AI literacy

A finance leader asks their AI whether a customer has renewed. “Yes, for another year.” The account record, however, says the renewal is still under discussion. Both sound plausible. Which is right?

Consider the source: Perhaps the AI summarized an old account. Perhaps it retrieved a new message that the account team has not recorded. Perhaps it inferred an agreement from language that expressed an intention.

Knowing how to navigate this uncertainty—how to reliably interpret what the AI is plainly telling you—is a practical test of AI literacy. Do you know what the system did, examine the evidence it used, and when to trust its output?

Follow the answer back to its source

“AI says” says nothing at all.

An AI might draft from the material supplied in a conversation, retrieve documents, query a business system, or combine multiple steps. The relevant question is what happened in this instance. (Can you see its reasoning? Can you probe its reasoning? How much can you trust its responses?)

In this example, you might ask to see the underlying record and its date. Check whether the text describes a signed agreement, a customer intention, or an account manager’s expectation. Note if the AI is using a tool call, pulling a response from training and/or tuing data, or merely reflecting what you said.

A source link helps… but how did that link get there? A hallucinated link that looks valid is not evidence of a renewal unless followed to the system of record.

Separate assistance from authority

A model’s explanation of its answer can suggest where to look. It should not be accepted as independent proof that the answer is correct. For consequential work, check the record itself and the process by which that record becomes authoritative.

This distinction also improves the task you give the system. “Tell me whether the customer renewed” asks for a conclusion. “List the dated evidence for and against renewal, distinguish signed commitments from expectations, and show what is missing” produces a more inspectable assignment. A person still needs to verify the result.

The same habit applies when comparing two products. Before choosing the answer that reads better, compare their inputs, source access, dates, and instructions. If those differ, the products may be answering different versions of the question.

Make the practice easy to repeat

An executive does not need to memorize every model or technical term. They do need enough understanding to establish a workable standard for their team. For an AI-assisted claim that affects a decision, ask for the source, the relevant date, the unresolved uncertainty, and the person who checked it. Apply the standard in proportion to the consequences; a private brainstorm needs less scrutiny than a board figure.

In the hypothetical forecast review, the strongest contribution may be a sentence that keeps uncertainty visible: “The customer has expressed an intention to renew; the signed agreement is still outstanding.” The executive can work with that distinction. Confidence without it is harder to use.