AI strategy01
Before you buy another AI tool, map one day of work.
A practical way to find a useful AI starting point: follow one task, identify the friction, and define a result worth measuring.
The journal / Practical AI / Note 03
Practical AI · Field note 03A draft is not a decision.
A well-written answer can feel finished before anyone has checked it. That is where a useful assistant can turn into an expensive distraction.
Decide what kind of draft you need
A summary, an argument, and a recommendation are different assignments. Tell the assistant which one you want, and give the human reviewer the same brief. Otherwise, a response can drift from organizing information into making a choice that no one asked it to make.
For an internal meeting summary, the job might be to separate decisions, open questions, and proposed next steps. A suggested action should not appear as an agreed commitment. The distinction matters more than whether the sentences sound elegant.
Check the statements that can be wrong
Read once for claims. Names, dates, amounts, links, quotations, responsibilities, and deadlines deserve direct checking against the approved source. If the source does not support a statement, remove it, qualify it, or find the missing evidence.
The Canadian Centre for Cyber Security identifies inaccurate or false output as a risk of generative AI. Fluency is therefore a poor substitute for verification. A confident tone does not tell you whether an answer has correctly represented the material you supplied.
Restore the context
A technically correct sentence can still mislead when it leaves out an important condition. “The team can deliver on Friday” means something different from “The team can deliver on Friday if approval arrives on Tuesday.” Look for conditions, uncertainty, and disagreements that the draft may have compressed away.
Then check the audience. An update for a colleague who attended the meeting can assume background knowledge. A message to a new stakeholder cannot. Good editing includes deciding what the reader needs in order to interpret a statement properly.
Make responsibility visible
Assign a reviewer who understands the subject and has the authority to approve the message. If the draft concerns an area where the reviewer lacks expertise, get the appropriate person involved. The assistant should not be the only bridge between an uncertain author and a consequential claim.
A useful review record can be as simple as an approved draft with a date and an owner. The point is to know which version was checked and who is responsible for the final communication.
Keep the learning
When the same kind of correction keeps appearing, change the task instructions or narrow the use case. Repeatedly fixing missing qualifications may mean the brief needs an explicit section for conditions and unresolved questions.
Judge the workflow by the reliability of the final result and the effort required to reach it. Generating a draft is one stage of that workflow. The decision to use it belongs to a person.
One thing to take away
Review claims, context and ownership separately. A polished draft still needs an accountable editor.
Further reading
Canadian Centre for Cyber Security — Generative artificial intelligence ↗
Background on generative AI risks and mitigation considerations.
Pasqualino Mazza
Entrepreneur, global AI consultant, and content creator. Writing about AI, business, and leadership from a Toronto perspective.