Giving generative AI a useful first job.
The arrival of ChatGPT at the end of November 2022 gives the familiar January question a particularly visible candidate. A system that can answer questions and draft text in conversation is easy to try and easy to imagine inside everyday work.
I would approach 2023 with curiosity and a specific task. The useful question is where this technology can help a person do a better job, and how we will recognise an answer that needs correcting.
Choose work you can review
Drafting an internal summary from approved notes is a more manageable first experiment than letting a model answer customers independently. A reviewer can compare the result with the source and identify omissions, invented details or an unsuitable tone.
OpenAI's introduction describes limitations, including plausible answers that can be incorrect. A fluent sentence therefore needs the same scrutiny as any other unverified statement. Give the reviewer responsibility for the final output and enough time to check it.
Start with information the team is allowed to use in the selected tool. Agree that boundary before the first enthusiastic experiment spreads customer records into an unapproved service.
Define what a good result looks like
Take a small set of representative tasks and keep the original material. Compare the assisted output with the team's existing approach. Record factual corrections, missing details, time spent reviewing and whether the finished result is useful.
Include awkward examples. Notes may contradict each other. A request may lack essential context. Sometimes the right response is to ask a question or decline to draw a conclusion. Those cases will teach the team more than a polished demonstration.
Give the experiment an owner
Someone needs to maintain the instructions, review failures and decide whether the process is ready for wider use. That person also needs a route to the people responsible for data, security and the business outcome.
The leadership lesson from 2019 remains relevant. An organisation learns when it gives people room to test an idea and expects honest evidence in return. It learns less when every experiment must be presented as a success.
For 2023, I would give generative AI one useful first job. Keep the scope small enough to understand the mistakes, then let the results determine the next step.
Context: OpenAI's November 2022 introduction of ChatGPT.
The annual outlook series
2019 · 2020 · 2021 · 2022 · 2023 · 2024 · 2025 · 2026
Annual series edition: 5 January. This edition was prepared or revised on 8 September 2026.
