What AI-assisted work has to show
The guide to showing your own part in a marketing result asks you to write down everyone who touched a result and what each of them did, then to find the single line that is yours. Its list of contributors names strategists, copywriters, designers, analysts, developers, agencies and the manager who argued for the budget. It does not name a tool. When a tool wrote the first draft of the copy, built the first version of the analysis or produced the variants that were tested, the honest answer to "what did you do?" changes, because the words, the table or the list may have come from the tool while the decisions that made them usable were yours.
That is a question about evidence, not about whether using a tool is acceptable. The skills hub explains that the Chartered Institute of Marketing frames competence as knowledge together with its demonstrated application, and the site's guide to building AI skills across a marketing team lists what AI work asks of a person: briefing, evaluation, fact and claim checking, data literacy, data handling, experiment design and knowing when not to use a tool. Each of those skills can leave a trace: a brief, a source pack, a list of checks, a rejected draft, a decision to do the task by hand. An output on its own shows what a tool can do. The trace shows what you did with it.
Your part also depends on the seat you held. The accuracy and review workflows guide names four roles for an AI-assisted piece: the operator, who prompts the tool and answers for what went in, the reviewer, the approver and the owner. In a small team one person can hold more than one seat, and the guide asks for sign-off to be recorded. Which seat you held decides which sentences you can truthfully write.
This tutorial covers describing the work. Learning AI skills is the manager's plan in the guide above and, for an individual, the practice stage the skills hub describes; writing a team's rules is the policy tutorial linked in the first step; and how any employer, client or hiring manager reads a candidate's use of AI is a question no page the site cites documents, so nothing here claims to answer it. Where an application, an employer or a client states its own rule about AI, follow it. Questions of contract, confidentiality, client terms, data protection and disclosure belong to your organisation's policy, its legal and compliance reviewers and, where needed, an adviser. Everything below is general practice, not advice.
Before you start
- The piece of work, with its brief. What it was for, who it was for, when it ran and what you were asked to do.
- Your organisation's AI rules, if they are written down. An AI use policyA team's written rules for AI: approved tools, what each use may do, what data may go in, what is reviewed, when customers are told and who owns problems., if there is one, is the first place to look. If the work was for a client, add the client's terms and the agreement between the client and your employer.
- Whatever you kept at the time. Briefs and promptsThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told., the source material you gave the tool, drafts in order, review comments and sign-off messages. If you kept very little, the steps below show what to keep next time.
- Your private record of the actual figures, plus the permission, if you hold one, to describe the work. The shared version of a case is written from this record and does not replace it.
- The measure the work was meant to move, such as the KPIA metric tied directly to an objective. it served, and the comparison you hold for it, such as a baselineAn estimate of what a metric would have done without a given change, campaign or promotion, against which the effect is measured. or an earlier period. If there is none, the work can still be described; it simply supports a smaller claim.
- Time to read your own notes. Allow time to read them, and to rebuild the record if it was not kept at the time.
The steps
Find out what your organisation lets you show
Before writing anything, find out what you may describe and to whom. Read the policy if there is one. The tutorial for writing a marketing AI use policy covers what such a policy decides. A case or a CV line is written outside the organisation, so the policy and your contract govern what you may say about the work in it, not the guide you are reading. Where the work was done for a client, or by an agency for a business, the agreement should settle which AI uses are permitted and which are excluded; the guide to AI rules for agencies and contractors sets out the points such an agreement covers. Whether and how you may describe the result is a question for your contract and, if you are unsure, your manager or the client, as the portfolio guide's confidentiality section says; the CV and contribution guides add your organisation's policy. Then ask the person who owns the policy or the client relationship, in writing: may you describe this work, in what detail, and may you name the kind of tool that was used? Keep the answer with the record. If no rule exists, say so in the record and describe the work at the level you would be comfortable showing your manager first.
List where a tool did part of the work
Write one line for each use, not for each tool: researching, outlining, drafting, rewriting, summarising a report, producing variants, generating an image, writing a formula or a query. Include the AI features built into tools you already use, since the policy tutorial notes that built-in features are routinely missed. If your team keeps a register of uses, find your rows in it; the AI use register template shows what each row holds, from the task and the data the tool saw to the risk tier and the owner. Note the tier each use sat in. A use whose output went to customers after an editor's review carried a different risk from notes written for yourself, and how much checking stood behind the work is part of what the case shows.
