The use cases: what happened, and the impact

In each of these something went wrong for someone: planning appeals, funding stopped, a regulator’s inspection, clinicians receiving the wrong dosing instructions, £350,000 spent on AI visibility for a site AI could not read. Nobody saw it coming. The pages and documents looked right to the people who owned them; a machine read them differently.

The cause is the organisation’s own published record: superseded, duplicated, forgotten, or laid out in a way a person can read and a machine cannot. That is also why it can be put right. One case shows the other outcome: a university whose digital team had warned about the sprawl for years mapped the estate, moved to a single platform under one point of control, and is now counting the saving.

The use cases: what happened, and the impact
Case	What happened	The impact	Why it matters
The annual report AI read was not the one the board signed	The report AI reads is not the report the board approved. Values come loose from headings, characters come through wrong, and the back of the file goes unread.	Investors, analysts and advisers act on figures the company never published.	The primary disclosure document becomes a market-facing statement the company did not make. Correction has to reverse an established position.
Brand information misunderstood by AI	Several versions of brand material sit online with nothing marking how the brand evolved. The version carrying most authority is treated as the master, not the latest. Third parties with more authority describe the brand in their own terms. Marks the brand no longer owns are still quoted as its own.	The brand AI describes is out of date and partly someone else’s.	Brand spend defeated by the organisation’s own older material. AI governance: knowing what the organisation owns and is permitted to use.
Planning decisions went wrong on a website the city had forgotten	Authorities merged into one and the old digital landscape came with them. A well-built, well-cited planning microsite from one original authority, forgotten, outranked the new portal, whose documents carried weak version, date and authority signals. The new authority’s own chatbots and agents read the old site too. In 2 instances blocking made it worse.	Incorrect advice, planning rejections that should not have happened, appeals, some litigation, poor citizen experience.	Multiple versions across the landscape, no map of the landscape, and weakly structured new information.
The chatbot was replaced twice before anyone looked at what it was reading	Successive chatbots failed, the software vendor was blamed for the first, the digital team for the wrong answers that followed. The cause was the landscape the chatbot read.	Internal disruption, the cost of 3 deployments, and time spent chasing the wrong cause, with wrong answers reaching citizens throughout.	A widespread problem. The chatbot answers from the estate as it finds it, and the organisation answers for what it says (Air Canada, cited briefly).
Chemicals: handling and storage information served wrong	Safety manuals and procedures as PDFs, a proportion from a third party, deep in old systems. Several versions with no version handling. One PDF fed both the intranet and the web. Storage temperatures out of date and step order scrambled were served by AI.	A dangerous position: handling, storage and mixing information wrong at the point of use.	The documents looked fine on screen and nobody checked how a machine reads them. 2 instances, one in Europe, one in the US.
A hostile version of the results was the one AI quoted, because the company's was locked away	The company blocks AI. Its results documents carry no titles or authors. A hostile party publishes its own version of the results, structured for AI to cite, and AI cites it.	The market reads someone else’s version of the company’s results.	The company’s own blocking created the vacuum. The sharpest market-integrity angle in the library.
Patients were told a withdrawn medicine was available, because the website still said so	Sites still carrying products withdrawn, no longer legal in the country, keep them evident and citable.	Patients and clinicians told a withdrawn product is available.	Patient safety and licensing consequence at once.
Retailer reach defeats the manufacturer’s safety information	A major US retailer’s out-of-date support and safety details, including which versions apply in which country, treated as correct because of the retailer’s reach. The manufacturer’s own details could not be cited as easily.	Customers receiving wrong safety information about the manufacturer’s product.	The manufacturer does not control the answer AI returns about its own product.
A 2-year-old partner promotion put the firm in front of the regulator	A joint promotion on a partner site 2 years earlier, for products no longer applicable to that market. No central record of the shared positioning. AI served the partner page as current product information.	The inspection cited misrepresentation of products to a market the firm could no longer serve.	Shared positioning lives on properties the firm does not control and has no list of.
Customers abroad received the wrong cover and the wrong number to call	Cover levels on the page conflicted with the contract. The emergency contact had changed to a global 0800 number not on the page AI served.	Customers receiving wrong cover and the wrong number at the point of need.	The readable page beat the filed policy.
Regulator embarrassment at poor document control	20% of links inside the regulator’s legal and regulatory PDFs broken, including the core legal basis of some documents. AI reads them on behalf of the organisations the regulator oversees.	A working interpretation of the law in circulation that the regulator cannot stand behind.	A body holding others to a published standard seen unable to keep its own record in order. Embarrassment at best.
A withdrawn medicine still carried full dosing guidance online, and AI served it as current	A site built at speed under pandemic conditions, the usual compliance oversight not applied, a reasonable trade-off at the time, then left online. Superseded documents sit beside current ones. One withdrawn product still carries full administration guidance.	AI returns withdrawn guidance as current clinical instruction.	The site outlived its purpose and its owners. Nothing retired it.
The results table AI read had its figures in the wrong columns	Tables reach AI as a stream of data. Which value belongs to which heading, column and period is lost, and worse where a table runs over several pages.	Financial figures ingested wrong and treated as right.	Several faults sit behind one word. Structure recovery agent.
