01What happens.
A chatbot answers from what it can read. Where the published estate holds old, duplicated and conflicting material, the chatbot serves it with confidence. Replacing the chatbot changes the interface and leaves the answers where they were.
02Why it matters.
If you sponsor automated citizen services, your chatbot’s answers are the authority’s answers, and the programme is judged on them. When the answers are wrong, the vendor and the digital team are the first suspects, and both can be replaced without touching the cause.
The estate the chatbot reads is its operating manual, and it is the part of the system that carries the authority’s history. In this case it was the one part that none of the 3 deployments tested.
03The case.
A local authority introduced a chatbot to handle citizen enquiries across its services. The chatbot drew its answers from the authority’s published estate, which included a planning microsite from a predecessor authority that the digital team had no record of. It treated that site’s structured documents as current guidance and served them to residents.
The answers were wrong and the complaints arrived. The authority concluded that the chatbot was at fault, held the software vendor responsible, and replaced the product with a second chatbot.
The second chatbot read the same estate and returned the same answers, and residents who had received wrong answers from the first received them again from the second. This time the content was blamed, and with it the digital team held responsible for the published estate. The team had been warning about the sprawl of the estate for years and had asked for it to be mapped, and the warning had been treated as a housekeeping matter.
A third chatbot was procured, and it read the same estate and failed in the same way. Each replacement had cost a procurement, an implementation and a period of parallel running, and each had been justified by the failure of the one before it. Only then was the cause examined where it sat, in what the chatbots were reading, and the digital team’s warning was recognised as the finding it had been. The first step in correcting the answers was understanding the authority’s actual digital footprint, and that step had been available before the first deployment.
Nobody had tested what the chatbot would read, because the estate looked like a website to the people who ran it and like a single source of truth to the system. Between those 2 views sat years of accumulated material that nobody owned as a whole and nobody had a list of.
04Exposure.
The chatbot’s answers are the authority’s answers, and the vendor’s contract does not move that. Where 3 products in succession returned the same wrong answer, the authority published the fault 3 times. The principle has been tested: in Canada a tribunal held an airline responsible for wrong information its website chatbot provided to a customer, whether the information came from a static page or from the chatbot (British Columbia Civil Resolution Tribunal, Feb. 2024), and a public body stands in the same position with its residents.
The cost was 3 procurements, 3 implementations, 2 investigations that ended at the wrong door, and residents’ complaints throughout. MIT research found 95% of enterprise generative AI pilots deliver no measurable return (MIT NANDA, 2025). AAAnow observed a cause that study did not examine, the material the systems were built on, and the 2 findings sit side by side. A chatbot written off as a failed AI project, when the estate it read was the failure, is that pattern in a local authority.
05What AI Readiness changes.
What the authority could not see from the inside was what its chatbot would find. AAAnow reads the digital landscape as AI reads it and evidences the forgotten properties, the superseded documents and the conflicting versions a chatbot will draw on, before the first deployment rather than after the third. With that evidence the authority can decide what is retired, what is corrected and whether the system is ready to go live.
Your team makes the changes; AAAnow provides the assessment, the evidence, the guidance and the review.
06Related use cases.
Planning decisions went wrong on a website the city had forgotten
The same inherited-estate failure in a US municipality, and what it cost residents.
New York City’s business chatbot told companies to break the law
The same failure on the public record, with the chatbot withdrawn.
The organisation’s own AI compounds its defective documents
The same material feeding internal systems, where the origin of the error drops out of view.
07References.
- British Columbia Civil Resolution Tribunal, Moffatt v Air Canada, Feb. 2024, reported by CBC. www.cbc.ca/amp/1.7116416cbc.caWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- MIT NANDA, The GenAI Divide: State of AI in Business, 2025. Used with attribution; the AAAnow observation sits beside it and is not merged with it.
Next step
Talk it through, in your terms.
The briefing sets this case in your organisation’s terms, and the discussion follows your questions.
Observed, anonymised. No organisation is identified.