Split each use into the tool's part, yours and the reviewer's
For each line on the list, write three short statements: what the tool produced, what you decided or supplied, and what a person other than you reviewed or approved. "What you supplied" covers the brief, the sources, the examples and the constraints. "What you decided" covers what you asked for again, what you rejected, what you changed and what you left out. A human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. is the step at which a named person checks what a tool helped to write before it is published or sent, so the reviewer's statement should carry a name or a role and a date. A grid in the style of the RACI matrix does the same job at a glance: each use gets a row, the tool, you, the reviewer and the approver are the columns, and exactly one accountable role sits in each row. The AI-assisted content review checklist is worth reading beside it, because its sign-off group asks for a named human reviewer and a recorded approval, and the accuracy and review workflows guide's rule that the person who prompted the tool does not review that output once it goes beyond internal use is what makes that reviewer someone other than you.
Check every fact and figure against a source you hold
The failure modes of AI-assisted output do not stay inside the work; they travel into the write-up. A fluent draft can contain an invented figure, a reference to something that does not exist or a claim nobody can support, and a case built on that draft inherits it. Treat the write-up as a document to be checked, not only the work. The fact-checking tutorial gives the claim-by-claim check; apply it to each statement about the work, with your own record as the source. A statement you cannot trace to the record, a report or a message from the time comes out, or becomes a question you put to the person who knows.
Keep the record while you work
Open one note for each piece of work at the start, and add to it as you go. Rebuilding it from memory later is how claims drift. The table in the next section lists the fields. Its last column borrows the four functions in the NIST AI Risk Management Framework, which the site's framework page turns into plain questions for a marketing team: govern, map, measure and manage. For one person's work they become the headings of a short record: who answers for the use and under which rules, where a tool was used and what could have gone wrong, how it behaved when it was checked, and what was done about what the check found.
Choose the verbs the evidence supports
The contribution guide's test of the verb applies to a tool's work too. Words such as briefed, sourced, checked, rewrote, rejected and decided describe your actions and can be backed by the record. "Wrote" is a claim about text you composed, so it needs support when a tool composed the first version, and "built" or "analysed" need the same care for a table or a query the tool produced. Say what the tool did in a plain phrase and then say what you did. The site's AI-assisted content guide describes disclosure in similar terms for published work: say in plain words what the tool contributed and state that someone named had checked and edited the output. If nobody else reviewed the work, say that plainly too, and do not describe a review that did not happen.
Write the case around the decisions, not the tool
Use the structure of the portfolio guide: context, decisions, results and learning. In the decisions, write what you asked the tool for and why, what you gave it, what you rejected and what you changed. In the results, keep to what the record supports. Time saved, defects caught before publication and drafts rejected are inputs you moved, and they can be stated with the comparison you hold for them. A business result such as orders or revenue needs the comparison the contribution guide sets out, and a case with none should say that it makes no claim about the result. In the learning, say where the tool failed and how the brief, the source pack or the review changed afterwards. For the CV, one line is enough: the kind of use, the part you played and the check that stood behind it, placed beside the skills it shows, as the CV guide advises for tools and credentials. A line that names a check and a decision shows what you did.
Check any text a tool drafts about you against the record
If you use a tool to tidy or draft the write-up itself, the same discipline applies. A prompt works like a brief, and a tool given a thin one fills the gaps with whatever is typical, which in a case study means responsibilities you did not hold, results you did not measure and features of tools you did not use. Give it your record rather than your memory, use groundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. in the same way as for any draft by telling it to draw only on what you supply, and share only what your organisation's rules let you put into a tool. For anything that touches customer or personal data, prepare the material first with the data minimisation routine. Then read each line against the record and delete what the record does not contain.
Prepare for the questions about how you worked
A reader who accepts the claim may still ask how you worked, so prepare short answers. What did the tool do, and what did you do? What did you check, and what did it get wrong? What would you have done without it? Which rule did you work under? Which parts of the task would you leave out of a tool, and why? The interview preparation tutorial gives the structure for a story: situation, task, action, result and what you learned. Choose one story in which a check caught a mistake the tool made, because it shows evaluation, one of the skills the AI skills guide lists, and it is a plain example of the work that stayed with you. Say what you do not know where you do not know it.
Revisit the record when rules or tools change
Policies, tools and the models behind them change, and a case written under one rule may describe a use that a later rule no longer allows. Before you reuse a case, read the current policy again, check that the permission you were given still covers what you are showing and update the wording if the tool or the rule has changed. A policy carries a review date, and your record can carry one too: the date you last checked that what you describe is still something you may describe. A case that is out of date on its rules is a weaker piece of evidence than a smaller case that is current.
