Most SME phone lines send calls to a mobile that goes to voicemail after 4 rings, with a 47-hour average callback time and a 60% lead loss rate before the first human conversation. AI call routing fixes that: a voice agent answers, qualifies and routes every inbound call instantly. It is the highest-ROI automation available to UK SMEs that rely on phone enquiries, and the most commonly under-deployed one.
The pattern isn't complex: an AI agent answers every call within the first ring, captures intent, qualifies against your ICP in 90 seconds, and either books a meeting instantly or routes to the right human with a structured summary in their Slack. The result (average first response under 5 minutes, 24/7 coverage, zero missed leads) is achievable in 1–2 weeks of deployment for a standard B2B inbound flow.
This guide covers the routing architecture we deploy for UK SMEs: the flow logic, intent classification, CRM integration, and the overnight queue that makes the after-hours path as effective as the in-hours one. Quantum builds it as inbound lead routing software wired to your CRM and calendar; see our custom AI agents service.
Do you need AI call routing? A 30-second check
Are you missing calls outside business hours?
YES → AI inbound routing is mandatory. Every missed call is a lost lead. Continue.
NO → is your average first-response to answered calls > 1 hour?
YES → automate the routing and notification layer. Continue.
NO → human-only routing is working. Tune the process before automating.
Is the inbound call mix mostly new leads, or existing customers?
MOSTLY LEADS → focus on qualification + booking flow.
MOSTLY CUSTOMERS → focus on account lookup + support routing.
MIXED → build both flows with caller identification at connection.
The Missed Call Problem
The data on missed inbound calls is consistent across UK SME sectors: 35–55% of inbound calls to a mobile or single-line number go unanswered or to voicemail during business hours alone. After hours, the figure is 90%+. Of the calls that hit voicemail, fewer than 10% leave a message, and of those messages, roughly 40% are ever returned.
The commercial impact: for a UK field services firm getting 30 inbound enquiries per week, a 40% miss rate means 12 leads per week that never enter the pipeline. At a 30% close rate and £2,000 average job value, that's roughly £24,000 per month in recoverable revenue from a problem that takes two weeks and under £1,000/month to fix.
The reason this persists is that the obvious fix (hire a receptionist) is expensive and leaves the after-hours problem untouched. AI missed call recovery at the point of the call handles both, and costs less than a part-time receptionist at most SME inbound volumes.
How AI call routing works
Every deployment follows the same four steps:
- Telephony → Agent runtime. The call connects to a Twilio number that forwards to the agent runtime (Retell AI, Vapi, or custom broker). The agent answers within the first ring. Barge-in is enabled, so callers can interrupt the greeting immediately. See the Voice AI Architecture guide for the full stack configuration.
- Intent and identity capture. The agent captures: caller's name, the nature of their enquiry (new service, pricing, support, existing customer), their preferred contact details if different from the calling number, and any time-sensitivity signal. This takes 60–90 seconds for a straightforward call.
- Routing decision. Based on intent, the agent routes to: (a) immediate self-serve booking if the intent is a known service type, (b) SDR callback queue if the deal value signal warrants human follow-up, or (c) knowledge answer if the caller has a factual question the agent can answer directly.
- CRM creation + notification. The structured capture is written to the CRM and a Slack alert is fired to the relevant team or rep. The notification includes a one-click "claim & call back" link. If a booking was made in step 3, the meeting is created in the calendar and a confirmation SMS/email is sent to the caller.
Intent Capture and Classification
Intent classification (determining why the caller is calling) is the decision that determines which path they take. The agent should determine this in the first 30–60 seconds without making the caller feel interrogated. AI lead qualification happens in the same pass, so the caller never answers a second round of questions.
The classification categories that cover 90%+ of B2B service inbound calls:
- New enquiry (demo / quote / discovery): route to qualification + booking flow. Capture: company size signal, urgency, who else is involved in the decision.
- Pricing question: give the ballpark if you have a published range; if bespoke, route to a "scoping call" booking. Never give a single number on an agent-handled call; it anchors without context.
- Support / existing customer: phone number lookup against CRM to identify the caller. Route to customer service flow with account context pre-loaded.
- Referral / partnership: capture the referring party and nature of interest; route to a specific rep or booking type if relevant, otherwise to a general "partnership" queue.
- Unclear / unusual: short hold ("one moment while I connect you") and route to the general human queue with the partial transcript. Don't attempt to force-classify ambiguous calls.
The agent's system prompt should include clear routing instructions for each category and explicit instructions about what NOT to capture (sensitive financial details, health information) to stay within GDPR data minimisation requirements.
Routing Logic
Routing rules determine which human gets the notification and what action they're expected to take. Simple rules outperform complex ones because they're easier to maintain and debug when something goes wrong.
The routing tiers most UK SMEs use:
- Tier A: self-serve booking. Caller fits a known service type, has a clear intent, and the booking type has availability. Agent books directly, no human required. SMS + calendar confirmation. Write to CRM with "booked" status and notification to assigned rep.
- Tier B: priority callback queue. High-intent call with signals of larger deal value (larger team size, multi-site, tight timeline). Human must call back within 5 minutes. Slack alert to SDR channel with priority flag and "claim" button.
- Tier C: standard callback queue. Genuine interest, standard deal size, no urgency signal. Human calls back within 2 hours during business hours. CRM task created with due time.
- Tier D: knowledge answer + follow-up. Caller had a specific factual question the agent answered. No immediate call-back required. CRM note created with the question and answer. Rep reviews at next opportunity.
CRM Integration
The CRM write is the record of the call: it must be created within seconds of the call ending or (for bookings) immediately during the call. Delayed or batch CRM updates defeat the purpose of the routing system.
