Lesson objective
Draw the onboarding orchestration as a map of steps, agents, data, human decision points and records, and test it for gaps before it runs on a live client.
Plain-language explanation
Why this matters now
Lessons 1 to 3 designed three parts of onboarding: the decisions people keep, the claims register and the list checks. In practice these run as one orchestration, often across several tools: a form builder, the agency's project tool, a file store, an AI assistant, a CRM and one or more ad platforms. When something goes wrong, the first question is always "what happened, in what order, and who decided?" A map drawn before the workflow runs is how you can answer it.
- Professional guidance (NIST AI RMF 1.0, January 2023; voluntary): the Core has four functions, GOVERN, MAP, MEASURE and MANAGE. GOVERN "applies to all stages" of AI risk management. NIST states that after instituting GOVERN outcomes, "most users of the AI RMF would start with the MAP function" and continue to MEASURE or MANAGE, iteratively. Appendix C says human roles "need to be clearly defined and differentiated." There is no NIST certification, and the framework binds no agency. NIST has announced a revision of 1.0.
- Professional guidance (OWASP GenAI Security Project, OWASP Top 10 for Agentic Applications for 2026, announced 9 December 2025; voluntary community document): its ten entries include ASI06 Memory & Context Poisoning, ASI07 Insecure Inter-Agent Communication and ASI08 Cascading Failures. This lesson uses the names only, as prompts for questions to ask about your own map; it does not describe OWASP's definitions of them.
- Law / rule (Texas, Business & Commerce Code chapter 541; binds controllers and processors within the chapter): the act assigns duties to controllers and to processors, including a data protection assessment for processing for targeted advertising. An orchestration map is where an agency shows, for each flow of personal data, which role counsel or the contract says it holds.
A question to sit with
If a client asked why a copywriter drafted around a claim nobody had approved, could you point to the step in your onboarding map where that should have been stopped, and to the record showing whether it was?
What goes wrong
The map is the vendor's diagram. The only picture of the workflow is the automation tool's canvas, which shows triggers and actions but not who decides, what data moves or where it is recorded.
Decision points are drawn as boxes, not as people. "Claims review" appears as a step. It does not say who, what they see, what they record, or what happens if they are on leave.
Files feed memory. Intake material (briefs, brand decks, evidence packs) is loaded into an assistant's persistent knowledge for the client. A superseded product sheet or a claim that was later rejected remains there and resurfaces in drafts. The OWASP agentic list names memory and context poisoning as a risk category; stale or rejected material is the ordinary, non-malicious version of the same problem.
Agents hand work to agents. An intake agent passes a brief to a drafting agent, which passes copy to a scheduling agent. Nobody drew the hand-offs, so a bad input at the start travels to the end.
Roles are unrecorded. Nobody can say, for a given list, whether the agency acts on the client's instructions or has started deciding its own purposes, or who holds the targeted advertising assessment.
The practical answer
Draw the map with five lanes. The format is an AI Coalition Network recommendation, and the downloadable integration map for this level uses it.
| Lane | What it shows |
|---|---|
| Steps | Each step in order, with its trigger |
| Actors | Which person, agent or tool performs each step, and at which authority level from Module 01 |
| Data | What data enters and leaves each step, including personal data, claims and evidence, with its trust level |
| Decisions | Each human decision point: who, what they see, what they record, and their deputy |
| Records | Where each step and decision is logged, and how long it is kept |
Mark every human decision point with four facts. For acceptance, scope, claim approval, list approval and access set-up: the named decider and deputy; the information they see (the pre-acceptance summary, the claims register row, the list checks); the record they create; and what the workflow does while it waits. The default while waiting is to hold, never to proceed.
Trace the hand-offs. Where one agent or tool passes work to another, draw the hand-off and write what is checked at it. At minimum: the receiving step checks that the input carries the approvals it depends on (a claim ID approved to draft, a list tagged as approved), and rejects anything that does not. This is how an error at intake stops at the next hand-off instead of reaching the public.
Control what intake puts into memory. If an assistant keeps a knowledge base per client, the map shows what goes in and who approves it. Load only current, approved material; tag each document with its status and date; remove superseded and rejected material when the register changes. Rejected claims should never be in a drafting assistant's context except as clearly marked "do not use".
Record data roles on the map. For each flow of personal data, note the role the agency holds as recorded by counsel or in the contract, the client's instructions, and whether a Texas data protection assessment or its equivalent elsewhere exists for targeted advertising the data will feed. Where the role is uncertain, mark it "for counsel".
Walk the map before go-live. Take one fictional client through every step with the named people present. For each decision point ask: what would this person see, what would they record, and what would happen if they were away? For each hand-off ask: what would stop a bad input here? Fix the gaps, then approve the map in the agent authority register.
