Lesson objective
Map an intake-to-engagement orchestration as explicit states with owners, permitted agent actions, exit conditions, timeouts, failure paths and stop conditions.
Plain-language explanation
Why this matters now
The previous three lessons each fixed one decision point: what the intake agent says, when conflicts run, and who decides engagement. In a real firm those decision points sit inside a single flow that crosses several systems, several teams and several days, and most failures happen in the gaps between them rather than at the points themselves.
An inquiry sits in a queue nobody monitors over a holiday weekend. A conflicts search times out and the retry never fires. A prospective client replies to a reminder, and the reply lands in an inbox the workflow does not watch. An inquiry is submitted twice and becomes two records with two different clearance statuses. None of these is dramatic, and each can leave a person with an urgent problem and no answer, or a firm holding information it should not have.
An orchestration map makes the gaps visible. It is the document that says, for every state an inquiry can be in, who owns it, what the agent may do there, what moves it on, how long it may wait, and what happens when something breaks.
Security standard: the OWASP Top 10 for Agentic Applications for 2026, published on 9 December 2025, lists agent goal hijack, tool misuse and identity and privilege abuse among its entries. An intake orchestration reads material written by members of the public by design, so it is a natural place for those risks to appear. The list is a voluntary standard, not law, and its entries were confirmed from a secondary summary rather than the OWASP page, so check the PDF before quoting. Injection defence in depth is Module 10; this lesson places the controls on the map.
A question to sit with
Pick any inquiry in your intake system right now. Can you say what state it is in, who is responsible for moving it on, and what happens if nothing happens for three days?
What goes wrong
1. States are implicit. The workflow has steps but no named states, so an inquiry can be half-way between two of them and belong to no one.
2. Handoffs have no owner. "Passed to the team" names nobody. A queue without a monitored owner and a response time is a place inquiries go to wait indefinitely.
3. Timeouts are missing. Nothing defines how long an inquiry may sit in a state, so nothing happens when it sits there too long.
4. The unhappy path was never drawn. The map shows inquiries flowing from contact to engagement. It does not show what happens when conflicts cannot run, a person does not respond, a decision is deferred or a duplicate arrives.
5. The agent has more reach than the state needs. The intake agent can read existing matter records because a broad connector was convenient. In the names stage it needs a conflicts search, not the contents of matters.
6. Public text steers the workflow. Content in a web form is treated as something the agent should act on, rather than as information to record. A cleverly worded submission changes what the orchestration does.
The practical answer
Name every state and give each one an owner. A workable map for intake to engagement looks like this.
| State | Owner | Agent may | Exit condition | Timeout and escalation |
|---|---|---|---|---|
| 1. Contact received | Intake team | Acknowledge with approved text; check for duplicates; route urgency | Contact details captured | Urgent: person within the stated response time; otherwise one business day |
| 2. Names collected | Intake team | Collect stage A fields; normalise names | All required names present | Reminder after two days; then state 10 |
| 3. Conflicts search | Conflicts analyst | Generate variants; run searches; assemble every hit | Every name searched successfully | Any error locks the inquiry and alerts the analyst |
| 4. Conflicts review | Conflicts analyst | Nothing beyond presenting hits | Every hit dispositioned; potential conflicts sent to state 5 or cleared | One business day, then supervising attorney alerted |
| 5. Conflict resolution | Licensed attorney | Assemble material for the attorney | Decision recorded with reasoning | Two business days, then practice head alerted |
| 6. Narrative collected | Assessing team | Send stage B link; record the account | Account and documents received | Reminder after three days; then state 10 |
| 7. Engagement decision | Licensed attorney | Assemble package; suggest template with reasons | Accept, decline or request more, recorded | Two business days, then practice head alerted |
| 8. Letter and signature | Licensed attorney approves; agent executes | Generate from approved terms; send approved package; track | Signed, or change requested and returned to state 7 | Reminder schedule; unsigned after stated period returns to attorney |
| 9. Engaged | Responsible attorney | Hand off to matter opening | Matter opening accepted | Handoff failure alerts the responsible attorney |
| 10. Closed without engagement | Licensed attorney approves | Send approved closing or non-engagement letter; end all sequences | Letter sent and record complete | None |
The numbers matter less than the principle: every state has an owner who is a role with a monitored queue, every state has an exit condition that can be observed, and every state has a timeout that does something.
Draw the failure paths as carefully as the happy path.
| Failure | Required behaviour |
|---|---|
| Conflicts system unavailable | Fail closed at state 3; alert the analyst; no narrative link issued |
| Duplicate inquiry | Detect by contact details and names; link records; one clearance applies to both or neither |
| Prospective client asks for a person at any state | Route to the intake team now, whatever the state |
| Timing mentioned at any state | Route to the urgent queue now, whatever the state |
| Reply arrives outside the watched channel | Intake team monitors every channel the workflow sends from |
| Handoff to matter opening rejected | Inquiry stays in state 9 with the responsible attorney alerted |
| Anything the map does not describe | Stop, hold the inquiry, alert the intake team |
The last row is the most important. An orchestration that meets a situation its designers did not foresee should stop and ask, not improvise.
Give each state only the tools it needs. The agent's permissions change with the state. In state 2 it can write stage A fields. In state 3 it can query the conflicts system. It never needs to read existing matter contents, and in stages before clearance it has no access to narrative storage. Least privilege and automation identities are the subject of Module 8; the map is where you first write down what each state requires.
