Most recognise the danger of AI. It delivers a slick, formatted persuasive response to a prompt. It has become way too easy to just accept the veracity of the response and move from the drivers seat to the back seat. This removes yourself from the sweat of doing the work, and the responsibility for the outcomes.
We see it all the time, often in spots where we accept that there has been a level of scrutiny that should deliver reliable outcomes. For example, the crapola pie Deloitte delivered to the government in June 2025, which then published the deeply flawed AI generated report without any review.
The answer is in the transparency of the process of assembling, analysing, and preparing the output. What is included, what has been ‘AI imagined’ to fill the gaps in the prompting and resulting workflow, and easiest to miss, what has been left out.
The challenge is rapidly compounding as we move into the ‘AI Agent’ world, where we expect a whole multi-step process to be executed on our behalf by a machine.
Following a few sensible steps can dramatically improve the quality of the output.
Set hard boundaries.
- Define in explicit terms what the process will and will not do. Is the workflow restricted to your own files, or can it go outside?
- If it is instructed to use outside sources, what are the boundaries?
- Explicitly instruct that there be no generation of conclusions before a human review of the sources and for/against arguments.
- Insert a series of ‘stop’ points beyond which the tool will not proceed until instructed to do so. Instruct the tool to act as a devil’s advocate at each stop/go point.
Transparent provenance.
AI tools extract information from the sources it finds or are directed to. Those sources define the potential of the output to deliver useful value of some sort to the user. Curating the sources the tool examines is therefore a fundamental step in the generation of that value.
Remain curious.
Just because a tool can repeat a workflow accurately every time does not mean that the workflow is perfect. It takes human curiosity and experimentation to test and retest a process that is designed to deliver an optimised outcome. The AI cannot do that optimising; it requires a curious human to be in the drivers seat asking that key ‘what if’ question. So, turn off the process from time to time, and go back to the old way, manual execution.
An exercise I did many times pre-AI to improve a record-keeping process was to imagine myself as a paperclip, attached to relevant documentation. I would follow the document through the process, documenting every point at which the document was delayed, added to, moved, and authorised, and the time lapse of each of those points. Map it out, and inevitably you will see improvement opportunities. AI cannot see those opportunities.
These steps will stop the tool trying to please you by being agreeable and synthesising conclusions.
