How these pages work
Most diagnoses here begin by collecting the failures, because one wrong answer proves little and a pattern names the layer. An accuracy problem starts with a log of recent inventions and the source each one should have come from; a wrong-answer problem starts with a sample of wrong answers traced to the passage each one cited; an unresolved-question problem starts with the conversations that ended in a request for a person; a creative problem starts with the rejection exactly as the platform gave it; a reporting problem starts by rebuilding the disputed figure from the source data; an automation problem starts with the run that sent the messages, paused; a quality problem starts by dating the change from the review log; and a data problem starts by telling the data protection owner, then listing every case. Once the fault is visible, a diagnosis table lists what you can see, the check that confirms each cause and the fix. Each cause then gets a short section with the fix, a way to verify it and the habit that prevents it, followed by a routine that keeps the whole set under watch.
The pages assume the methods are taught elsewhere and link to them. Why a model produces likely text rather than checked facts is in the guide to how language models generate text; how GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. works and how a tool finds passages in a library is in the grounding and retrieval guide; building a test setA collection of real inputs for an AI task, each noting what a good output must contain, run to compare prompts or models and rerun when they change. and rerunning it after every change is in the guide to testing prompts and models; and review tiers, fact checks and claim checks are in the accuracy and review workflows guide.
Find the page for your symptom
Choose the row that matches what you are seeing. Diagnosis pages in this section are also listed beneath this introduction automatically; the last three rows send situations that belong somewhere else to the page that owns them.
| What you are seeing | Start with | Background |
|---|---|---|
| Drafts or published pages contain figures, quotes, awards or features that do not exist | AI-written content contains invented facts | Accuracy and review workflows |
| A customer-facing assistant quotes an old policy, the wrong plan or a price that has changed | AI assistant giving customers wrong answers | Preparing documents for an AI assistant |
| Customers keep asking the assistant for a person, repeat the question or give up and phone or email instead | AI assistant not resolving customer questions | AI in customer service |
| An ad platform rejected, labelled or removed a generated image, video or copy variant | AI-generated ad creative rejected | AI for creative |
| A figure an AI tool produced cannot be reproduced by the person who checks it | AI-generated report numbers do not match | AI for analysis and reporting |
| A workflow with an AI step sent an unapproved offer, a wrong name or a message to the wrong people | AI automation sent the wrong messages | AI for marketing automation |
| Customer lists, transcripts or order histories have been pasted into an AI tool | Customer data pasted into AI tools | Minimising data before prompting |
| Output from a prompt or assistant that used to pass review now needs heavy editing | AI output quality getting worse | Testing prompts and models |
| The team uses AI tools every day, yet deadlines, output and workload look the same as before | AI tools not saving time | Measuring AI's impact on team output |
| An AI error has already reached customers and needs containing today | Handling an AI mistake that reached customers | Writing an AI use policy |
| AI-assisted pages are live but do not rank or convert | AI-written content not performing | AI for content |
| You need to check one answer a tool gave before anyone relies on it | Fact-checking an AI answer | Prompt patterns for marketing |
Source, prompt, model, review or permission: decide which layer failed
Most symptoms in this section trace to one of five layers, and naming the layer first decides who fixes it. A source failure means the material the tool drew on was missing, stale or wrong: no sources supplied, superseded documents left in the library, a help centre that was out of date before the assistant read it. A prompt failure means the instruction never carried the audience, the material, the constraints or the rule to answer only from the library. A model failure means behaviour changed under a setup that used to work, through a new version, a vendor update or a prompt edit nobody retested. A review failure means a person read for fluency rather than tracing facts, or the gate before upload or sending was skipped. A permission failure means the tool was allowed to do more than its risk justified: an agent free to send or change records where a fixed workflow with an approval step was needed, or offers and legal copy left to the model instead of locked as rules.
The test is to ask what would have stopped the failure. If a correct source would have, it is the source; if the source was there and the tool ignored it, look at the prompt and the retrieval; if the same setup worked last month, look at what changed; if a person saw the output and passed it, look at the human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. step; and if nobody saw it before customers did, look at permissions. The NIST AI Risk Management Framework, as the site's governance guides apply it, organises the same work into four functions, govern, map, measure and manage: name an owner, know where each use sits and what could go wrong, measure how it behaves, and change the setup on that evidence rather than on one bad answer. Every statement these pages make about what models and tools do is the one the foundations section makes.
When to review instead
A diagnosis page fixes one failure. When several of these symptoms arrive together, when nobody can say which uses of AI the team runs or who owns them, or when a new assistant or workflow is about to go live, a review is the better tool. The AI use register template records every use with its risk tier, owner, data, permissions, reviewer, versions and incidents, and every control in the governance section attaches to it; the content review checklist is the gate every AI-assisted article, ad or email passes before publication; and the customer service launch checklist is the gate before an assistant answers customers, from a test on real past questions to a weekly review booked in advance. The AI marketing hub places foundations, applications and governance side by side.
