AI adoption: stop treating it as an experiment
AI adoption is a leadership decision. A subscription for everyone won't do it: settle data, access and review first, then run a focused trial on one workflow.
AI adoption is a leadership decision, not an experiment you leave to whoever happens to be curious. GPT-4 has been out since March, and right now I see two typical reactions. Some companies block ChatGPT, others buy a subscription for everyone and wait to see what happens. I think both are wrong. Blocking means using AI too little. Handing out subscriptions means trusting it too much while barely scratching the surface. This year I started moving SUMAX onto AI together with the team, and I'm at the very beginning. But my order is already set: first the foundation for everyone, meaning users, permissions and access, then tools for specific tasks.
A subscription for everyone isn't adoption.
With ChatGPT anyone in a company can try out what a language model can do. That's a fine way to start. But a subscription settles none of what matters in daily operations: which data may go in, who checks a result and which task gets better. Everyone tries things on their own, and the know-how stays with whoever happens to be good at it.
A ban doesn't help either. People stay curious, and the usage easily moves to private accounts where nobody can see what data goes in. Treat AI as a change to how work gets done rather than a software purchase, and set the frame before anyone gets going.
Customer data stays out of any tool nobody has vetted.
With public AI tools it's often unclear what happens to the data you type in and who gets to see it later. As long as that's open, customer data and personal data don't belong there. Security has to be there from the start, because once habits set in, they're hard to walk back.
That's why the foundation comes first. Who uses which tool, with which permissions, through which access? It sounds like admin work. Without those answers, though, no leadership team can sign off on AI with a clear conscience. Decide which data may go into which tool before you hand out the first subscription.
Without a written-down workflow there's no yardstick.
A trial starts with a task whose result you can check. An impressive demo shows what a model can do. It doesn't show whether it helps in everyday work. For that you need to know how the work runs today and what a better result would look like.
At our company a lot of that lives in individual heads and scattered spreadsheets. As long as it stays there, nobody can say whether a trial beats the old way. So for me, AI transformation starts with unglamorous work: process knowledge belongs in one place where people, and later tools, can find it. Write the workflow down before you compare it with AI.
If you pass on an AI result, you own it.
Language models write fluently even when they're wrong. When they're missing a fact, they tend to make one up. That's called hallucination. Still, I see the bigger risk on the human side. Someone types in a request, gets a result and forwards it without reading it. People like that turn a useful tool into a liability. So reading and checking start with the first trial.
Only someone who knows the subject can judge what a result is worth. You learn that best on real tasks from your own work. That's why the shift includes new role descriptions and training on those tasks, and why tools get built with the people who use them. AI is there to take load off people. Put time for practice and review in the plan, or learning only happens on the side.
Ending a trial is a good outcome.
A trial ends with three options: expand it, rework it or end it. Leadership decides based on criteria set before the first run. A trial that shows a task isn't worth it has done its job.
A hypothetical example: a trial is supposed to prepare tasks from meeting notes. Before the first run it's settled who reviews the results, which errors are unacceptable and how follow-up questions get handled. After the trial, the decision becomes a comparison. Does the reviewer have less rework, or do the same errors keep coming back?
Without these settings you can't justify anything at the end, even if the output looks good. It stays a demo. Set the criteria for expanding, reworking and ending before the trial starts.
Leadership decides based on criteria set before the first run.
Five questions before the first trial.
- Which task should get better, and how will we recognize a better result?
- How does this task run today, and is that written down anywhere outside someone's head?
- Which data may go into the tool, and which never?
- Who checks every result before it moves on?
- What will we base the final call on: expand, rework or end?
If one answer stays blank, the trial isn't ready.
Waiting is a decision too.
A lot of leadership teams want to wait until the market sorts itself out. I get it, the models are changing every few months right now. But every model needs the same groundwork: settled data, permissions and access, written-down workflows and people who can judge a result. You can start that work today. It stays right whichever tool wins in the end.
AI belongs on leadership's desk before it spreads through the company on its own. Pick a workflow that depends on a single person, write it down and turn it into your first limited trial. How the next steps fit together is on the AI roadmap page.