AI assistants: can you really undo what they do?
In SHR — Swiss Human Resources’ programme “Creating your own AI assistants and agents without coding”, one management question deserves attention before an assistant is connected to workplace tools. Can the previous state be restored after an incorrect action, even when its effects have already crossed team boundaries?
What the literature establishes
Ben Shneiderman (1983, Computer) presents direct manipulation as an approach based, among other things, on incremental, reversible actions with immediately visible results. This classic design article does not address AI agents, but it provides a useful criterion: making action easier is not enough if users cannot see its effects or undo it.
The distance between intention and effects
Edwin Hutchins, James Hollan and Donald Norman (1985, Human–Computer Interaction) analyse the distances users must bridge to translate intentions into actions and interpret the results. Applied to an assistant, this framework suggests distinguishing a simple instruction from a genuinely intelligible operation: “update the shipment” may conceal several changes across different tools.
The premature conclusion
One might conclude that an “Undo” button solves the problem. Yet restoring a field does not retract a message that has already been read, nor does it necessarily stop a process triggered at a partner organisation. Reversibility must therefore be assessed at the level of consequences, not just the interface. (our executive and employee training programmes)
What these studies do not demonstrate
These articles offer design frameworks; they do not measure the effectiveness of undo functions in current no-code agent platforms. An execution history makes an action traceable, but does not prove that it is reversible. What can be restored, what requires a compensating action and what remains irreversible must be checked separately.
A practical check in Bâle
In Basel, a pharmaceutical or life sciences team could test, in an isolated environment, an assistant coordinating a fictional sample shipment request between international R&D and Rhine logistics. Within the SHR programme, the exercise would involve introducing an incorrect change and then attempting to restore the initial state of every affected system. The measure would be the proportion of cases fully restored, verified by comparing before-and-after states, with recovery time and effects that cannot be undone recorded separately. A correction in the originating tool would not count as complete restoration if a notification or instruction remained active elsewhere. To go further: explore the Building AI Assistants and Agents Without Coding training in Basel in the canton of Basel-Stadt, or browse our executive and employee training programmes in Switzerland.
In pictures: Building AI Assistants and Agents Without Coding in Basel
- building ai assistants and agents without coding
- Basel
- research
- training basel