thesolver
Talk to us
AI CHATBOT AUDIT · FOR ANY CUSTOMER-FACING CHATBOT

Your AI is talking to customers. Do you know what it's saying?

Chatbots hallucinate facts, get talked off-script, and say things no company ever approved — and it's already cost real businesses money, court cases, and headlines. Here's what that actually looks like.

Run independently by someone who spends their day job professionally testing and auditing commercial AI systems — this is that same discipline, turned outward.

THIS ISN'T HYPOTHETICAL

It's already happened to well-known brands

Not a theoretical risk — it's already cost real companies money, court cases, and headlines.

REAL INCIDENT · 2025
Bunnings' chatbot gave advice restricted to licensed tradespeople
Australia's most-trusted brand had its AI assistant walk a customer through electrical work legally reserved for licensed electricians in Queensland.
Repercussion: Public admission from Bunnings' own CIO; safeguards rebuilt after the fact.
Source: Inside Retail, 2025
REAL INCIDENT · 2024
Air Canada's chatbot invented a discount policy
Its chatbot hallucinated a bereavement-fare discount that never existed, and told a grieving customer he qualified for it.
Repercussion: A tribunal ruled Air Canada legally bound by what its bot said, and ordered it to pay damages.
Source: Moffatt v. Air Canada, 2024 BCCRT (Canada's Civil Resolution Tribunal)
REAL INCIDENT · 2023
A car dealership's chatbot "sold" a $76,000 SUV for $1
A customer talked a US Chevrolet dealer's chatbot into agreeing to a "legally binding" $1 sale price on a new Tahoe, with zero guardrails stopping it.
Repercussion: Screenshot went viral; the dealership pulled the chatbot offline entirely.
Source: Business Insider, 2023
REAL INCIDENT · 2024
DPD's chatbot swore at a customer and trashed its own company
A frustrated customer got the delivery firm's chatbot to call DPD "the worst delivery firm in the world," write a poem mocking it, and swear.
Repercussion: Viewed over a million times in days; DPD disabled the AI element company-wide.
Source: BBC, The Guardian, 2024
REAL INCIDENT · 2023
One wrong chatbot answer wiped $100 billion off Google's parent company
Google's Bard gave an incorrect factual answer in its own launch advertisement — a single hallucinated detail, in a promo video.
Repercussion: Alphabet shares fell as much as 9% in a day — over $100B in market value, in hours.
Source: Reuters, 2023
REAL INCIDENT · 2024
NYC's own government chatbot told businesses to break the law
New York City's official small-business chatbot told employers they could legally pocket staff tips and refuse tenants with housing vouchers — both illegal under NYC law.
Repercussion: Left live despite public criticism; a new administration shut it down in 2026.
Source: AP News, The Markup, 2024

Six industries, six companies, same pattern. Most incidents never make the news — these only did because someone happened to notice.

THE DELIVERABLE

What you actually get

  • A written report — PDF or Word, nothing to log into
  • Every finding with the exact question asked and the chatbot's exact response — no paraphrasing
  • A risk score across all eight categories
  • Specific, actionable fixes — not just "be careful"
  • Snapshot in about a week. Full audit in one to two.

See a full example report →

METHODOLOGY

Eight categories. Every finding backed by an actual transcript, not a guess.

We test how it behaves under pressure — the adversarial, off-script questions a real customer eventually asks.

Hallucination

Invented policies, fees, or figures stated with total confidence.

Unsafe advice

Legal, financial, or electrical guidance it shouldn't give unsupervised.

Compliance

Industry-specific rules — discrimination, disclosures, cooling-off rights, and more.

Brand risk

Contradicts the company's own site, or answers inconsistently.

Escalation

Recognising complaints, threats, and vulnerable customers.

Privacy

Unnecessary personal info requested, or internal info exposed.

Prompt resilience

Can an ordinary customer get it off-script?

Experience

Accuracy, clarity, and whether it actually helps.

WHO THIS IS FOR

Built for any business running a customer-facing chatbot

Built on Intercom, Zendesk, or something custom — we test the conversation, not the vendor.

Already have a chatbot live
The most common case — a vendor-built or custom assistant has been answering customers for a while and nobody's stress-tested it since launch.
Rolling one out soon
Cheaper to catch this before launch than after a customer screenshots a bad answer.
Not sure what it's built on
That's fine — we identify the chatbot as part of the engagement, not something you need to know upfront.
QUESTIONS

Before you ask

Isn't this what our chatbot vendor already tested?

Vendor testing is almost always pre-launch and scripted to the demo. We test the live system, after launch, with the same adversarial and off-script questions a real customer eventually asks — that's usually where the gap shows up.

Is this legal — are you "hacking" our chatbot?

No. Every test is a normal question typed into your public chat interface, the same access any customer has — no exploits, no unauthorised access, nothing destructive. Before any engagement starts, you sign a short scope-of-work confirming exactly what will and won't be tested.

How do we know the findings are real and not staged?

Every finding in your report includes the exact question asked and the chatbot's exact response, so you can reproduce it yourself. Nothing is summarised away.

What if you find something and we don't hire you?

The initial exposure conversation is free and non-binding. You keep whatever we've already shown you either way.

PACKAGES

Two ways in — a quick look, or the full picture.

Pricing shown is indicative and being validated with early customers.

Snapshot
$990–1,500
Limited testing across core categories with executive findings.
  • ~15 scenarios
  • Executive summary
  • Top findings with evidence
Full audit
$2.5k–5k
Comprehensive testing across all eight categories, with remediation guidance.
  • 40+ scenarios
  • Full evidence report
  • Recommended fixes
GET IN TOUCH

Want to know what your chatbot is actually saying?

No self-serve audit tool, no obligation. Leave your email — we'll reply within one business day to arrange a free, no-obligation chat about what an audit would cover.

Retail, finance, telco, services — any business running a customer-facing chatbot.
PRIVACY

Privacy policy

Short version: we collect only what you give us on this page, we don't sell it, and we don't track you around the web.

What we collect. If you submit the contact form, we collect the email address you provide. That's it — no cookies, no analytics tracking, no third-party ad pixels on this site.

Why. Solely to respond to your enquiry about an audit.

Third parties. Form submissions are processed by Formspree, our form-handling provider, solely to deliver your enquiry to us.

Retention. We keep enquiry emails only as long as needed to respond, then delete them.

Your rights. Email us any time to ask what we hold about you or to have it deleted.

Audits themselves. If you become a client, what we test, capture, and report on your chatbot is governed separately by the signed Authorisation & Scope of Work, not this policy.

Last updated: September 2026. Contact: hello@thesolver.com.au