POST /api/ingest/lead
{
"name": "Alex Rivera",
"email": "[email protected]",
"phone": "+447700900123",
"intent": "new_enquiry_demo",
"tier": "B",
"source": "inbound_call",
"call_sid": "CA...",
"transcript_id": "tr_123",
"timezone": "Europe/London",
"notes": "Evaluating AI appointment booking for 15-person sales team. Current stack is Salesforce. Timeline: next quarter.",
"routing_owner": "team_sdr"
}The notes field is the most valuable part: a structured summary of what the caller said, not a raw transcript excerpt. This is what the human rep reads first, and it determines whether they call back with context or cold. Train the agent's system prompt to generate a 1–3 sentence summary in the notes field: "Alex is evaluating AI appointment booking for a 15-person sales team currently using Salesforce. Timeline is next quarter. No budget discussion; he wants a demo first."
De-duplication is the most common CRM integration bug: the agent creates a new contact for a number that already exists in the CRM. Build a phone-number lookup before creating the contact: upsert if found, insert if not. A CRM with 3 duplicate records for the same prospect (one from the form, one from the first call, one from the re-call a week later) is a lead management failure waiting to happen.
Booking via Calendar
Direct booking during the call is the highest-value outcome: the meeting is in the diary, the calendar invite is sent, and no human action is required for the initial scheduling step.
// Agent proposes slots during the call
const slots = await cal.getAvailability({
owner_id: routedOwnerId,
from: startOfNextBusinessDay(),
days: 7,
timezone: callerTimezone
});
const topTwo = slots
.filter(isWithinWorkingHours)
.slice(0, 2)
.map(formatSlotForSpeech);
// Agent says: "I have Tuesday at 10:30 or Wednesday at 2pm —
// which works better for you?"
// On confirmation:
await cal.createBooking({
eventType: '30min',
attendee: { name, email },
start: selectedSlot.start,
end: selectedSlot.end,
metadata: { source: 'voice_inbound', call_sid, transcript_id }
});
await crm.createMeeting({
contact_id: contactId,
start: selectedSlot.start,
end: selectedSlot.end,
subject: `Voice inbound: ${intent}`,
notes: callSummary
});Hold the proposed slot for the duration of the booking interaction (typically 60–90 seconds) before releasing it back to the availability pool. Cal.com's hold API handles this; build hold-and-release logic if using a custom booking system. A caller who says "Tuesday 10:30" and then gets told "sorry, that slot is no longer available" is a brand damage moment that erodes trust in the agent.
Overnight Queue and Morning Handoff
The most consistently skipped component in inbound routing deployments is the overnight queue, the mechanism that gets after-hours leads to the right human at the start of the next business day with enough context to act immediately.
The failure mode: the AI agent handles an after-hours call perfectly, writes to the CRM, and sends an email notification to a shared inbox. Nobody reads the shared inbox until 10am. By then, the prospect has received three other callbacks from competitors. The lead was handled. It just never got prioritised.
The overnight queue that works:
- A dedicated Slack channel (#inbound-overnight or similar) that receives one message per after-hours lead
- Each message contains: caller name, intent, tier classification, call summary (1–3 sentences), deal value signal, and a "Call back now" button that opens the dialler with the prospect's number pre-loaded
- The channel is sorted by tier (Tier B leads first) and the first rep into the office claims and acts on Tier B leads before anything else
- An escalation alert at 9:15am if any Tier B lead hasn't been claimed in 15 minutes
This pairs directly with the speed-to-lead framework: the agent's after-hours capture is the automation; the morning queue is the human's 5-minute window. Both layers are required for the system to work end-to-end.
Tracking and Metrics
The four metrics that tell you whether the inbound routing system is working:
| Metric | What it measures | Target |
|---|---|---|
| Answer rate | % of inbound calls answered by agent | 100% (no missed calls) |
| Booking rate on answered calls | % of qualifying calls that result in a booked meeting | Depends on offer; typically 20–35% |
| Median first human contact (business hours) | Time from call end to first human callback (Tier B/C) | < 5 minutes |
| Overnight queue claim rate | % of after-hours Tier B leads contacted before 10am | > 90% |
Track by call time-of-day: an 80% overall booking rate that drops to 30% on calls after 5pm tells you the overnight queue isn't working, not that the agent is underperforming. The channel-mismatch rate (caller used phone, notification went to email) is the silent killer of overnight performance.
What works, what breaks, what hurts
Good. Phone-number CRM lookup at call connection to identify existing customers. Instant CRM record creation with structured notes. Tier-based routing with different SLA expectations per tier. Chat-native overnight queue with one-click callback. De-duplication check before contact creation. Booking confirmation by SMS (not just email) for mobile-first callers.
Bad. No lead de-duplication: three CRM contacts for one person. Timezone mismatch in slot proposals, such as booking a 9am meeting in the system but confirming "9am" to a caller in Scotland when the rep is in London is the same timezone. A routing rule that assigns all Tier B leads to the same rep, who is on annual leave. A Slack notification that goes to a channel no one monitors outside business hours.
Ugly. Voicemail loops: the inbound call hits voicemail, the voicemail system records a confused AI greeting, the agent's platform interprets the callback as a new inbound call and repeats the cycle. Implement call SID cooldown and state tracking. An overnight queue that fills up over a bank holiday weekend because nobody set an escalation rule. A "claim this lead" button in Slack that opens the HubSpot contact record (not the dialler), so the rep has to manually find the number and dial. Measure the time from Slack notification to first dial; fix the UX before optimising the script.