Reliability

AI can be wrong. Here is exactly what we do about it.

Every AI product ships with a line of small print saying it can make mistakes. That line is true and it is also not good enough on its own — a caller who is told the wrong balance acts on it. This page is the longer answer: the specific failures that matter on a phone call, and the specific controls that catch each one.

The failure that actually matters

Not rudeness, and not a clumsy sentence. A confident, specific, wrong fact.

Why "hallucination" is the wrong word for the risk

The word suggests something obviously strange. The real failure is the opposite: a sentence that sounds exactly like every other sentence on the call, delivered with the same steady confidence, and wrong. "Your balance is three hundred and forty pounds." "That was issued on the fourteenth." Nothing about how it sounds tells the caller anything is amiss — which is precisely why it is dangerous, and why the control has to sit in the system rather than in the caller's judgement.

Why a prompt is not a control

"We told the AI not to do that" is a hope, not a safeguard. Instructions in a prompt are weighted, not enforced: they compete with everything else in the context, and a long or unusual call is exactly when they lose. Anything that genuinely must not happen has to be impossible in code rather than discouraged in language. That principle runs through everything below.

Four layers, each catching what the last one can't

They are deliberately different in kind. Stacking four versions of the same idea gives you one control with extra steps.

Layer 1

It cannot do the worst things at all

There is no function anywhere in the system that can change, waive or zero a balance. Not a restricted one — no such function exists, so there is nothing to be tricked into calling. Payment arrangements are calculated in code from your own policy: the minimum instalment, the longest term, how soon the first payment must fall. The agent can only read out what that calculation produced.

Account information is gated the same way. Nothing about an account reaches the model until verification has passed in code, and a caller who has not verified cannot be told the account exists, let alone whose it is.

Layer 2

Every figure is checked before it is spoken

Everything specific the agent is entitled to say came from somewhere: your systems, the caller, or your configuration. So before a sentence is turned into speech, every amount, date and reference in it is checked against that record. A figure that cannot be traced to any of them is not spoken — the agent offers to have it confirmed instead.

This runs in microseconds, in code, before the audio exists. It is the only point at which a wrong number can still be stopped, and it is the reason this layer is deterministic rather than another model's opinion: asked whether £340 is a number anyone mentioned, a model guesses. A lookup knows.

Layer 3

A second, stronger model watches the call

The model holding the conversation is chosen for speed, because a second of dead air on a phone call reads as a broken line. A larger model reads each exchange while the reply is being spoken, checking it against your rules. It cannot unsay a sentence — nothing can — but it can put a correction in front of the next one, which is what a person does when they realise they have misspoken, and it can pull a human onto the call.

It costs no latency, because it never sits in front of the caller. It is also deliberately narrow: it reports fabrications, broken rules, premature disclosure, commitments you never authorised, and missed distress signals. It is explicitly not asked for opinions on phrasing.

Layer 4

Every call is reviewed after it ends

Once the call is over there is nothing left at stake in it, so the review can be thorough. The whole transcript goes back through the supervising model against the full set of rules for your business, and anything it finds lands on the call record where you and we can both see it.

That record is the point. It turns "AI can make mistakes" from a disclaimer into something with evidence behind it: what was caught, whether the caller ever heard it, and what happened next.

What each layer actually catches

The honest version, including where a layer stops being useful.

An invented balance, fee or payment amount

Caught by layer 2, before the caller hears it, because the figure appears nowhere in what your systems returned. Layer 1 already makes it impossible for the agent to act on such a figure — it could never apply it to an account — so the exposure was always the caller acting on what they were told, and that is the exposure layer 2 closes.

A date that never appeared on the record

Caught by layer 2. Dates are checked in both directions — a record holding 2026-07-14 and an agent saying "the fourteenth of July" are the same fact in two alphabets, and the check knows that.

A payment plan the business never agreed to offer

Prevented by layer 1. The terms are computed from your policy before the model sees them, and the model has no way to produce different ones. It relays; it does not decide.

Advice it was told never to give

This is layer 3's territory. Whether a sentence crosses from "taking details" into "advising" is a judgement, and a judgement is what a second model is genuinely better at than a rule. The guardrails are agreed with you in writing during onboarding and are the standard it is checked against.

Missing that a caller is struggling

The primary control is not a guard at all — vulnerability signals escalate a call to a person before verification has even finished, so nobody in difficulty is made to pass an identity check first. Layer 3 is the backstop: if the signal was subtle and the agent carried on taking details, the supervisor flags it and a person is offered.

Simply mishearing someone on a bad line

Not a hallucination, and not something any of these layers fix — it is an audio problem, and it is handled as one. Long codes get a free re-read; spoken numbers and the police alphabet are folded to their written form; a reference that keeps getting mangled can be keyed in on the phone's keypad instead; a couple of turns the recogniser struggles with prompt the agent to ask the caller to come off speakerphone; and if the line itself is failing, the agent says so rather than leaving silence or blaming the caller. A person on a bad line does exactly the same things.

A sentence that is technically true and badly judged

Layer 3 sees it; whether it flags it depends on whether it crossed one of your rules. This is the honest limit of the system. Tone, timing and tact are not solved problems, and we would rather say so than claim a control we do not have. It is one of the reasons we read every transcript from a new line's first week.

The data we refuse to keep

Reliability is also about what never lands in a database. Callers tell a phone line things they would never type into a form — the discipline is holding on to as little of it as possible.

Care flags, not case histories

When a caller's circumstances mean careful handling — a bereavement, an illness — the record your team sees is a fixed category, applied in code: handle with care — bereavement. The caller's own words about their health or their loss are never what gets written to the account. Your team gets everything they need to act well, and nothing they don't.

Transcripts swept after every call

The post-call review does one more job: any transcript line containing special category data — health, disability, religion and the rest of the UK GDPR's protected list — is withheld from storage and replaced with a marker saying so. What someone volunteered in a difficult moment doesn't sit in a database waiting to become a breach.

Minimised, then deleted on schedule

Personal identifiers are redacted before transcripts are stored at all, and everything is purged automatically on a fixed retention window — a nightly job, not a policy that depends on someone remembering. A call that becomes evidence in a complaint can be placed on hold, so it survives the purge for exactly as long as it's needed. Every action on an account stays audit-logged, so less data never means less accountability.

If your compliance team wants this in writing rather than on a web page, it is: the special-category handling above is a commitment in our Data Processing Agreement, alongside every sub-processor, transfer basis and retention rule — including an honest statement of where automated recognition has limits.

What we are not claiming

The parts a vendor usually leaves out.

Not that it never gets anything wrong

It will. The claim is narrower and more useful: the failures that could cost a caller money or harm them are caught by something that is not the model's own judgement, and the rest is recorded so it can be found and fixed.

Not that it replaces your team

It answers the front of the call, handles what follows a pattern, and hands over what doesn't — with the details already taken. Anything needing professional judgement in the moment goes to a person, by design and not as a fallback.

Not that you can stop looking

Call records and findings are made available to you, and reviewing them is part of running the service properly. We read every call from your first week and report back; after that the volume drops but the habit shouldn't.

All of this is contractual, not aspirational. The AI clause in our Terms of Business sets out what the service can and cannot be relied on for, what we commit to operating, and where responsibility sits between us. If a supplier will not put that in writing, that tells you something.

Test it yourself.

Ring the demo line and try to make it say something it shouldn't. Ask it for a figure it has no way of knowing. That is a more useful five minutes than anything on this page.