A workflow does not need one AI strategy. It needs mode decisions.
The lazy version of AI adoption asks whether a task should be automated. That creates a false binary: manual or automated. Real work is messier. The same workflow often needs several modes depending on risk, confidence, complexity, customer impact, and reversibility.
Four modes are enough for most operators to start: assist, review, exception, automate.
Assist means AI helps a human do the work, but the human remains the primary actor. This is the safest starting point for messy judgment work. AI can summarize the case, retrieve context, draft options, check for missing fields, or propose next steps. The human still frames the problem and owns the output.
Assist works well when context is fragmented, quality requires judgment, or the work is high-trust. A customer escalation, executive update, product decision memo, legal negotiation, or sensitive people issue usually belongs here. AI can reduce friction, but pretending the work is ready for automation will create risk.
The design question for assist mode is: what should AI remove from the human's cognitive load without taking over judgment?
Review means AI checks work before or after a human does it. This mode is underrated. AI can catch missing evidence, policy mismatches, tone problems, calculation errors, stale assumptions, incomplete records, or deviations from a playbook.
Review mode is useful when humans still produce the work but consistency matters. Think support replies, implementation plans, sales notes, renewal risks, forecast comments, code reviews, contract redlines, and operating-review packets.
The design question for review mode is: what errors should never reach the next step unnoticed?
Exception means the system handles the normal flow and humans handle the cases that do not fit. This is where AI can create serious leverage, but only if exception rules are explicit. Without them, exception mode becomes a dumping ground for weird cases and low-confidence outputs.
Exception mode works when the happy path is stable and the risk thresholds are clear. A standard billing question, routine support classification, low-risk data cleanup, or known policy answer may fit. High-value customers, legal ambiguity, emotional tone, missing data, unusual requests, or low confidence should route to a human.
The design question for exception mode is: what conditions require human judgment before the work continues?
Automate means the system completes the work without human review in the normal case. This mode should be earned. It is appropriate when inputs are clean, outcomes are reversible or low risk, quality can be measured, and failure paths are visible.
Automation is not a moral victory. It is simply one mode. A company that automates too early often spends the savings on cleanup, trust repair, and manager supervision.
The design question for automate mode is: can the system complete this work safely, observably, and with a clear rollback or escalation path?
The same workflow may use all four modes.
A support workflow might use AI to assist the agent with context, review the proposed answer for policy issues, route normal password-reset cases through automation, and escalate angry enterprise customers as exceptions.
A finance workflow might use AI to assist with variance narratives, review formulas and source references, automate low-risk reconciliations, and escalate unusual movements above a threshold.
A sales workflow might use AI to assist with account research, review CRM hygiene, automate enrichment, and route strategic account changes to a manager.
The mistake is assigning one mode to the whole process.
Mode decisions should happen at the step level. For each step, ask:
- What is the risk if this is wrong?
- How reversible is the action?
- How much judgment is required?
- How clean is the input?
- How measurable is quality?
- Who owns the result?
- What exception path exists?
This gives managers a practical language. Instead of arguing about whether AI should "do" a job, they can redesign the workflow step by step.
The point is not to maximize automation. The point is to put the right work in the right mode.
That is where trust comes from. People trust AI systems when the design matches the risk of the work. They distrust them when everything is treated like either magic or danger.
Assist, review, exception, automate. Four modes. Use them deliberately.
This is part 4 of 10 in Work Design for the AI Era.