The impact of AI finding what you cannot see	Fonts not embedded, specialist fonts in particular. Characters come through wrong and do not line up. Headings, titles and table structure are invisible to AI.	Data disordered on reading, and again when moved into other systems.	One specific fault, invisible to a person.
AI read half the annual report and made up the rest	A large file read only in part, worse on a slow or throttled connection. AI fills the gap it did not read.	Misinformation built from half the report.	How AI works, the basis of hallucination. With half the information, the fault does not sit with AI.
2 AI tools, 2 different answers, and consequences the organisation could not see in full	Search and Gemini present materially different pictures of the same organisation. Access, source selection and processing can contribute; no single cause is asserted without evidence.	Applicants told different courses and dates by different tools. Wrong mileage repayment details from government information.	Lost confidence, and the human cost of dealing with the challenges that follow.
The organisation’s own AI compounds its defective documents	Poorly structured documents ingested automatically into internal AI and agents. Errors compound as information is retrieved, summarised and consolidated. By the time an answer reaches a decision its origin has disappeared.	Decisions taken on answers nobody can trace.	Digital Estate DEBT™ entering the organisation’s own systems. The 9 in 10 construction: MIT measured the failure rate, AAAnow observed a cause the study did not examine.
Forgotten sites nobody was patching: AI finding and highlighting vulnerabilities	Forgotten subdomains, microsites, platforms, forms and scripts sit outside patching, access review, certificate renewal and monitoring. Unmaintained software is unpatched software.	An attack surface nobody owns. Takeovers documented across government and Fortune 500 sites.	Organisations that know their landscape and remove what no longer earns its place save money and time and run more efficiently. Coverage and cyber exposure move with it.
At renewal, the firm could not show what it had online, and the terms showed it	Exposure becomes financially visible at underwriting and renewal. Organisations that can show levels of control and reasonable steps taken are better placed on risk quality, cost, limits and terms. No outcome is automatic.	Terms, exclusions and cover worsen where control cannot be evidenced.	Not knowing what is online is a risk in itself, with ever greater impact.
Duplicates and untracked copies: the correction that never lands	19% of PDFs on organisational websites are duplicates. 37% become untracked once they leave the producer’s site. Correcting one copy leaves the others in place. Some are found, some are not, and the chase continues.	Old figures stay retrievable and the management burden never ends.	Governance in how documents are controlled, and solid foundations. The AI Accelerator’s internal education exists so governance across document control is understood.
Careers material shapes what candidates are told	Old careers pages, past roles and benefits no longer offered are assembled into the employer AI describes.	Candidates told the wrong thing about the employer, and acting on it.	73% of talent acquisition leaders are concerned candidates receive outdated or inaccurate AI information; 33% are confident their employer brand is accurately represented.
The buyer's AI knew things about the company that the company did not	The acquirer’s tools read the target’s conflicting public record and unknown assets during diligence.	Valuation and warranty consequence from a record the target did not know it had.	Industry commentary puts thorough cyber diligence at 10% of acquirers, while AI-era diligence frameworks now reach digital assets.
The documents that sank the internal AI project	A generic internal AI system was fed the conflicting safety documents from 5. Answers disagreed because the sources disagreed. The project failed, the spend was written off and filed as proof that AI projects fail. Nobody traced it to the documents.	Spend written off, and the wrong conclusion drawn.	MIT measured the rate. AAAnow observed a cause the study did not examine. Side by side, never merged.
Funding stopped on submissions that looked fine	Applicants submitted PDFs with no structure defined. 64% carried faults in how tables and data were laid out. The system misread them and funding for key projects stopped. The system was blamed, restarted on the same documents, and failed again.	Projects unfunded, applicants lost out, hours of review, a second failure.	The anchor case. Faults invisible to people decided outcomes by machine.
£350,000 to be seen by AI routes the estate was turning away	£350,000 committed to AI visibility while AI routes were turned away or left without a reply at the digital doorway. The website looked fully operational to people throughout.	Spend amplifying foundations that could not be read.	Evidence base: 5 anonymised doorway assessments, 4 to 14 of 15 AI routes turned away, people served.
The external repository that became a dumping ground	A contracted third-party document store holding annual reports and safety manuals. Drafts uploaded beside finals, versions multiplying, no management over time. AI selects from the lot.	Draft and superseded material served as the record.	The landscape extends beyond the organisation’s own sites.
Investors acted on results the company had already superseded	The published record answers investor questions through AI. Superseded results, legacy pages and guidance read with equal confidence, each repetition reinforcing the wrong position.	Investors and advisers on wrong numbers. Correction has to reverse an established position.	Confidence in management, regulatory attention, correction cost on investor relations.
Regulator's questioned governance where the compliance register stopped	Held to what is published, and AI restates it without supervision. The gap between the compliance team’s list and the real estate is where inquiries begin.	Demonstrating control after the fact is slow and expensive because the record was never assembled.	The widest case. Regulatory attention from the other cases routes here.
