Using AI is not enough: what your team needs to learn

The working paper by Randazzo and colleagues, including Harvard Business School's Karim R. Lakhani, examines 244 BCG consultants. It distinguishes integrated collaboration, selective use and delegation with limited engagement, associated with different learning outcomes. It does not establish one approach as best for every business task.

Research on the jagged technological frontier also finds different performance across tasks. These are bounded studies, not forecasts for every model or team. The operating approach below is my interpretation of the evidence, not a Harvard protocol or a guarantee of improvement.

An active subscription or many conversations do not establish better work. Choose a bounded task and specify the expected result: a report with reconciled figures, a product page consistent with its catalog, or a proposal that meets documented constraints. Identify who can assess the deliverable and what information they need.

Record the starting point using comparable cases. Consider total time, corrections and quality, including review effort. If AI drafts quickly but the team spends longer checking the draft, the benefit remains unproven. Changing the process without adding AI can also be a valid alternative.

Write the brief before starting. Asking for research or a better report is not enough: specify permitted materials, the expected output, boundaries and acceptance criteria. Define which actions are authorized and which require a person. Access to information and permission to act are separate decisions.

The process owner agrees these boundaries with the people doing the work and, where needed, IT and security. An agent may prepare a comparison without permission to send it to a customer. The ability to build a demo does not establish the authorization and controls needed for production use.

Imagine an e-commerce team preparing a product comparison from an approved catalog. This is an illustrative scenario, not a client case or a measured result. An agent proposes a table; the team checks whether specifications, variants and availability match the sources and whether the comparison answers the original question.

Then introduce a variant missing an essential attribute or two conflicting documents. A useful outcome is not a complete table at any cost. It is recognizing the limit, flagging the missing information and not presenting an assumption as a fact. Reviewers must explain which criterion they applied, rather than simply approve the text.

Asking for objections can surface alternatives, but it does not make the responses true. A model's stated confidence is not a validated measure of your process risk. Two agents may share the same mistaken premise or rely on an incomplete source.

Use checks appropriate to the output: reconcile figures with original data, trace claims to sources, test and review code, and require approvals for consequential actions. Keep encountered errors as test cases. When the model, context or an integration changes, assess which checks to repeat before extending its use.

I suggest observing behavior, not just output or satisfaction after a session. Ask people to explain a decision, recognize an unsupported proposal and adapt instructions to a similar task. These are signals to document for the chosen case, not a general certification of AI competence.

If the process has not been selected or its data, roles and checks remain undefined, AI Strategy & Integration helps establish the priority use case and operational plan. Development and integration are included only where specified in the agreed scope.

If the project is already chosen and the need is to learn how to work on it, Agentic Execution Lab supports your team on the real case. They keep the agreed deliverable, files, instructions and reusable checklists. The services can complement each other, but buying both is not required. Ongoing operation and maintenance are agreed separately.