Why marketers need to know how the tools work
Most of what goes wrong with AI in a marketing team is not a tool failure in the technical sense. The tool did what it was built to do; the team expected something else. A model asked for a statistic produces a plausible one because producing plausible text is its job. A help-centre assistant cites an old returns policy because the old policy is still in the documents it searches. A keyword clustering feature puts two queries together that search engines treat quite differently, because it grouped them by meaning rather than by the results pages. In each case, knowing the mechanism would have predicted the failure and shown where to put the check.
That knowledge also changes how a team talks to vendors. A claim that a tool "knows your brand", "never makes things up" or "learns from your data" means something specific in each case, and the right follow-up questions only occur to someone who knows what the possible meanings are. The AI in marketing pillar sorts AI into generative, predictive and classification-and-retrieval kinds; this section opens up the first and third, which are the ones a marketer briefs directly.
None of it needs mathematics. The pages here describe mechanisms at the level of general principle, the level that holds across tools and survives the next product release, and they make no claims about how any particular vendor's system is built.
How models produce text and errors
How language models generate text is the starting point. It explains training on large amounts of text, the way input is broken into small pieces, and how a model produces an answer one piece at a time by choosing among likely continuations. It then explains the consequence that matters most for marketing: a model optimised to produce likely text will produce a likely-sounding figure, quote or citation when it does not have a real one, and nothing in the output marks the difference. That is the mechanism behind the failure modes that the accuracy and review workflows guide catalogues and builds checks around.
The same page covers what a model can and cannot see: the instructions and material in front of it, but not the team's files, last week's price change or anything after its training unless a tool supplies it. Most "the AI got it wrong" moments trace back to that boundary.
Grounding, retrieval and similarity
Two mechanisms extend what a model can do with a team's own material. How grounding and retrieval work explains what happens when a tool answers from documents: the documents are split and indexed, the passages most relevant to a question are fetched, and they are placed in front of the model with an instruction to answer from them and cite them. It also explains where that chain breaks, and why the state of the source library usually matters more than the model. The tutorial on preparing a document library for an AI assistant puts the library in shape and keeps it there: one current source per fact, owners and review dates, sections that stand alone, and test answers traced back to the passages they came from.
How embeddings group similar text explains the idea underneath semantic search, keyword clustering and the grouping of customer feedback: text represented as a position, so that passages with similar meaning sit close together. It also shows where similarity of meaning misleads, for instance when two phrases mean nearly the same thing but signal different search intent.
Working with models: prompts, libraries and tests
Three pages cover the practical skills. Prompt patterns for marketing treats the prompt as a brief and sets out named structures that work across content, research, data and advertising tasks: the brief prompt, extraction to a table, critique and revise, generate then rank, and rewriting for a reader. Building a team prompt library is a tutorial for turning the prompts people keep retyping into shared templates with named inputs, test cases, versions and owners. Testing prompts and models sets out a repeatable method for deciding whether a prompt, model or tool is good enough: a test set drawn from real work, a scoring rubric, blind comparison and a pass bar, rerun whenever something changes.
Testing is the thread that ties the section together. The NIST AI Risk Management Framework names "measure" as one of its four functions alongside govern, map and manage, and for a marketing team measuring mostly means knowing, with evidence from its own tasks, how often a tool's output is usable. It is the same discipline as an A/B TestControlled comparison of two variants., applied before anything is published rather than after.
Choosing an approach: rules, agents, retrieval or tuning
Two comparisons help with decisions that vendors often frame for you. Rule-based automation vs AI agents sets fixed workflows against agents that plan and act, on predictability, auditability, cost of an error, upkeep and oversight, and covers mixed designs where an agent drafts inside a rule-based flow with approval. Retrieval vs fine-tuning for brand knowledge compares the two ways of making a tool know a brand's products, policies and voice, and sets out what a promise of a model "trained on your brandTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself." should mean before anyone signs.
| Question the team is asking | The approach it usually points to | Page |
|---|---|---|
| Can this run the same way every time? | Rule-based automation | Automation vs agents |
| Are the steps impossible to write down in advance? | An agent, with approval before customers see the result | Automation vs agents |
| Do the facts change often? | Retrieval from maintained documents | Retrieval vs fine-tuning |
| Is the need a stable style or format? | Prompts and examples first, tuning only if measured gaps remain | Retrieval vs fine-tuning |
| Is the output good enough to rely on? | A test set and rubric before any rollout | Testing prompts and models |
A yarn retailer works out why its assistant invents stock
Pellwyn Wool Co, an invented online yarn retailer with a marketing team of 4, adds an AI assistant to its product pages. In the first 2 weeks, staff log 31 answers that mention colourways the shop has never stocked. Reading the how-it-works pages, the team sees that the assistant is grounded in a folder of 260 product documents, of which 70 describe discontinued lines and none carries a date. When a question has no good match, the model fills the gap with a likely-sounding colour name. The team archives the discontinued documents, adds a stock-status line to each remaining one, and changes the instruction so the assistant says it does not know when nothing relevant is found. It then builds a test set of 40 real customer questions, scores answers with a 3-point rubric before and after the change, and keeps the change only after the logged invented colourways drop from 9 of the 40 test answers to 1.
The example is the section in miniature: the mechanism explains the symptom, the fix sits in the documents and the instruction rather than in a new tool, and a test on real questions decides whether the fix worked. Each KPIA metric tied directly to an objective. a team sets for AI use, such as the share of drafts that pass review first time, depends on that kind of evidence rather than on a vendor's description.
Further guides
Further guides turn the mechanisms into routines a team can run:
- Preparing a Document Library for an AI Assistant — a tutorial for taking an existing pile of help articles, policy pages and internal guides to a library an assistant can answer from: list what the assistant can reach, take out personal and internal-only material, keep one current source per fact, give every document an owner and a review date, rewrite sections to stand alone, trace test answers to passages and tie updates to policy and price changes.
Common foundations problems
Most of the questions marketers bring to AI tools trace back to one of a handful of mechanisms. The table maps each symptom to the page that explains and fixes it. When the symptom is already a pattern, such as a help-centre assistant answering customers wrongly week after week, the AI in marketing troubleshooting section traces a sample of failures to the source, prompt, model, review or permission that let them through.
| Problem | Usual cause | Guide |
|---|---|---|
| The tool states a figure, quote or source that does not exist | The model produces likely text and has no fact to draw on | How language models generate text, accuracy and review workflows |
| A help-centre bot gives an outdated or wrong answer | Stale, duplicated or badly split source documents | How grounding and retrieval work, Preparing a document library |
| AI keyword groups disagree with what ranks | Meaning-based similarity is not the same as shared search results | How embeddings group similar text |
| Output is generic and off-brief | The prompt lacks audience, material, constraints or a quality bar | Prompt patterns for marketing |
| Everyone prompts differently and results vary by person | No shared, tested templates | Building a team prompt library |
| A tool looked great in a demo and disappointed in use | Judged on one good answer rather than a test set | Testing prompts and models |
| An automated agent did something nobody approved | An agent used where a fixed workflow or an approval step was needed | Automation vs agents |
| A vendor offers a model "trained on your brand" | Unclear whether that means retrieval, tuning or neither | Retrieval vs fine-tuning |
The section explains how the tools work; the applications section covers what to use them for, and the governance section covers how to keep that use accurate, honest and accountable. Content that has to demonstrate experience and expertise, summed up as E-E-A-TExperience, Expertise, Authoritativeness, Trust., still needs people with that experience whichever mechanism drafted it.
