By function

A research agent is valuable when it produces a packet someone can decide from, not a longer brief nobody asked for.

Analysts and operators re-research the same questions across documents, CRM, and the web. The cost is time-to-decision, not a shortage of text.

Adnan Boz
Adnan Boz

How I see it

Research agents

The done state of a research agent is a recommendation with evidence, gaps, and a suggested next action. It is not a chat that feels informed.

This is a strong fit for account research, vendor comparison, underwriting support, property or deal packets, and internal knowledge that is currently trapped in files.

Citations, source freshness, and a human who owns the decision are the control points. The agent should make the uncertainty visible.

If the research does not change a decision speed or a downstream workflow, it is content, not leverage.

Common mistakes

What teams usually get wrong.

01

Unsourced fluency

A confident memo with no evidence is a risk product.

02

Research with no decision owner

Packets pile up. Nobody acts.

03

Boiling the ocean

An agent that tries to know the whole company will not finish a case.

A useful diagnostic

Five questions before you fund the work.

  1. What decision is delayed by research today?

    If you cannot name it, do not fund a research agent.
  2. How often is the same research repeated?

    Repetition is the economic case.
  3. What sources must be included, and which are forbidden?

    Source policy is part of the design.
  4. Can a decision-maker act from a two-page packet?

    If they need a 30-page brief, the process is the problem.
  5. Will the packet be used in a real meeting this month?

    If not, you are building a library.

Economic model

Research leverage

recurring questions × hours per packet × loaded cost = the research-agent case

Add decision-cycle time if you can measure it. Do not invent a strategy value.

Three credible paths

How far should you go?

Do not force one solution. Choose the path the economics, the risk, and the organization can support.

01

Improve the work you already have

Keep the process mostly intact and use AI on the bottlenecks that create delay, rework, or follow-up.

Best when

The workflow is already sound and a few steps create most of the friction.

Limitation

Gains are usually incremental. The operating economics do not change much.

02

Redesign the workflow around AI

Question every handoff, queue, and duplicate step, then rebuild the process around what AI can now do.

Best when

The process grew over years and coordination now costs more than the work itself.

Limitation

Requires process change, clearer ownership, and a willingness to retire old steps.

03

Put an agent on a high-value outcome

Give an agent responsibility for one valuable result across systems, with humans at the control points that require judgment, authority, or risk acceptance.

Best when

The workflow is high-value, variable, multi-step, and worth engineering for production.

Limitation

Needs stronger architecture, evaluation, controls, and monitoring. This is not a prompt project.

When this is the wrong next step

Do not fund an agent here.

  • The question is truly one-off and a senior person should do it once.
  • Sources are not available or not allowed.
  • Leadership wants a general company brain instead of a packet for a real decision.
Adnan Boz

A useful next step

Bring one workflow. Get guided into production.

We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.

Discuss an AI opportunity