repeat consistently and require no professional judgement to answer:
| # | Query type | % of total calls | Primary data source |
|---|---|---|---|
| 1 | Self-assessment submission status | 22% | Xero PM / IRIS |
| 2 | Payment reference for HMRC | 18% | Practice billing record |
| 3 | Missing reminder or engagement letter | 15% | Document management log |
| 4 | Request to speak to named partner | 12% | Staff directory + availability |
| 5 | UTR number retrieval | 9% | Client record |
| 6 | Companies House filing deadline | 8% | CH record or internal log |
| 7 | VAT return status | 7% | Xero / IRIS |
| 8 | Payroll run confirmation date | 5% | Payroll module |
Those eight cover 96% of eligible calls. The remaining 4% — HMRC investigation notices, complaints, and distressed callers — need a human on the line immediately. Designing the triage system well means being precise about that boundary, not maximising the automation percentage.
Call flow design for accountancy triage: routing by intent without misclassifying urgent escalations
The failure mode here is over-routing. Some agents send everything that mentions "HMRC" or "investigation" to a callback queue. That is not a safety net — it is a broken handoff that leaves a stressed client waiting three days while their investigation letter sits unanswered.
Design the intent classifier as a three-tier system:
- Tier 1 (Look up and respond): Intent is clear, data is in the system, answer is deterministic. Agent retrieves and reads back.
- Tier 2 (Acknowledge and schedule callback): Intent is ambiguous or data requires professional review. Agent confirms receipt, logs the query, and offers a specific callback slot from the staff calendar.
- Tier 3 (Live transfer immediately): Intent signals an emergency, distress, or escalation. Agent says "Let me connect you to someone right now" and bridges the call — no voicemail, no recorded message.
The routing logic lives in a structured configuration fed to the agent's system prompt. Here is the core segment from our production setup:
{
"intent_routing": {
"tier_1_phrases": [
"when will my return be filed",
"has my return been submitted",
"what is my UTR",
"payment reference",
"have you received my documents",
"when is my VAT return due"
],
"tier_3_phrases": [
"HMRC investigation",
"compliance check",
"code of practice",
"COP8",
"COP9",
"been contacted by HMRC",
"I can't pay",
"going bankrupt",
"bailiffs",
"I'm really worried",
"desperate"
],
"default_tier": 2,
"escalation_override": "Any phrase in tier_3_phrases takes priority over all other intent signals regardless of surrounding context"
}
}
The tier_3_phrases list primes the LLM, but the system prompt also instructs the agent to escalate on any expression of financial distress regardless of exact phrasing. "Desperate" and "bailiffs" appear rarely; when they do, Tier 2 routing is too high a cost.
For a deeper look at call flow JSON design, see our post on call flow design for voice agents.
Integration with Xero Practice Manager and Iris: surfacing the live client record during the call
Both Xero Practice Manager and IRIS have API access. IRIS OpenSpace exposes a REST interface; Xero Practice Manager works through Xero's standard API token framework via the Xero Developer documentation.
The no-developer route uses a middleware layer: n8n or Make sits between the voice agent and the PMS. The agent fires a webhook with the caller's phone number or UTR (collected in the opening seconds via speech or DTMF), the workflow queries the PMS API, and the result lands back in the agent's context before the client has finished stating their query.
Latency matters here. Our benchmarks show the round-trip adds 600–900ms. That is acceptable if the agent's opening sentence buys the time — "Let me pull that up for you right now" covers the gap. If the PMS API is slow, pre-fetch the client record at call open using the CLI match; most practices already have a phone-to-client mapping in the PMS.
For practices on neither system, a CSV export synced nightly to a lookup table works — we have run this for three clients — at the cost of a one-day data lag. That is acceptable for filing deadlines, not for same-day payment references.
See our voice AI and document analysis case study for a fuller example of how real-time data retrieval integrates with live calls in a regulated context.
Handling self-assessment deadline queries: the seasonal spike pattern and how to pre-brief the agent
The January deadline creates a predictable volume shape: calls rise from around 5 December, peak in the week of 20–24 January, and drop sharply on 1 February. Query distribution shifts too — submission status and payment reference account for a higher share as clients seek confirmation that their return reached HMRC.
Pre-brief the agent for this period with two specific adjustments:
Tone shift. The default acknowledgement should reflect seasonal awareness. "We know January can feel stressful — let me check that for you right now" performs better than neutral phrasing. Clients calling in the last week before the deadline are stressed; a flat tone reads as indifferent.
Proactive data surfacing. For submission status queries in January, have the agent read back the status and the HMRC acknowledgement reference number if the return is filed. Clients are often confirming what they can tell a mortgage advisor; the reference number closes the loop so they do not need to call again.
HMRC's Self Assessment overview for agents documents the submission timeline, including the 72-hour HMRC confirmation window that clients frequently confuse with a processing delay — brief the agent on this distinction explicitly.
Identifying urgent escalations: HMRC investigation notices, debt management, and distressed client signals
A voice triage agent that fails to escalate a distressed caller is worse than no agent at all — it is a professional standards complaint and a reputational risk packaged together.
