AI decision support7 min read · 13 / 20

‘Build an AI agent’ sounded exciting. It was not a product requirement.

The agent could search, summarize, draft and act. What nobody had defined was the decision it was responsible for improving.

Written by Mitesh DabhiAI Adoption Specialist
Hand-drawn business decision at the center while an AI agent gathers evidence and prepares a recommendation around it.
A useful agent is organized around a decision, the evidence it needs and the boundary of what it may do.

An agent without a decision to support can stay endlessly busy without becoming useful.

The request included a long list: connect systems, understand context, generate recommendations and take action. Each capability was technically possible.

As I tried to shape the product, the scope kept expanding. A general agent needed access to more data, more tools and more permissions, while the expected business outcome remained vague.

The project became clearer when I replaced ‘what should the agent do?’ with ‘what should a person be able to decide better because this agent exists?’

01 / Starting with the decision

A specific decision gave the agent a responsibility.

Should this lead receive immediate attention? Is the document ready for submission? Which project requires escalation? Each question implied a different evidence set and a different risk boundary.

  • Decision owner
  • Evidence needed
  • Frequency
  • Risk level
  • Action boundary
  • Success measure

The agent stopped being a collection of features and became a focused workflow with a reason to exist.

02 / Evidence before recommendation

The quality of the answer depended on the evidence the agent could organize.

For project risk, that meant deadlines, recent activity, blockers, capacity, communication, scope changes and pending approvals. Designing retrieval around the decision was more useful than giving the agent broad access to everything.

The three boundaries

I define what the agent informs, what it recommends and what it may execute.

01

Inform

Gather and present relevant, source-linked evidence without converting it into an unsupported conclusion.

Facts · history · gaps · sources
02

Recommend

Suggest a next step with reasoning, uncertainty and alternatives visible to the decision owner.

Option · rationale · confidence · risk
03

Act

Execute only inside defined, reversible rules or after an accountable person approves the action.

Prepare · approve · execute · audit
03 / Measuring decision quality

Fluent output was no longer enough.

We evaluated whether the correct evidence was used, important risks were identified, sources were represented honestly and the person accepted or improved the recommendation.

01Correct evidence02Source accuracy03Risks identified04Useful recommendation05Human acceptance06Decision time07Unsafe action prevented08Outcome tracked

Decision time and prevented mistakes mattered more than how sophisticated the agent appeared in a chat window.

The product question“What decision should become faster, clearer or safer because this agent exists?”
04 / Useful restraint

A focused agent created more impact than broad autonomy.

A reliable weekly risk review needed fewer permissions and was easier to test than a general agent with access to every system.

Defined purposeRelevant evidenceClear authorityLower riskMeasurable qualityGreater trust

The narrower design created a clear expectation for users and a manageable responsibility for the product team.

What I carry forward

I build agents around responsibility, not activity.

The business decision comes first. Evidence, reasoning, interface and action boundaries follow from it.

That order prevents an impressive list of capabilities from becoming an expensive product nobody can evaluate or confidently use.

A good agent is not useful because it can do many things. It is useful because it improves one clearly defined responsibility.