Lesson objective
Build a step-by-step map of one content, media or PR workflow from the people who do the work, showing every actor, input, tool, decision point, exception and record before any automation is proposed.
Plain-language explanation
Why this matters now
Module 1 ended with a rule: map before you manage. This module is how. Most agencies have a picture of their workflows already. It lives in a new-business deck, a process slide or the head of a department, and it shows the work as it is supposed to run: brief, draft, review, client approval, publish. The real workflow has more steps, more people, more exceptions and more waiting, and it is the real one that any AI tool will be plugged into.
Professional guidance (voluntary framework): NIST's AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, January 2023) organises its Core into four functions: GOVERN, MAP, MEASURE and MANAGE. NIST says that, after instituting GOVERN outcomes, "most users of the AI RMF would start with the MAP function" and continue to MEASURE or MANAGE, iteratively. The framework is voluntary, binds no agency, and cannot be certified against; NIST states that version 1.0 is being revised as part of the White House AI Action Plan. Its value here is the order of work it recommends: understand the context before measuring or controlling anything.
There is a second reason to map first. Research (not binding): Humlum and Vestergaard's working paper on Denmark (NBER Working Paper 33777, May 2025, revised March 2026) reports that employers absorb AI through task reorganisation, including new tasks in content generation, AI oversight and AI integration. It is not peer reviewed, uses Danish data, and does not list the occupations studied, so it says nothing directly about US agencies. It does make a practical point: adding AI rearranges tasks. If you cannot see today's tasks, you cannot see where the rearrangement will put new work, or which checks it will quietly remove.
A question to sit with
Ask the head of content to draw how a blog post moves from brief to publication. Then ask the most junior writer, the freelance editor and the person who uploads it to the client's site. How many of the four drawings will match, and which one will the AI tool be connected to?
What goes wrong
Mapping the ideal. A map drawn in a meeting room from memory describes the process people believe in. It leaves out the step where a writer checks last year's post for the client's preferred spelling, and the step where the account executive rewrites the headline because the client always does.
Missing the exceptions. Rush jobs, client-supplied copy, weekend posts, reactive statements and "just this once" requests often skip steps. They are also where errors cluster. A map that shows only the normal path cannot show where an automated step will meet an exception.
Invisible decisions. Many decisions in agency work are made in passing: a writer decides a claim is fine because it was used last month; a community manager decides a complaint does not need escalating. If the map does not mark those decisions, automation can remove the person who was making them without anyone noticing.
Workarounds and personal tools. People fill gaps in a process with their own fixes: a personal spreadsheet of banned words, a private AI account used to shorten copy, a message thread where approvals really happen. These are part of the workflow, whether or not they are allowed, and a map that omits them is wrong.
Mapping the tool, not the work. A vendor's diagram of where its product fits is a sales document, not a map of your workflow.
The practical answer
Map one workflow at a time, for one type of deliverable, from the people who do it. The method below is an AI Coalition Network recommendation, not a legal, professional or framework requirement, and no source measures whether it improves results.
Gather the real process from four sources.
| Source | What it shows | How to collect it |
|---|---|---|
| Walk-throughs | The steps people say they take | Ask each role to talk through their last three jobs, not a typical one |
| Records | What actually happened | Tickets, approval histories, file versions, sent-email logs |
| Observation | What people do but do not mention | Sit with someone for one full deliverable |
| Exceptions | Where the path breaks | Ask: "When did this last go differently, and why?" |
Record each step in a consistent form. Use the Agency Workflow Map resource or your own version with the same fields.
| Field | What to write |
|---|---|
| Step | A verb and an object: "check prices against rate card" |
| Who | The role, not the person's name |
| Input | What they work from: brief, claims file, prior post, client email |
| Output | What they hand on |
| Tool | Every tool used, including personal ones |
| Decision | Any judgement made at this step, and on what basis |
| Record | What evidence the step leaves behind, if any |
| Exceptions | When this step is skipped, changed or rushed |
| Time | Touch time and waiting time (Lesson 4) |
Mark the decision points. Put a clear symbol on every step where a person decides something: that a claim may run, that a comment needs escalating, that an image is cleared, that a send list is ready. Professional guidance (voluntary framework): NIST's AI RMF 1.0, Appendix C, says human roles in decision making and oversight "need to be clearly defined and differentiated." Marking decision points is how you make that visible for your own workflow.
Mark the checks with no record. A step that catches errors but leaves no evidence is the most fragile step on the map. Flag it now; Module 4 turns it into an enforceable SOP step.
Draw the exception paths separately. Show the rush path, the client-supplied-copy path and the out-of-hours path as their own routes. Note which checks each one skips.
