We bought AI. What now?

“I don’t use AI” explains very little. One person cannot see a use for it in their tasks. Another does not know which data they may share. A third has tested the tool and found no benefit once corrections are included. Each needs a different response.

I would avoid permanent labels of enthusiasts and opponents. Someone may happily draft presentations with AI while reasonably questioning its use in a process requiring high accuracy. A concern can reveal something that needs fixing before wider adoption.

If a team has no time to experiment, its manager should agree the priority and scope of a trial. If confidentiality is the obstacle, a decision on permitted working practices is needed. Encouraging people to be “more open-minded” cannot replace either decision.

The board should define the intended outcome and investment boundaries. The business process owner establishes what should improve, which results will be sufficient and how the team’s work will change. IT is responsible for its agreed technical scope: the environment, access, integrations and ongoing operation. Data and risk rules require the appropriate people, including security and other functions depending on the use case.

The team manager must also provide time to learn, agree how outputs will be checked and enable people to report problems. Employees should not have to decide for themselves whether they may share a customer document or who will be accountable for an incorrect answer.

This division of responsibilities is consistent with the voluntary NIST AI RMF 1.0: leadership is accountable for risk decisions, while the organisation establishes the context, assessment and management actions. It is a reference for organising work, not a certificate of compliance or evidence of an investment return.

I discuss the reciprocal responsibilities of the business and IT in 7 changes in the relationship between IT and the business. With AI, that partnership becomes very tangible: someone must accept the output and decide whether it is fit for use.

I would start with a bounded use case that has an available baseline. For instance, drafting a response to a recurring enquiry using approved reference material. This illustrates how to plan a trial; it does not describe an implementation I carried out.

Before starting, agree who will check the response, which errors are unacceptable and what happens when the tool cannot help. In this example, the response only reaches its recipient after verification by an authorised person. The trial should include harder cases rather than only those that look good in a demonstration.

I would record the tool version, dates, task scope and method of selecting cases. If results improve after changing instructions or source material, we need to be able to reconstruct what was compared. I also cover practical starting principles in AI in your company without chaos.

I am interested in the time for the complete task. Preparing a response includes collecting data, checking the output, making corrections and handling exceptions. A few seconds of generation may look impressive, but the complete workflow shows its effect on work.

Quality needs an equally precise definition. If we measure the proportion of responses needing correction, we record that number alongside the total number of cases assessed. Serious errors are recorded separately. An average should not conceal a situation that could harm a customer or the company.

A before-and-after comparison is meaningful with comparable task types, quality requirements and working conditions. In a small trial, it is useful to show individual observations and the spread of results. Where comparability is limited, the conclusion should remain within what was actually tested.

Costs include licences, integration, data preparation, training, verification and ongoing operation. I separate cash expenditure from the use of existing staff time to avoid counting the same item twice. Time released creates capacity for other work; it becomes a cash saving only when a mechanism for reducing expenditure is identified. I develop this distinction in the article and calculator on AI implementation ROI.

I would agree decision criteria before the trial. Expansion needs evidence that outputs meet the agreed requirements, people can work in the new way, and costs and risks are acceptable to those accountable for the decision.

If improvement looks promising but the sample is small or unrepresentative, a further test may be appropriate. If verification consumes the entire benefit or errors cannot be reduced to an acceptable level, it is worth stopping that use case. This need not rule out other uses of AI.

Without an agreed owner, quality measure or rules for using data, the decision lacks a sufficient foundation. Team enthusiasm is valuable, but it cannot fill those gaps.

A successful trial does not end a leader’s responsibility. Current instructions, user support, exception handling and periodic review of the outcome are needed. A change to the tool, data or process may change the result. NIST’s MEASURE function addresses assessment before deployment and monitoring during operation; the 2024 NIST AI 600-1 Generative AI Profile develops this for generative AI.

The costs of continued use should enter routine IT planning. The IT budget model: plan, actuals and forecast helps structure this side of the decision. The budget model itself does not measure AI effectiveness: that still requires process data.

Returning to the board’s question, a useful answer should show what changed in the selected process, the conditions in which it was tested and who owns the ongoing outcome. A discussion about further licences then has a concrete foundation.

I have prepared an AI use case assessment record. It helps document the problem, responsibilities, baseline, data handling rules, costs and decision rationale. It includes a blank template and a hypothetical worked example in Polish and English. It does not calculate an automatic readiness score or replace a use-case-specific risk assessment.

You can complete the HTML version locally in a browser, save a separate file containing your answers or print it. The Markdown version can be edited further in your own documentation environment. The resource is free and does not require an email address.

This is my proposed approach for IT leaders and process owners. The opening scene and use case example are hypothetical. The article does not report an adoption effectiveness study or client results.

Framework sources: NIST AI RMF 1.0, 2023, GOVERN, MAP, MEASURE and MANAGE functions and NIST AI 600-1, Generative AI Profile, 26 July 2024. Checked on 1 November 2026. The assessment record is my own work, with no claim of certification or full NIST compliance.