The phone call your customer expected to be useless.

The phone call worth taking

Most “AI voice agents” demo well in the boardroom and fail the second a real customer interrupts with an accent, a background compressor, and a machine model number spoken in three syllables instead of five. The call agent worth deploying is the one that knows the corpus and can finish the work — open the ticket, draft the SMS, escalate when it must. Everything else is a demo.

How it works

Three modes, one corpus, no separate knowledge silo.

Inbound. A customer calls the dealer support line. The voice agent identifies the machine from the serial or model number the caller states, pulls the relevant service bulletin, and either answers the question or opens a work order — part SKU pre-attached — without a human picking up the headset. When the case is beyond what the agent should handle unattended, it hands off cleanly: call transferred, full transcript ready, citations visible to whoever takes it.

Outbound. The agent calls back when a part lands at the warehouse, when an SLA window is about to open, when a post-service survey is due. AHT, deflection rate, and callback rate are the numbers your procurement committee will ask for; they move in the right direction when outbound is handled by a system that doesn’t go off-shift.

In-workshop. A technician is hands-busy, eyes-busy, gloves on. Torquing a fitting, both hands on the wrench. Voice is the right interface — not because it is convenient, but because it is the only one that works. The agent takes the query, answers from the same corpus, and logs the exchange. The workshop voice piece covers why this matters for field operations specifically.

All three modes run against the same grounded corpus as the Knowledge Agent. Same citations, same access-control lists, same audit log. A technician with read-only access to a specific product line gets that access on the phone the same way they get it in the portal.

The system also handles 40+ languages and regional accents, including mid-sentence code-switching. A German-Polish bilingual technician can say “der Werkzeugträger ist locker” mid-call and the agent tracks it. This is not a feature note — it is the difference between a system that works in a real multilingual dealer network and one that routes every non-English speaker to a queue.

A call that closed in under 90 seconds

A Mitsubishi Logisnext dealer in Poland. Service desk, mid-morning. A customer calls: forklift down at a busy dock, production stopped, VIN on the plate.

The voice agent takes the call. VIN identifies the chassis variant. The agent narrows to the relevant service bulletin, walks the customer through a 30-second diagnostic check — power cycle, indicator code, connector seat — and identifies a likely failed proximity sensor. The work order opens with the part SKU pre-attached. A technician callout is scheduled. A confirmation SMS goes to the customer in Polish.

By the numbers

AHT on that tier-1 case: under 90 seconds. The senior technician on shift never picked up the headset. Across a full pilot, that compounds to 11 hours per engineer per week recovered — typical pilot result, varies with corpus quality and language mix. AHT, deflection rate, and callback rate all move when inbound volume stops requiring a human for tier-1.

What you configure

Voice persona — tone, pace, hold music. Languages enabled per line. Hours during which the agent operates unattended versus hours that route to a human. Which actions the agent may take without confirmation (open a work order, send an SMS) versus which require a human to approve (commit a parts order above a spend threshold, escalate to a named technician).

The audit log captures the full call transcript, the citations the agent cited during the call, and every action it took. Nothing happens off the record.

What it connects to

Telephony via Twilio, Vonage, or your existing gateway — inbound and outbound. Work-order creation goes to whichever FSM or CRM is in place: ServiceNow, Salesforce FSL, your bespoke ticketing system. Parts lookups run against the ERP — SAP, Dynamics, IFS, whatever is the source of truth.

The full connector list is at Integrations.

Where to go next

• Construction & Material Handling — dealer-hotline case patterns and call-volume benchmarks.
• Customer Service use case — how deflection rate is framed for service leadership.
• Voice in the workshop — the detailed argument for voice as the primary interface when screens are not an option.