Treat everything the public submits as content, never as instruction. Fields from a web form, chat messages and emailed attachments are recorded and presented to people. They never change the orchestration's state, choose a tool, or alter a recipient. State transitions come from the workflow's own rules and from recorded human decisions only.
Keep the map and the implementation in step. The map has a version and an owner. Any change to a state, a timeout or a permission changes the map first, and a licensed attorney who owns intake approves it.
Remember this
- Map intake to engagement as named states, each with an owning role, permitted agent actions, an observable exit condition and a timeout that escalates.
- Draw failure paths as carefully as the happy path, including conflicts outages, duplicates, requests for a person and mentions of timing at any state.
- When the orchestration meets anything the map does not describe, it stops, holds the inquiry and alerts a person.
- Give the agent only the tools each state needs, and never access to matter contents or narrative storage before clearance.
- Public submissions are content, never instruction; only the workflow's rules and recorded human decisions move an inquiry across a decision point.
Legal example
The fictional firm Marrowgate Keane LLP has redesigned its intake script, its conflicts sequencing and its engagement step, and the automation lead asks the intake manager, Delphine Asante, to prove the whole flow works before it goes live across three offices. She draws the orchestration as ten states and then spends most of a week on the failure paths. Three findings change the design. First, the conflicts analyst, Rory Callender, is the only owner of states 3 and 4, and he takes two weeks of leave every August. The map gains a named deputy and an escalation to the supervising partner, Imogen Faraday-Holt, after one business day. Second, the reminder emails are sent from a no-reply address that forwards replies to an unmonitored mailbox. In a test, a fictional prospective client replies "I have a hearing next Tuesday, please call" to a reminder, and the reply sits unread. The sending address is changed to a monitored intake mailbox, and replies are scanned for timing words and routed to the urgent queue whatever state the inquiry is in. Third, a test submission includes a line in the "other parties" field written to look like an instruction to mark the conflicts review as complete. The agent records it as a party name, which is the correct behaviour, but the review shows that nothing in the design prevented a field from influencing a state change. The orchestration is changed so that states 4, 5 and 7 can only be exited by a recorded decision from a named person, and the agent's tool for updating state is restricted to states 1, 2, 3, 6 and 8. Delphine also adds the catch-all row: anything the map does not describe holds the inquiry and alerts the intake team. In the first fortnight it fires twice, once for an inquiry submitted in a language the script does not cover and once for a prospective client who is also an existing client's employee. Both are handled by a person the same day.
Responsible-use warning
Never let an intake orchestration improvise when it meets a situation its map does not describe. The default behaviour of most workflow tools is to continue, and continuing is exactly how an inquiry reaches narrative collection without clearance or an urgent reply disappears. Stop, hold and alert. Equally, never allow content submitted by the public to move an inquiry across a decision point. Conflicts review, conflict resolution and engagement exit only on a recorded decision by a named person.
Human-review requirement
The licensed attorney who owns intake approves the orchestration map and every version of it. The reviewer confirms that every state has an owning role with a monitored queue and a deputy, that every exit condition is observable, that timeouts escalate to a named role, that failure paths cover outages, duplicates, requests for a person and mentions of timing at any state, that decision states exit only on recorded human decisions, that the agent holds only the tools each state needs, and that the catch-all stop is in place before go-live.
Practical exercise
Map a fictional intake orchestration end to end.
- Invent a firm with at least two offices and a practice area that receives urgent inquiries.
- Draw the orchestration as named states from first contact to engagement or closure. Aim for eight to twelve.
- For each state, write the owning role, its deputy, what the agent may do, the observable exit condition and the timeout with its escalation.
- Mark the states that may only be exited by a recorded decision from a named person, and explain why for each.
- Write the failure path for at least six failures, including a conflicts outage, a duplicate inquiry and a reply arriving on an unwatched channel.
- Write the catch-all rule for situations the map does not describe.
- List every channel the orchestration sends from and every channel a prospective client might reply to, and name who monitors each.
- For each state, list the tools the agent needs, and identify one tool the current design gives it that no state requires.
- Describe, at the level a defender needs, how you ensure that text submitted by the public can never change a state, choose a tool or alter a recipient.
- Give the map a version number, an owner and the attorney who approves changes.
Knowledge check
-
What four things must every state in an intake orchestration map define?
Answer guidance: An owning role with a monitored queue, what the agent may do in that state, an observable exit condition, and a timeout that triggers a stated escalation. Each state also needs a defined failure path.
-
What should the orchestration do when it meets a situation the map does not describe?
Answer guidance: Stop, hold the inquiry and alert a person. It should never improvise or continue by default, because continuing is how an inquiry reaches a later stage without the decisions that should precede it.
-
Which intake states should only be exited on a recorded human decision?
Answer guidance: The decision states: conflicts review, conflict resolution and the engagement decision. Their exits must come from a named person's recorded decision, never from an agent, a timeout or content in a submission.
-
How should an intake orchestration treat text submitted by members of the public?
Answer guidance: As content to record and present to people, never as instruction. Submitted text must not change a state, choose a tool or alter a recipient; transitions come only from the workflow's rules and recorded human decisions.
Sources and further reading
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications for 2026 (2025-12-09). Accessed 2026-09-14.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-14.
- American Bar Association — Formal Opinion 506: Responsibilities Regarding Nonlawyer Assistants (client intake) (2023-06-07). Accessed 2026-09-14.
- American Bar Association — Model Rules of Professional Conduct. Accessed 2026-09-14.
- Smith.ai — Smith.ai. Accessed 2026-09-14.