Diagnosing adoption opportunity and work context · Lesson 2

Map workflow and human boundaries

6 min read

Lesson objectives

  • Map an AI-assisted decision flow, including handoffs, exceptions, human authority, and failure checks.
Reading tools

“Will staff use the assistant?” is too broad a design question. Ask where its output enters the workflow, who sees it, and which decision it could influence. A useful workflow map follows one case from trigger to correct closure; it is more than a diagram of departments. Trace an ordinary request and an exception. Record the triggering event, the inputs available to each role, each step’s output, the waiting point, and the receiving role. If the map represents the assistant as a person who decides, it has hidden human authority rather than designed it. Name the assistant’s function precisely: retrieve a passage, draft wording, propose an initial category, or flag missing information. Each function has a different failure mode and a different effect on time.

At every handoff mark four things: permitted data, source of truth, editing authority, and escalation condition. In a customer complaint, the assistant might retrieve a reply rule from the current manual. A cited passage is not automatically applicable. The tool may confuse an exception with a general rule or retrieve an expired version. The employee who sends the reply must check the original passage, its date, and the conditions of the case. If the complaint concerns a discretionary reimbursement, a person with delegated authority must decide it; the assistant cannot create a financial promise. This is a human boundary: a system suggestion does not become an organizational commitment until an informed person with authority acts.

Do not put an unexplained “human review” box on the map. Specify what the reviewer checks, how much time they have, and which evidence they see. Requiring one employee to approve 150 replies in an hour could turn review into a label rather than a real control. A signature is also weak if the original source is hidden or approval is effortless while challenge is cumbersome. A practical review screen could show the passage supporting the suggestion, let the employee edit or stop a reply, and record objections without punishing the reviewer for raising them. The precise interface and review intensity should follow the seriousness and reversibility of a possible error; this is a design choice, not a rule for every AI use.

Measure the whole path, not the model’s response time. Before the pilot, intake takes two minutes, policy search six, drafting four, and review two; 20% of cases also wait 18 hours for approval. Assistance may cut search to two minutes and drafting to two, while raising review to four. Ordinary active work then takes 2+2+2+4=10 minutes instead of 2+6+4+2=14, a saving of four minutes. The 18-hour wait remains in cases that require approval. Reporting a 29% reduction in resolution time for all cases from active minutes alone would mislead. Segment the cases and check whether appropriate escalation changes. A decline in escalation caused by overconfidence might make the time metric look better while increasing false commitments.

Before launch, ask at least two affected workers where actual practice differs from written procedure. One may use an exception table absent from the manual; another may reopen cases after a second team closes them. That finding does not automatically validate an undocumented practice. It reveals a dependency that must be understood before removal. Record where evidence is held, who can correct a policy, and how error reports reach that owner. The MAP function of the NIST AI RMF emphasizes purpose, context, and affected parties, while governance calls for clear responsibilities. Our map applies those ideas; it is not a mandatory NIST process chart. A strong result is a document that service and risk teams can test against one ordinary and one exceptional case.

Stress-test the map before measuring speed. Give a reviewer a case with a correct transaction number and a misleading “subscription” word, then make the latest policy temporarily unavailable as a search failure might do. Which step prevents a false reply? If the employee cannot see the billing record or open the source passage, the assigned review is impossible however diligent they are. If the source is visible but its version is unclear, the failure lies partly in document control, not merely drafting skill. Test again after redesign: show the policy date, prevent a final reply when transaction type is missing, and provide a route to the authorized manager. Record three separate outcomes: whether the conflict was detected, whether it reached the right decision maker, and how long correct closure took. This small stress test says more than “human in the loop” because it states what the human must do and what information they need. Also test the opposite case: a valid recurring subscription should not wait for unnecessary financial escalation. Excess control can create delay and erode adoption.

Lesson summary

Map an AI-assisted decision flow, including handoffs, exceptions, human authority, and failure checks.