Remember this
- Map before you run: NIST AI RMF 1.0 describes most users starting with MAP once governance is in place.
- Use five lanes: steps, actors, data, decisions and records.
- Every human decision point has a named decider, a deputy, the information they see, the record they make, and a hold while waiting.
- Check approvals at every hand-off between agents and tools so intake errors stop early.
- Control what intake loads into an assistant's memory, and record the agency's data role for each flow.
Marketing agency example
Marrowfield Organics, a fictional snack brand, runs an in-house team that brings in two outside agencies each year for seasonal campaigns. Its marketing operations manager, Jonah Whitley, maps the onboarding orchestration the team uses to brief an agency. The vendor's workflow canvas shows nine steps and three AI actions. Jonah redraws it in five lanes. The Actors lane shows the brief-writing assistant acting at level 2, drafting inside the team's own workspace, and a file-sync step pushing the approved brief and brand pack to the agency's shared folder, which is an outward act at level 3. The Decisions lane exposes a gap: the claims decision is shown as a box labelled "legal", but nobody is named, and the brief is pushed to the agency whether or not it has happened. Jonah names the claims reviewer, Mara Keane, and her deputy, and adds a hold: the push step checks that every claim in the brief has an approved-to-draft ID and rejects the brief otherwise. The Data lane shows the team's customer list travelling to the agency for a paid social campaign. Jonah marks the agency's role as "processor, per contract clause 9, for counsel to confirm" and notes that Marrowfield's targeted advertising assessment for the campaign is held by the privacy lead. The Records lane shows that the assistant keeps a per-season knowledge base. During the walk-through, a claim rejected last spring, "boosts immunity", is found still in it and appears in a test brief. Jonah removes it, tags every document with status and date, and adds a rule that register changes trigger a knowledge-base review.
Responsible-use warning
Do not treat the automation tool's canvas as the map. A canvas shows triggers and actions; it rarely shows who decides, what personal data moves, which role the agency holds, or where each decision is recorded. And do not let a decision point proceed by default when the decider is away. The workflow holds until a named person or deputy decides, and every hand-off checks that the approvals it depends on are present.
Human-review requirement
The agent owner draws and maintains the map. Each named decider confirms what they will see and record at their decision point, and names a deputy. The technology owner confirms that hand-off checks and holds are actually enforced in the tools. Counsel confirms the agency's data role for each personal data flow and whether an assessment is required. The account lead or in-house owner approves the map in the agent authority register before the orchestration runs on a live client.
Practical exercise
Use a fictional client, or your own onboarding with nothing confidential.
- List every tool your onboarding touches, from forms and file stores to AI assistants, CRM and ad platforms.
- Draw the orchestration in five lanes: steps, actors, data, decisions and records.
- Mark each actor's authority level from Module 01, and highlight every outward act.
- For each human decision point, write the decider, deputy, what they see, what they record and what the workflow does while waiting.
- Draw every hand-off between agents or tools and write the approval check the receiving step performs.
- List what intake loads into any assistant's memory or knowledge base, who approves it, and how superseded material is removed.
- For each flow of personal data, record the agency's role and mark it "for counsel" if uncertain.
- Run the rejected-claim walk-through and list every place the claim could reach.
- Record the fixes, then note who approves the map and when it will be reviewed.
- Hold a thirty-minute walk-through with the named deciders using one fictional client, and for each decision point record what they said they would see, what they would record and what they would do if the information on screen were incomplete.
Knowledge check
-
What are the five lanes of the onboarding map in this lesson?
Answer guidance: Steps with their triggers; actors with their authority levels; data entering and leaving each step with its trust level; human decisions with decider, deputy, information and record; and records showing where each step is logged and for how long.
-
According to NIST AI RMF 1.0, where do most users start after instituting GOVERN outcomes?
Answer guidance: With the MAP function, before continuing to MEASURE or MANAGE iteratively. The framework is voluntary, has no certification, and NIST has announced that version 1.0 is being revised.
-
What should the workflow do while a human decision point is waiting for its decider?
Answer guidance: Hold. The default is never to proceed; the workflow waits for the named decider or deputy, and downstream steps check for the approval before accepting the work.
-
Why does intake material loaded into an assistant's memory need control?
Answer guidance: Because superseded product sheets or rejected claims can remain in the knowledge base and resurface in later drafts. Only current, approved material should go in, tagged with status and date, and removed when the register changes.
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.
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications for 2026 (2025-12-09). Accessed 2026-09-15.
- Texas Legislature — H.B. No. 4 (88th Legislature), Texas Data Privacy and Security Act, Business & Commerce Code chapter 541 (enrolled text). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.