Applications by task
The table groups the common applications by the task they support and names the guide that covers each. "Output" is what the tool hands back; "the person's job" is what no tool removes.
| Task | Typical AI output | The person's job | Guide |
|---|---|---|---|
| Research and outlining | Topic maps, question lists, outlines from a brief | Supply the angle; reject generic structure | AI for content |
| Drafting and editing | First drafts from supplied sources, rewrites for length and level, translations | Check every claim against the source; restore voice | AI for content |
| Image and video production | Generated and edited images, variations, resized and extended assets | Brand fit, rights, disclosure, review before publishing | AI for creative |
| Querying data | Queries written from a plain-language question; tables and charts | Confirm the definition and the filter match the question | AI for analysis and reporting |
| Monitoring and alerts | Flags on unusual movements in a metric | Decide whether the movement is real, a tag problem or seasonality | AI for analysis and reporting |
| Bidding and audiences | Bids set per auction; delivery expanded beyond chosen segments | Conversion tracking, goals, exclusions, holdout measurement | AI in ad platforms |
| Ad assets | Assembled and generated headlines, descriptions, images and video | Asset quality, pinning where needed, review of generated enhancements | AI in ad platforms, Reviewing AI enhancements in ad accounts |
| Classification and routing | Leads, tickets and replies tagged and routed by predicted category | Thresholds, approvals, a path for the wrong call | AI for marketing automation |
| Personalisation and triggers | Content chosen at send time; sends fired from a prediction | Guardrails on frequency, data and tone | AI for marketing automation |
The core guides
- AI for content — where AI fits the content process from research to localisation, where quality fails, and what Google's published guidance says about AI-generated content. It takes the application view and leaves the production workflow to the content marketing section.
- AI for creative — image and video generation, variations for testing, adaptation across sizes, brand consistency, rights and disclosure, and the creative tools the ad platforms document in their own interfaces.
- AI for analysis and reporting — plain-language querying, anomaly detection, report narration and forecasting, with the verification each one needs before a number is acted on.
- AI in ad platforms — the automation already running inside Google Ads and Meta Ads, as the help centres describe it: bidding, audiences, assets and the campaign types that bundle them, plus what the advertiser still controls.
- AI for marketing automation — models inside workflows: classification, send-time personalisation, agents that draft and act, prediction-driven triggers, and the guardrails that keep them bounded.
Governance runs across all of them. The governance section covers the policy, review tiers and disclosure rules that each guide assumes, and the hub page sets out how applications and governance fit together.
Task guides
Beneath the core guides sit pages for specific jobs. Each one names the task, the review it needs and the failure to watch for. The foundations section explains how the models behind them work, and testing prompts and models shows how to check a workflow before relying on it.
- AI for Alt Text, Captions and Transcripts: how to use AI to draft alt text, video captions and transcripts that meet accessibility needs, what a person must check, and where generated text gets it wrong.
- AI for Keyword Research and Clustering: where AI helps keyword research, from seed ideas and intent labels to grouping, and where it fails: invented volumes, stale data and groups results contradict.
- AI in Customer Service and Chat: answers from approved content, agent reply drafts, triage, summaries, handover, disclosure and review.
- AI in Search Results: What It Means for SEO: what Google's published guidance says about AI-generated content and spam, how AI answers change results pages, and what that means for content and measurement.
- AI-Assisted Ad Copy Testing: how to use AI to draft ad copy variants that test a real hypothesis, check claims before launch, fit platform formats, and read results honestly.
- Building a Brand Voice Guide for AI Drafting: turn a brand's voice into instructions an AI tool can follow: traits with paired examples, word lists, reference passages, test prompts and review rules.
- Checking AI-Generated Structured Data: how to use AI to draft structured data for a page template from its data fields, then check it against Google's guidelines, the visible page and the feed.
- Checking AI-Written Spreadsheet Formulas and Calculated Fields: a routine for checking the spreadsheet formulas and dashboard calculated fields an AI tool writes: definitions, ranges, lookups, blanks, dates and a hand total.
- Checking AI-Written SQL Step by Step: a routine for using AI to write SQL against marketing data and proving the answer right: schema context, definitions, joins, filters and reconciliation.
- Checking AI-Written Tracking Code Before It Ships: a routine for using AI to draft data layer pushes, tags, triggers and regular expressions, then proving them in preview against the tracking plan and consent.
- Generating Product Descriptions from Catalogue Data: how to generate product descriptions with AI from catalogue data: attribute hygiene, templates by category, banned claims, sampling review and refresh.
- Producing AI-Generated Ad Images with Review: brief, references, generation, selection, editing, rights and likeness checks, labels and final QA.
