A polished answer can still solve the wrong problem
A life-science company asks AI to recommend a laboratory information system. Within seconds, it receives a comparison of platforms, features, and implementation considerations.
The response looks useful. But the organization has not agreed on what is wrong.
Are scientists struggling to find prior results? Are sample identifiers inconsistent? Is information difficult to retrieve from an external partner?
Each problem could require a different intervention. Some may need technology. Others may need clearer ownership, consistent records, or better handoffs.
A stronger model or more sophisticated prompt might improve the comparison. Neither establishes whether comparing software was the right task.
What problem formulation means
Problem formulation is the process of turning an observed difficulty or opportunity into a clear, testable statement of what needs to change, for whom, under which constraints, and how success will be measured.
It distinguishes symptoms from causes, makes assumptions explicit, and identifies the decision or workflow that needs improvement—without prematurely committing to a solution.
Oguz A. Acar’s HBR article, “AI Prompt Engineering Isn’t the Future,” argues that defining problems is a more durable capability than mastering prompt techniques.
For operating leaders, that distinction extends to model selection. Discussions often begin with which model to use and how to instruct it. Both matter, but they answer downstream questions:
- Problem formulation: What are we trying to change, and why?
- Model selection: Which tool is suitable for that task, given its capabilities, risks, cost, and operating requirements?
- Prompt engineering: How do we communicate the task, context, constraints, and expected output?
These are complementary activities. Starting with the problem gives the other two a useful direction.
Separate the symptom from the operating need
Consider an illustrative research team spending hours preparing its weekly program update.
The initial request might be: “Which AI model can write our status reports?”
That could save drafting time. But the real delay may come from reconciling inconsistent milestones, locating vendor updates, and determining who owns unresolved issues.
Automating the narrative could make the report faster without making it more dependable.
A better formulation would be:
“Our program leaders need a weekly view of milestone changes, unresolved dependencies, and decisions requiring escalation. Source information is scattered, and conflicting updates are not consistently resolved.”
That statement exposes work beyond writing. The team may need shared definitions, designated sources, and a consistent update process. AI could then summarize approved inputs and highlight discrepancies for review.
The objective is a more useful operating review—not simply faster text generation.
Write a problem brief before choosing the technology
A one-page brief should answer six questions:
- What needs to improve? Identify the workflow, its users, and the decision or action the result should support.
- What happens today? Describe observable friction, delay, error, or uncertainty rather than prescribing a solution.
- What evidence is available? Identify authoritative sources, missing information, conflicting definitions, and untested assumptions.
- What boundaries apply? Specify permitted data, approved environments, access restrictions, and prohibited actions.
- What would success look like? Define measurable improvement, such as less preparation time or fewer unresolved discrepancies.
- Who remains accountable? Name the reviewer, decision owner, and escalation path.
This brief remains useful even if AI is not the answer. It reveals whether the workflow is ready for automation or still needs basic operating clarification.
Let the problem guide model selection and prompting
Once the task is clear, evaluate candidate models against representative examples rather than choosing solely by reputation or a compelling demonstration.
A tool suited to summarizing approved documents may not be the right choice for another task requiring different reasoning, integrations, or data protections. Assess output quality alongside security, deployment arrangements, responsiveness, cost, and review effort.
Then develop prompts that express the agreed problem: relevant context, permitted sources, required distinctions, uncertainty handling, and output format.
The goal is not to find one universally “best” model or perfect prompt. It is to establish a dependable combination of tool, instructions, information, and human oversight for a defined use.
Use AI to challenge the framing
Teams do not need a perfect problem statement before experimenting. AI can help identify missing questions, alternative explanations, and assumptions worth testing.
For example:
“What explanations other than insufficient laboratory space could account for this scheduling bottleneck? What evidence would distinguish among them?”
Treat the response as hypotheses, not a diagnosis. People familiar with the work must test those ideas against actual conditions.
The process is iterative: formulate, explore, examine evidence, and refine. Model and prompt adjustments support that process rather than replace it.
Establish boundaries before expanding use
In life sciences, intended use matters. Organizing questions for exploratory research differs from producing information relied upon in a controlled process.
Distinguish exploration, drafting, decision support, and authorized action. Keep confidential information within approved environments, verify important conclusions against sources, and involve Quality where regulated work may be affected.
Neither model selection nor prompt wording establishes compliance by itself. The surrounding workflow, controls, and accountability matter.
Clear boundaries give teams a practical place to experiment—and a recognizable point at which additional review is required.
Operational takeaway: define success before optimizing the technology
Choose one recurring workflow where AI might help. Formulate the problem with the people doing the work, establish the evidence and boundaries, and agree on how results will be evaluated.
Then select the model, develop the prompt, and test the complete workflow—including review and correction.
Did the approach reduce total effort? Preserve important exceptions? Support traceable conclusions? Improve the decision?
If the answer sounds impressive but the operating problem remains, revisit the formulation.
The lasting advantage is not access to a particular model or a library of clever prompts. It is the ability to define consequential problems clearly and organize the right tools and people to solve them.