What the record holds
The record can be one page per piece of work. The rows are general practice, built from the roles and logs the site's guides describe for a team and scaled down to one person's work. Add or drop rows to suit your organisation's rules.
| Field | What to write | Question it answers |
|---|---|---|
| The rule | The policy, tier and data rule that applied, and who gave permission to describe the work | Who answers for it, and under which rules? |
| The uses | One line for each task a tool did, with the tier it sat in | Where was a tool used, and what could go wrong? |
| What you gave it | The brief, the source pack and any examples, with their versions | Where was a tool used, and what could go wrong? |
| What it produced | The drafts or outputs, in order, with dates | How did it behave when checked? |
| What was checked and found | The facts traced, the defects by kind and what was removed or rewritten | How did it behave when checked? |
| What you changed and why | Edits, rejected options and the reason for each | What was done about what was found? |
| Who reviewed and approved | The reviewer and approver by role, the date and where the sign-off is kept | Who answers for it, and under which rules? |
The shareable version of a case is a summary of this record. The measure, its comparison, what else changed and what is confidential belong in the record the contribution guide describes, kept beside this one.
A marketing executive writes up AI-assisted product copy (illustrative)
Tukond writes up a quarter of product copy
Tukond, an invented maker of alpaca-wool blankets with its own online shop and 11 stockists, has a marketing team of 3. Its marketing executive uses an AI writing assistant, approved under the team's AI use policy at the tier for published work with an editor, to draft 36 product descriptions and 6 care emails in one quarter. She writes each brief, assembles a source pack from the maker's fibre, size and care specifications and tells the assistant to draw only on that pack. As her own pre-check, before anyone else sees a draft, she traces each statement in it to the pack. The head of marketing, who did not write the prompts, then checks each draft's claims against the same source pack and approves it by email, which the executive saves.
Her pre-check covers 214 statements against the pack: 187 match it, 21 are reworded and 6 come out, among them an invented fibre origin and a "hypoallergenic" claim the business cannot substantiate. Her time log from the previous quarter shows about 40 minutes to write and check a description by hand; this quarter it shows about 15, including her pre-check. She kept the briefs, the source pack versions, the drafts and the log in one note, and asked the head of marketing in writing whether she could describe the work. The answer was yes, in anonymised form, naming the kind of tool but not the supplier, and without the stockist list or the maker's unpublished prices.
Her case reads: "Over one quarter an AI writing assistant drafted 36 product descriptions and 6 care emails under my team's policy. I wrote the briefs and the source pack from the maker's specifications, traced 214 statements to them, rewrote 21 and removed 6, including an invented fibre origin and an unsupported claim; my head of marketing checked each piece against the same source pack and approved it. Drafting and checking time per description fell from about 40 minutes in the previous quarter's time log to about 15. I make no claim about sales, which I did not compare, and I did not separate out what else changed that quarter." Her CV line is one sentence of the same facts. For interview she prepares the story of the invented fibre origin, an answer on what she would do without the tool, and a short account of why she kept the claims check by hand.
The figures in the example do not carry over; the pattern does. The rule is checked and the permission asked before anything is written, each use is split into the tool's part, hers and the reviewer's, each statement is traced to a source she holds, the record is kept at the time, the wording names what the tool did and what she did, the result claimed is an input she measured against her own log and not a business outcome, and the questions about how she worked are prepared in advance.
Common mistakes
- Leaving the tool out of the story. A claim that reads as if you composed what a tool drafted is the version you cannot explain when asked how you worked. Say plainly what the tool did.
- Leading with the tool. A case that is largely about the product used shows what you logged in to, not your judgement. Lead with the decision, the check and the result the record supports.
- Describing a review that did not happen. If someone else reviewed the work, name the role; if nobody did, say so. Do not borrow a reviewer's part, and do not claim a check you did not make.
- Showing outputs you may not show. Prompts, source packs and drafts can hold confidential or personal data and commercial detail. Ask first, anonymise, or describe the work instead.
- Claiming a business result from a process gain. Time saved or errors caught are inputs you moved. Orders, revenue and retention need the comparison and the wording ladder in the contribution guide.
- Letting a tool write your record. Each line it drafts must trace to your own record, or it comes out.
- Writing the record afterwards. A case rebuilt from memory loses the briefs, the versions and the defects, which are the evidence. Keep the record as you work.
- Reusing a case after the rules changed. A permission and a policy have dates. Read both again before the case goes anywhere new.
