Keeping a manual path has long-term value

What it is

Once automation works, products often hide or delete the old manual entry in the name of simplicity. Keep a manual path treats that entry as infrastructure for skill and for faults: people who need to practise need a place to practise; when the machine is down there must be a road that does not go through the agent to put the world right.

A UI whose only move is “let the AI try again” has no second causal chain when the agent fails.

Why it happens

A manual path does three jobs. A practice ground: under low load a person can walk it, feedback still exists, procedural knowledge does not fall all the way off. A fault bypass: when the agent is down, the business need not stop at “wait for the model.” A landing for an adjustable level: when a person dials back to manual, manual is really there, not a greyed button. If Bainbridge’s irony is taken as a constraint, design has to leave a road a living person can walk besides automation, not design the person as someone who can only wait.

The path has to be real: permissions, data, tools still there, steps maintained. Fishing an expired handbook from backup on the day of the fault is not a path.

Studying it

Compare: delete the manual entry after automation versus keep it and periodically force a manual walk versus keep it but never force. Later measure: completion under fault, manual scores, time to find the entry. Independent variables: visibility of the entry, frequency of forced practice, whether the handbook is in sync with the current system.

Fault drills must cut the agent. An entry that still exists but whose permissions have expired and whose fields no longer match counts as no path.

Where it stops holding

A brand-new capability users never had (they could not do it before) has nothing to keep; a degraded path has to be built from zero. Production acts safety forbids a hand from touching may have a manual path only for a controlled repair role, not as every operator’s button. Whether skill drops, and when the drop is seen, are the previous two claims; this is the countermeasure: the road stays, and it can be walked.

Applying it

• For every core task the agent substitutes, keep a completion path that does not go through the agent, in a place operators can find — not only on an internal wiki.
• Walk the path with product revisions. Fields, permissions, external interfaces are the same set the agent uses, so the bypass is not a system from another decade.
• Under low load, offer “I will do this one myself” without punishing throughput scores.
• Check: cut the agent service and see whether the business can finish along the manual path inside the time limit. If not, the path is fake. Then time three operators who have not walked it in three months finding the entry — if they cannot, it has been deleted in effect.

Related

• Same group: L4.09.1 Long-term reliance on automation weakens human skill · L4.09.2 Degradation only shows when takeover is needed
• Nearby: L4.01 Levels of Automation · L1.06 Graceful Degradation of AI Failure · L4.04 Takeover and Handoff Design
• Search terms: manual fallback · graceful degradation · adaptable automation