Clinicians received instructions the company had replaced	Administration, dosing and instructions for use conflict across versions and AI answers from whichever it retrieves. Control covers the core site while exposure lives at the edges: event microsites, old programme pages, partner routes, regional variants.	Patients and clinicians receiving wrong instructions.	The sharpest form of the risk in a regulated sector, and a major D&O problem.
Applicants acted on entry requirements the university had already changed	Course pages, fees, entry requirements and prior-cycle prospectuses stay reachable. Different tools return different answers to applicants. Applicants decide on them.	Complaints and consumer-protection questions, cycle after cycle.	73% of student AI users received incorrect information, including wrong grade requirements.
Intellectual property disputes	Claims, names and assets across regional sites, campaigns, partner material, historic records and third-party repositories. AI blends them, attributes claims to the wrong product, revives abandoned campaigns, repeats third-party wording as the brand’s own, and makes things up.	In a dispute, material published under the organisation’s name reads as admission. Contested wording spreads faster than it can be withdrawn.	Developed in depth. An accurate position of the landscape first, the AI Accelerator establishing it, then monitored with the technology overseeing at scale. Nothing worse than AI inventing under your name.
Announcements buried in the page	Big, heavy pages carrying little information. The announcement is lost in the page furniture, so the systems that should carry it widely miss it and the superseded figure keeps being repeated. Companies that bought AI visibility to fix low reach without finding the root cause amplified the problem.	Important news, product and investor announcements not heard.	Under most people’s radar.
20 years of documents were still answering for the company as if each were current	Decades of documents, much superseded, most still reachable, read with the same confidence as the current label. Retired material stays online, mirrored or cached. More is out there than the organisation’s records show.	Superseded material repeated as current.	Route removal after the landscape stage. AI discovery an option.
Residents get different answers to the same question, and the cost lands on the council	Services, districts and arms-length bodies each publish their own version. A person asking AI gets one, served with confidence. Nobody is wrong; the estate as a whole is.	Conflicting information to the public, then repeat contact, complaints, rework and remediation.	The cost consequence, missing from the page today, is the point.
The estate the digital team had warned about for years	The digital team warned of the sprawl for years; the first step in correcting AI was the footprint. Higher education instance: a university replatformed to a single platform, single point of governance and control, and removed the old sites. A 9-month project the team had wanted for 3 years, brought forward by what AI was surfacing.	Ongoing savings: operating cost, and beyond it better information, more confidence, better-informed applicants and control of the message.	The positive case. The team was right, and the return is measurable.
The corrected storage instructions lost to the document they replaced	Storage information was updated into a new PDF but the versioning was not correct. The original carried more authority and an agentic system treated it as current.	Correct information published and defeated by the document it replaced.	Directors’ liabilities increased through failures of internal systems. PDFs carry more authority in AI results than the format’s machine reading is tested for.
Steps in the wrong order, temperatures transposed	A 9-step procedure returned out of order and temperatures swapped, from a document that looked fine to a person. Second instance: pharmaceutical equipment operating instructions reconstructed out of order after human review.	Potentially dangerous results at the point of use.	The general reading order problem lives here, never led by its technical name.
Investors were quoted guidance the company had already revised	The company issued guidance and later revised it. The earlier range stays online and AI cites it when an investor asks.	Investors acting on a withdrawn range.	Written from the mechanism as an illustrative case.
A withdrawn announcement was quoted as the company's current statement	A withdrawn or corrected regulatory announcement or release stays reachable and AI quotes the original as the current statement. Labelled by jurisdiction.	A withdrawn statement treated as the company’s position.	Illustrative.
Old capital return figures served by third parties	Dividend and buyback figures investors rely on are held by third parties. Where the company’s own site cannot be read well enough, the third party’s out-of-date figures take priority.	Investors receiving historic figures as the current position.	The third-party reach mechanism from 10, applied to a poorly performing website.
Investors were told the former chair still ran the board	Directors and officers who have left are named as the current board.	The governance picture wrong at the first question.	Outdated facts persisting in AI descriptions is a documented category (Lantern).
Discontinued division described as trading	A business the company closed or sold is described as part of the group today.	The business AI describes no longer exists.	Discontinued products persisting in answers is a documented category (Lantern).
Regional sites contradict the group position	Country sites carry different figures, products or statements from the group site, and AI answers from either with equal confidence.	2 positions in the market at once.	The blending mechanism from 36, applied to the group.
An old annual report answered as this year's after the page that dated it was removed	The web page that framed a report as historic has been removed. The PDF lives on, carries more authority than the page did, and answers as current.	A historic report treated as the present.	PDF authority in AI results.
New York City’s business chatbot told companies to break the law	The city’s business chatbot misstated policies and advised businesses to break the law. Identical questions got different answers at different times. The chatbot was taken down.	Citizen frustration, wrong information acted on, legal action against the municipality, withdrawal.	Public-record confirmation of 4 and 39. // Table B. What stands behind each case // Evidence status, figure with its locked scope, directors’ and officers’ liability, insurance and commercial exposure, links.

The use cases: whathappened, and the impact.

The use cases: what
happened, and the impact.