Record personal tools without blame. If a writer uses a personal AI account, note it on the map as fact. The aim is an accurate picture. Whether that tool is approved is a separate decision (Module 7 covers client segregation), and a map that punishes honesty will stop being honest.
Validate the map with the people who do the work. Show each role the finished map and ask what is missing. Expect at least one correction per role.
Remember this
- Map before you automate; NIST's voluntary AI RMF 1.0 suggests most users start with MAP after GOVERN.
- Map the workflow as it runs, from walk-throughs, records, observation and exceptions, not from the deck.
- Mark every decision point and every check that leaves no record.
- Draw exception paths and record personal tools and workarounds as facts.
- Validate the map with the people who do the work before anyone proposes a tool.
Marketing agency example
Brackenmoor Content Studio is a fictional agency that writes weekly blog posts for a fictional garden-supplies retailer, Ashcombe Garden Co. The process slide shows five steps: brief, draft, edit, client approval, publish. The operations lead, Delphine Marsh, maps the workflow by asking the account executive, two writers, the freelance editor and the web producer to walk through their last three posts with the ticket history open. The real workflow has fourteen steps. Three matter most. First, one writer checks every product price against the retailer's weekly rate card, which she keeps in a personal spreadsheet; the other writer does not, and his posts account for most of last quarter's price corrections. Second, the client's approval often happens in a message thread, not the approval tool, so the ticket shows "approved" for posts the client approved only in part. Third, rush posts for weekend promotions skip the freelance editor entirely. Delphine also finds that the freelance editor uses her own AI account to shorten introductions, which nobody at the agency knew. She records all of it on the Agency Workflow Map without blame, marks the price check and the partial approvals as decision points with no record, and draws the rush path separately. The proposal on the table had been to add AI drafting between brief and edit. The map shows that doing so would remove the one person who checks prices, on the exact path where the editor is already skipped, and that the most urgent fix is the approval record, not the drafting step. She takes the map back to all five people, who add two more steps before it is final.
Responsible-use warning
Do not map a workflow from a process slide, a vendor diagram or a meeting of managers. Those show the process people believe in, not the one that runs, and they omit the exceptions, workarounds and unrecorded checks that automation is most likely to break. Do not skip the exception paths because they are untidy; rush jobs and client-supplied copy are where errors gather. And do not use the mapping exercise to punish people for workarounds, or the next map will leave them out.
Human-review requirement
The operation owner owns the map and signs off that it reflects how the work actually runs, after every role has seen it. Each role confirms its own steps, decisions and exceptions. The account lead confirms the client approval steps as the client actually gives approval. Where the map shows a decision about whether a claim may run, that decision stays with the client, advised by its counsel where needed, and the map should show it. No automation proposal goes forward on a map that has not been validated this way.
Practical exercise
Use a fictional agency, client and deliverable only. No real client, colleague or campaign.
- Choose one deliverable type your team produces, such as a blog post, a paid social set, a press release or a media plan.
- List every role that touches it, including freelancers, vendors and the client.
- For each role, write a walk-through of the last three jobs, not a typical one, using tickets, approvals and file versions as prompts.
- Record each step in the map fields from this lesson: step, who, input, output, tool, decision, record, exceptions and time.
- Mark every decision point with a symbol, and note the basis on which each decision is made.
- Mark every check that leaves no record, and write what evidence it could leave instead.
- Draw the rush path, the client-supplied-copy path and one other exception path separately. Note which checks each skips.
- Record any personal tools or workarounds as facts, without naming individuals.
- Write down how you would validate the map with each role, and what you would do if two roles describe the same step differently.
- Mark the final review step and confirm that it names a role and leaves a record. If it does neither, write what it should record.
Knowledge check
-
What does NIST AI RMF 1.0 say about where most users start, and does it bind an agency?
Answer guidance: NIST says that after instituting GOVERN outcomes most users would start with the MAP function, then continue to MEASURE or MANAGE. The framework is voluntary, binds no agency and is being revised.
-
Why ask people to walk through their last three jobs rather than a typical one?
Answer guidance: A typical job is described from memory and leaves out awkward parts. The last three real jobs, with records open, surface the exceptions, workarounds and unrecorded checks that automation is most likely to break.
-
Why is a check with no record the most fragile step on a map?
Answer guidance: Because nothing shows it happened, so when a step around it is automated it can stop without anyone deciding to stop it, and nobody will notice until an error reaches the client or the public.
-
How should a map treat a writer's personal AI account or private spreadsheet?
Answer guidance: Record it as fact, without blame, because it is part of how the work really runs. Whether the tool is approved is a separate decision, and punishing honesty would make the next map inaccurate.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- National Bureau of Economic Research (NBER) — Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (NBER Working Paper 33777) (2025-05-01). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.