- Reviewing AI Enhancements in Your Ad Accounts: how to find, preview and decide on the AI enhancements and generated assets in Google Ads and Meta accounts, with an owner, a record and a re-check routine.
- Standalone AI Assistants vs Built-In AI Features: general-purpose AI assistants compared with AI features inside the tools a team already uses: data access, context, control, terms, cost and where each fits.
- Synthetic Personas and Simulated Respondents: what AI-generated personas and simulated survey answers can and cannot tell a marketer, the few safe uses, and why they never replace real research.
- Using AI for Research Synthesis: how to use AI to synthesise interviews, open survey answers and reviews into themes that trace back to evidence, with consent, quote checks and bias controls.
- Writing the Standing Instructions for an AI Assistant: how to write the standing instructions an AI assistant follows in every conversation: scope, sources, refusals, handover, disclosure, tone, tests and versions.
Judging an application: value, review cost, risk
Any proposed use of AI can be judged on three questions, and a use is worth adopting only when the answers line up.
Value. How much effort does the task take today, and how much of that effort is the first draft rather than the thinking? AI removes the first-draft cost; it does not remove the thinking. An application that saves an hour of drafting but adds an hour of review is not a saving, and the only honest way to know is to time the task with and without the tool, review included. The KPIA metric tied directly to an objective. for a pilot is time per acceptable output, not time per output.
Review cost. How cheaply can a person verify the result? An outline is judged at a glance. A rewrite is compared to its source. A draft that makes factual claims has to be checked claim by claim against something the reviewer can open. A generated figure that cannot be traced to a query and a table cannot be verified at all, and has to be treated as unknown. The cheaper the check, the sooner the application is safe to use widely. Where no one reads the output before the customer does, as with an assistant answering on a website, the checks move into its standing instructionsThe business's rules placed before every question an AI assistant answers: scope, sources, refusals, when to hand over and what never to promise. and the tests run against them.
Risk. What happens when the output is wrong and nobody catches it? A bad outline wastes an afternoon. A wrong product claim reaches customers. Customer data pasted into a prompt may leave the organisation's control. Automated bidding chasing a mistracked conversion spends real money on the wrong outcome, and the ROASRevenue generated per unit of ad spend. it reports will look fine while it does so. The qualities readers and search engines reward, E-E-A-TExperience, Expertise, Authoritativeness, Trust., are a useful shorthand for the content side of this: an application that removes the expertise from a piece raises its risk whatever it saves.
Scoring three proposed uses
A software company's marketing team scores three proposals on a 1 to 5 scale for value, review cost (5 is cheapest) and risk (5 is lowest). Headline variations for existing search ads score 4, 5 and 4: quick to judge, and the worst case is a weak ad that the auction will not favour. Drafting case studies from customer interview transcripts scores 5, 2 and 2: high value, but every quote has to be checked against the transcript and a misattributed quote would need a correction and an apology. Routing inbound leads to sales reps by predicted deal size scores 3, 3 and 2: useful, but a mis-routed enterprise lead sits with the wrong person for a week. The team adopts the first immediately, adopts the second with a named reviewer who checks every quote, and defers the third until the CRM can show a holdout comparison of routed versus unrouted leads.
Common application problems
Most problems with AI applications show up as one of a few symptoms. The table maps each to the guide that deals with it.
| Symptom | Usual cause | Guide |
|---|---|---|
| Published text contains an invented fact, quote or feature | Drafting without supplied sources; review that read for fluency rather than truth | AI for content |
| Every piece sounds the same | One prompt reused across topics; no brief behind it | AI for content |
| Generated images are off-brand or carry the wrong details | No brand references in the brief; no review of extended or edited areas | AI for creative; for platform-generated assets, Reviewing AI enhancements in ad accounts |
| A reported number cannot be reproduced | Query generated against a different definition or filter than the question meant | AI for analysis and reporting; for a spreadsheet or dashboard figure, Checking AI-written formulas |
| Alerts fire constantly or never | Thresholds set without reference to normal variation; seasonality not modelled | AI for analysis and reporting |
| Platform spend appears efficient but revenue does not move | Bidding optimised to a weak or double-counted conversion; no holdout | AI in ad platforms |
| Ads reach people far outside the intended audience | Expansion features left on with no exclusions or controls | AI in ad platforms |
| Automated messages sent to the wrong people or at the wrong time | Prediction-driven triggers without guardrails or approvals | AI for marketing automation |
| Customers keep asking an assistant for a person or come back by phone or email | Contact reasons the help content does not answer; a route to a person that is hard to find | AI in customer service; for a pattern, AI assistant not resolving customer questions |
