Most workflow diagrams show the sequence of work but skip a harder question: who, or what, should handle each step? The answer is rarely "AI" for the whole process. A required-field check, a messy description and a safety decision have different evidence needs and consequences.
The UK government's AI playbook tells teams to choose the right tool for the job and to plan for the resources needed to maintain it.[1] NIST's AI Risk Management Framework playbook goes further: context mapping should consider non-AI and non-technology alternatives, including mostly manual methods.[2] Those are useful design constraints. Start with the work, then choose the simplest handler that can do each step safely.
The MAJLS task-routing matrix below separates three handlers: deterministic rules, AI-assisted preparation and a named person. It also records the handoff. That last part matters because a sensible routing decision can still fail if the next owner receives a bare output with no evidence, deadline or exception reason.
Task-routing framework and matrix
Create one row for each meaningful workflow step. A row should be small enough that it has one input, one outcome and one accountable recipient.
Record these fields before choosing a handler:
- Step and input: What arrives, in what format, and from which approved source?
- Rule clarity: Can the correct outcome be expressed as a stable, testable condition?
- Evidence quality: Are required records present, current and consistent?
- Consequence: Could the step affect safety, money, access, rights or an external commitment?
- Reversibility: Can an incorrect action be undone cleanly, and by whom?
- Exception pattern: Which observable conditions make the normal route unsafe or incomplete?
- Handler: Rule, AI-assisted preparation or named person.
- Approval owner: Who authorizes any consequential action?
- Handoff: What output and evidence move next, to whom, by when, and on which escalation trigger?
Do not collapse the handler and approver into one field. An AI-assisted step may prepare a recommendation, but the named human process owner keeps consequential approval.
Route to a deterministic rule
Use a rule when the input is structured, the condition is explicit and the output can be tested without interpretation. Examples include checking that mandatory fields are present, matching a known location code or sending a complete case to a predefined internal queue.
A rule is not automatically safe because it is predictable. Record which policy or configuration defines it, what happens when data is missing and who owns changes to that rule. If a rule would create an irreversible or externally binding outcome, the matrix should still name a human approval point.
Route to AI-assisted preparation
Use AI assistance when the input is unstructured and a draft, summary or suggested category would save a person time. The output should remain inspectable. Pass the source material, extracted evidence and any unresolved ambiguity forward with the result.
This route fits preparation, not authority. AI may condense a long description or suggest likely categories. It should not approve spend, decide a safety response or make a supplier commitment. A model confidence score is not an escalation rule. Escalate on observable conditions such as missing records, conflicting facts, an out-of-scope category or a consequence that requires named authority.
Route to a named person
Use a person when the step requires judgment about consequences, policy interpretation, competing priorities or a commitment on behalf of the organization. Name the role rather than writing "human review." The facilities manager, budget owner or safety lead is accountable in a way that a generic review queue is not.
The person should receive a decision packet, not a mystery alert. Include the source description, checks already performed, suggested route, failed condition and exact decision required.
Filled illustrative example: facilities work-order intake
This example is illustrative. It does not describe a client deployment, operational result or measured saving.
| Step | Observable conditions | Handler | Output and handoff |
|---|---|---|---|
| Validate intake | Location ID, requester and problem description must be present | Deterministic rule | Return an incomplete form to the requester, or pass the complete record and validation log to preparation |
| Summarize description | Complete free-text description is available; original text remains attached | AI-assisted preparation | Produce a short summary with links to the source fields for the facilities coordinator |
| Suggest category | Description matches the approved category list; ambiguous or conflicting details are flagged | AI-assisted preparation | Suggest a category and show the phrases that support it; do not dispatch work |
| Apply known internal routing | Site and category map to a current, owner-approved queue | Deterministic rule | Move the case to that internal queue and record the rule version; unmatched cases stop |
| Decide safety response | Injury, exposed wiring, fire, structural damage or another defined hazard is reported | Named facilities manager or safety lead | Review the original report and authorize the response under the organization's safety process |
| Approve spend or supplier dispatch | Work may incur cost or create an external commitment | Named budget owner or facilities manager | Approve, reject or request evidence; only the named person can commit the organization |
| Resolve ambiguity | Records conflict, no category fits, the site is unknown or required evidence is missing | Named facilities coordinator | Correct the record, request information or assign an accountable owner |
The handoff for the safety row could read: "Send the original description, location, requester details, validation log and highlighted hazard terms to the duty facilities manager immediately. Escalate if the manager does not acknowledge within the response time set by the safety policy." The time limit comes from the organization's own policy; the matrix should not invent one.
Review the matrix before selecting tools
Read the rows left to right and look for weak boundaries:
- A step marked as a rule still depends on interpretation.
- An AI-assisted output loses its source evidence before review.
- A consequential step has no named approval owner.
- An escalation depends only on model confidence rather than a failed check or authority limit.
- A handoff names a queue but not the person responsible for clearing it.
- A connected system cannot confirm the change or support reversal.
Fix those gaps in the operating design first. Then decide whether existing workflow software, a narrow AI service or a new operational system is justified. The matrix is finished when every step has a suitable handler, every consequential action has a named human approver, and every exception has an observable route to an accountable owner.