The detection strategy uses three layers:
Keyword layer: Explicit phrases — "compliance check", "code of practice 9", "COP9", "HMRC letter", "investigation notice" — trigger Tier 3 regardless of the surrounding conversation. These phrases have almost no benign context in an inbound call to an accountancy firm.
Sentiment layer: High-arousal negative sentiment in the opening ten seconds. The main LLM system prompt instructs the agent to treat expressions like "I'm panicking", "I'm really worried", or "I've got a serious problem" as escalation signals, without requiring a separate sentiment model.
Repeat contact layer: Any caller who explicitly mentions a previous callback was not received, or who references having called more than twice about the same issue, routes to Tier 3. Repeated contact is the clearest distress signal in the call log data we have seen across multiple practices.
HMRC's Compliance Handbook COM10100 outlines the investigation types clients may reference. Brief the agent on COP8 and COP9 specifically — these are the terms a client holding a formal investigation letter will use.
Compliance layer for UK accountancy firms: ICAEW guidance on automated client communications
ICAEW does not prohibit automated client communications. Their Code of Ethics and Technology and the Profession guidance confirm that automation is acceptable for routine enquiries provided confidentiality obligations are met and the client can access a human when needed.
Three compliance requirements apply for any UK accountancy firm deploying a voice triage agent:
Call recording consent. RIPA 2000 and the ICO's recording guidance require notification before collecting any client information. The agent must deliver this in its opening statement.
Data minimisation. The agent should collect only what is needed for the specific query — CLI match or UTR confirmation. Do not prompt clients to repeat account balances or tax liabilities not required for the request.
Human fallback guarantee. ICAEW's client care standard requires access to a qualified member of staff when the matter needs professional judgement. "Press 0 at any time to speak with the team" must route to a staffed line, not voicemail, during practice hours.
Transfer-to-human design: the handoff that preserves client context without asking them to repeat themselves
The worst handoff is the cold transfer: the client explains their situation to the voice agent, gets transferred, then explains it again to the staff member. Our post-call survey data shows this is the top complaint in voice triage implementations, ahead of wait time and resolution accuracy.
The solution is a call summary injected at the point of transfer — read aloud to the staff member before bridging, or sent as a screen pop via webhook.
Our production handoff template:
TRANSFER SUMMARY — [timestamp]
Client: [name from PMS, confirmed by caller]
CLI: [inbound phone number]
Query type: [intent classification from tier routing]
Data retrieved during call: [fields surfaced — e.g. "return filed 14 Jan, HMRC ref SA123456"]
Reason for escalation: [tier_3 trigger phrase / sentiment flag / repeat contact flag]
Call duration so far: [seconds]
The staff member picks up already briefed. The client does not repeat themselves. When escalation is due to a distress signal, the summary flags this so the staff member adjusts their tone before the caller arrives.
For a detailed breakdown of transfer mechanics and context preservation patterns, see our post on transfer-to-human handoff design.
What changed in 2025–2026: HMRC and ICAEW guidance on AI-assisted client communications in regulated practices
Two developments directly affect UK accountancy firms deploying voice agents.
HMRC updated its agent authorisation framework in early 2025 to clarify that automated systems acting on behalf of an authorised agent practice remain covered by the practice's existing agent authorisation — no separate registration is required for the automation layer. This removed a meaningful ambiguity that had made some practices cautious about deploying automated inbound handling.
ICAEW published updated Technology and the Profession guidance in late 2025 explicitly addressing AI-assisted client communication. The guidance recommends — though does not require — that practices maintain a log of automated interactions for professional indemnity purposes. Any production deployment should include a call log with intent classification, fields surfaced, and escalation outcome per call. That log is also the diagnostic data that improves the agent over time.
One counterpoint worth reading: ACCA's research on client trust in automated advisory services finds that transparency about automated handling correlates with higher trust scores, not lower ones. Name the system — practices that do tend to see less client resistance, not more.
Good / Bad / Ugly: three accountancy triage architectures and their client satisfaction impact
We have audited or built voice triage systems for seven UK accountancy practices since 2024. The variance between implementations is large.
The Good: IVR replacement with live PMS lookup
A 22-partner regional firm replaced its four-option IVR with a natural-language agent connected to IRIS via n8n. Average call resolution time dropped from 4.2 minutes to 1.8 minutes for Tier 1 queries. Client satisfaction scores on post-call SMS surveys moved from 3.8/5 to 4.5/5 in the first quarter. The decisive factor was live data retrieval — clients received specific answers about their return, not instructions to check their email and call back.
The Bad: Intent classification without escalation override
A 6-partner London practice deployed a voice agent with a standard FAQ-style intent system and no Tier 3 override. Two calls involving HMRC compliance checks routed to the callback queue. Both clients had been waiting three days already. One filed a professional standards complaint. The fix took forty minutes. The complaint process took four months.
For the architecture differences that prevent this, see our post on IVR replacement for UK SMEs and on inbound customer service voice agents.
The Ugly: Off-the-shelf bot with no PMS integration
A 10-person practice purchased an off-the-shelf voice bot from a telecoms reseller with no PMS integration. Callers asking for submission status received: "I don't have access to your account details — please email us." That is functionally worse than voicemail. The practice reverted to manual handling after six weeks.
Voice triage without live data retrieval is not triage — it is a conversational dead end. The PMS integration is the product.