What AI voice agents get wrong, and how to design around it

AI phone agents got good quickly. A system that sounded obviously robotic three years ago now handles a booking request in natural speech, interrupts and recovers gracefully, and closes the call without the caller thinking much about it.

This has produced two opposite bad decisions. Some businesses route everything to an AI agent and discover their customers hate them. Others hear one bad demo and refuse to touch the technology at all, then keep paying people to read the same address off a screen four hundred times a month.

The useful position is in between, and it depends on understanding the specific ways these systems fail. Not the general “AI isn’t perfect” hedge — the actual failure modes, which are consistent enough to design around.

What they genuinely do well

Credit where it’s due. Current voice agents handle these reliably:

Structured information capture. Name, phone number, address, order details, appointment slot. Consistently, without transcription drift, at 3am.
Answering known questions. Hours, location, pricing, policy, “do you deliver to Winkler.” If the answer exists in a document, the agent can find and speak it.
Triage. Determining what the call is about and routing it correctly is something AI does better than a phone tree, because the caller can just say what they want instead of navigating menus designed in 2011.
Absorbing volume spikes. Forty simultaneous calls during a lunch rush is a staffing crisis for humans and a non-event for software.
Never being in a bad mood. The agent is exactly as patient on call 200 as on call 1.

For a restaurant taking orders, a clinic booking appointments, or a trades business capturing after-hours callouts, that covers a large share of total call volume.

Where they fail

1. Emotional state

This is the big one, and it’s not really solvable with better models.

A caller who is angry, frightened, grieving, or in crisis needs a person. Not because the AI can’t produce sympathetic-sounding words — it can, fluently — but because the caller can tell, and because being handled by software in a bad moment is itself an insult. The customer whose order was wrong for the third time does not want an efficient resolution from a machine. They want a human being to acknowledge that this was bad.

Getting this right requires detecting distress early and escalating fast. Getting it wrong produces the single worst customer experience in the whole system, and it’s the one that ends up on social media.

2. The unusual case

Voice agents are strong on the distribution’s centre and weak at its edges. The call that combines three issues, references a conversation from last month, involves an exception someone made verbally, or requires judgement about whether to bend a policy — these go badly. The agent will confidently do something reasonable-sounding and wrong.

The tell is that these calls are rare individually but common collectively. Any given edge case is 0.2% of volume. All the edge cases together might be 15%.

3. Accents, noise, and speech variation

Recognition has improved enormously and is still uneven. Callers with strong regional or non-native accents, callers with speech differences, elderly callers, callers on a bad mobile connection, callers from a construction site or a moving vehicle — error rates rise for all of them.

This matters beyond the technical failure. If your AI agent handles some customers noticeably worse than others, you’ve built a system that provides worse service to specific groups of people. That’s worth thinking about carefully before deployment, and worth measuring after.

4. Confident invention

If the information isn’t available, a poorly configured agent will produce a plausible answer anyway. It will quote a price that isn’t your price, promise a delivery window you don’t offer, or state a policy you don’t have.

A human agent who doesn’t know something usually says so, or puts the caller on hold and asks. This is a behaviour that has to be deliberately engineered into an AI system, and it’s engineered by constraining what the agent is permitted to assert — not by asking it nicely to be accurate.

5. Memory across calls

Most deployments treat each call as isolated. The caller who phoned yesterday about the same problem starts from zero. For a first-contact booking system that’s fine. For anything resembling ongoing support it’s actively worse than a person with a CRM open, and it produces the specific frustration of explaining your situation for the third time.

6. Silent failure

A human agent who’s out of their depth escalates, or at least sounds uncertain. An AI agent hits its limit and keeps going in the same confident tone. Nobody in the business finds out unless someone is reviewing transcripts — which is why the deployments that go wrong tend to go wrong quietly, for months.

Designing around it

The failures above are consistent, which means they can be planned for.

Give the agent a narrow, well-defined job. “Take delivery orders and answer menu questions” works. “Handle customer service” doesn’t. Scope is the single biggest predictor of whether a deployment succeeds.

Make escalation the default under uncertainty. The agent should hand off when it hits anything ambiguous, not attempt a rescue. Over-escalating costs a few minutes of human time. Under-escalating costs customers.

Escalate with full context. The handoff has to carry the transcript and the captured details to the human. If the caller has to repeat everything, you’ve built the worst of both systems — machine latency plus human cost.

Escalate on emotion immediately. Frustration, distress, and repeated attempts to reach a person are all triggers. And “let me speak to someone” must work the first time, every time. Trapping callers is the fastest way to turn an efficiency gain into a reputation problem.

Constrain what it can assert. The agent answers from a maintained knowledge base and says “let me get someone who can confirm that” for everything else. Test this adversarially before launch — ask it things it shouldn’t know and see whether it invents.

Tell people it’s AI. Callers work it out anyway, and discovering it after the fact feels like being tricked. A brief disclosure at the top costs nothing and removes the whole issue. In some contexts disclosure is also a regulatory expectation.

Review transcripts weekly. This is the discipline that separates deployments that improve from deployments that quietly rot. Sample real calls, look for the ones that went sideways, adjust. It takes an hour and it’s the highest-value hour in the whole operation.

The number that matters

The metric worth tracking is not containment rate — the percentage of calls the AI handles without a human. Optimising for containment produces systems that trap frustrated callers.

Track resolution rate: the percentage of calls where the caller got what they needed, whether or not a person was involved. Then track escalation separately, and treat a high escalation rate on a specific call type as information about scope rather than as a failure.

A well-scoped deployment on routine inbound volume commonly handles a majority of contacts end to end. The rest reaching a person quickly, with context, is the system working — not the system failing.

Why this arrangement is worth building

The economics are the reason any of this matters. Routine calls handled by software cost a fraction of routine calls handled by people. That saving is what makes it possible to staff the hard calls properly — with people who are paid reasonably, aren’t burnt out from reading the same address four hundred times, and are actually good at the conversations that need a human.

Used well, this makes support better rather than cheaper-and-worse. Used badly, it’s a phone tree with a nicer voice, and everyone can tell.

DNOTCH builds AI voice agents and runs inbound contact centre services with people handling every escalation. If you’re weighing up whether your call volume suits this, book a call — we’ll tell you if it doesn’t.