Before your team trusts an AI answer, ask these five questions

If you could ask your organisation’s data one question on Monday morning, what would it be?

Why are customer enquiries taking longer to resolve? Which projects keep running over budget? Where is your team spending hours on work that could be simplified?

Those questions offer a useful way to look at the latest AI announcements.

On 10 September 2026, OpenAI announced a Data agent in ChatGPT Work. According to the company, it can connect to business data, investigate changes and produce interactive dashboards through a conversation. These are vendor-described capabilities, rather than an independent assessment of how it will perform in your organisation. Read OpenAI’s announcement.

For business leaders, easier access to analysis raises a practical question: what needs to happen before an AI answer deserves to influence a decision?

1. What decision are we trying to make?

“Let’s use AI in the business” is an ambition. It is difficult to turn into a useful task.

“Let’s understand why our customer response times increased last month” is a starting point. It identifies a problem, a period to examine and a result that someone can check.

Before choosing a new tool, write down one decision your team regularly makes. Then identify the information needed to make it well.

A customer service manager might need to decide where to allocate staff next week. Useful evidence could include enquiry volumes, the types of requests coming in and the time taken to resolve them.

An attractive dashboard is only useful if it helps that manager make a better decision.

2. Are we working from reliable information?

Consider a simple question: “How many customers did we gain last month?”

Does “customer” mean someone who signed a contract, paid an invoice or started using the service? Are cancellations included? Are different teams counting the same thing?

A conversational interface makes a question easier to ask. It does not settle those definitions.

Before using AI to analyse a business problem, agree:

• Which source is authoritative.
• What the important terms mean.
• Whether the information is current and sufficiently complete.
• Who can explain inconsistencies.

This work may sound ordinary. It is also what makes an answer worth trusting.

For a broader view of readiness, see my 5 P’s framework for AI strategy, which connects people, process, platforms and proprietary data.

3. How will we check the answer?

Suppose an AI tool reports that customer response times rose because enquiry volumes increased.

That may be a reasonable explanation. But did staffing change? Was there an outage? Did a new product generate more complicated questions?

A useful review would check the underlying figures, look for missing context and distinguish an observed pattern from a possible explanation.

For an initial trial, choose a question where someone on the team can independently check the result. Ask the tool to identify its sources, explain its calculations and state what information is missing. Check those sources and calculations against the original material.

The person reviewing the answer needs enough subject knowledge to challenge it. Fluent language is not evidence that the conclusion is correct.

4. Does our team know how to challenge it?

Access to a tool is the beginning of adoption. People also need to learn how to frame a question, recognise an incomplete answer and decide when to ask a colleague for help.

A practical team exercise could start with an approved dataset and a familiar business question. Participants could compare their approaches, review the results together and discuss where the tool helped or created extra work.

The aim should be specific: by the end of the exercise, people can complete and check a task they understand.

That is a more useful training outcome than simply having seen a demonstration. Give the review a named owner, and make clear who is responsible for the eventual decision.

I explore the distinction between decision support and handing over judgement in Who Decided? AI decision accountability.

5. Has AI improved the whole task?

Choose one recurring task with a clear owner. Record how it is completed today, including the time spent checking and correcting the result.

Then trial an AI-assisted approach using information your organisation has approved for that purpose.

Compare the two approaches on:

• Accuracy: Were the figures and conclusions correct?
• Total effort: How much time did preparation, analysis and review take?
• Usefulness: Did the result help someone make a decision?
• Repeatability: Could another colleague follow the same process?

Include the checking time. A fast first answer can still lead to a slow overall task.

A trial that reveals limitations is useful too. It helps the team decide where AI belongs in the process and where human work remains essential.

When a trial is ready to grow, my two-track enterprise AI operating model explores how employee adoption and specialist AI work can develop together.

The question to take into Monday

The latest announcement offers another opportunity to examine how we work.

The most useful response is to choose a question that matters, establish the evidence needed to answer it and give someone responsibility for checking the result.

Start with this: Which decision could our team make better if we had a clearer answer—and how would we know the answer was right?

That gives AI adoption a concrete purpose.

Bring this conversation to your leadership team

Planning a conference, leadership offsite or company event? Explore Mark Kelly’s AI keynotes for business leaders to bring practical questions about AI, people and decision-making to your audience.

Check Mark’s availability for your event. Share your audience, preferred date or timeframe, and what you want attendees to take away.