Operating Brief
| Item | This lesson |
|---|---|
| Decision | Decide how the workflow will move from experiment to normal work without transferring hidden burden to frontline staff. |
| Output | An adoption and transition plan. |
Workflow Problem
An AI pilot can improve a model metric while making the team’s work worse. Review queues may grow, responsibilities may become unclear, corrections may disappear, and staff may distrust the system because they were shown a finished design rather than involved in defining it.
Adoption is not a communication campaign after deployment. It is part of workflow design.
Identify the Role Changes
For every affected role, state:
- what work is removed;
- what new work is added;
- which decisions remain theirs;
- which evidence they will see;
- how they correct the system;
- what happens when they disagree;
- how performance will be measured fairly.
A system that “saves time” by adding invisible review and exception work has not proven value.
Use Shadow Mode
Before automated routing or action, run the system beside the existing process:
- The normal workflow remains authoritative.
- AI produces its proposed output.
- Staff compare the proposal with the actual decision.
- Corrections are categorized.
- Review volume, time, and confusion are measured.
- Only selected low-risk cases move to assisted mode after a gate review.
Shadow mode separates learning from operational risk.
Train Around Decisions, Not Features
Training should answer:
- When should I accept the suggestion?
- What evidence must I check?
- When must I reject or escalate?
- How do I record a correction?
- Who can change labels, rules, sources, or thresholds?
- Who may pause the workflow?
A product tour is not operating training.
Measure Adoption Honestly
Useful measures include:
- acceptance, edit, rejection, and escalation rates;
- reviewer time per item;
- queue or backlog effects;
- repeated correction categories;
- staff confidence by risk tier;
- use of unsupported workarounds;
- outcomes by team, shift, language, or case type.
Low usage may indicate weak value, poor training, or rational resistance to bad design.
Practice: Write the Transition Plan
| Area | Required decision |
|---|---|
| Stakeholders | Who defines labels, controls, metrics, and exceptions? |
| Shadow mode | Duration, sample size, baseline, and exit criteria |
| Training | Decisions, evidence, corrections, and escalation |
| Capacity | Expected review volume and service level |
| Feedback | How corrections reach the owner and change process |
| Incentives | How staff are measured while the workflow changes |
| Gate | Conditions for assist mode, partial automation, expansion, or rollback |
Include one explicit promise to affected staff: what the pilot will not automate or use for performance management without a separate decision.