Skip to content
Editorial image of a graduated row of machined steel components on a red felt mat along a daylit workbench.

Essay 03 · AI and Operations

Useful AI starts
with the workflow

Start with the work that needs doing. Then ask what information it requires, who can check the result, and how you will know whether it helped.

01The essay3 min read

01Start with the task

A convincing demonstration can make a capability look like a solution. But the business question comes earlier: which part of the work needs to change? Name the task, who performs it, the handoffs around it and what a good result requires. If the work is poorly defined, adding a model may simply make the confusion faster.

Consider a team that repeatedly turns product information into answers for sales inquiries. Is the slow part finding an approved specification, shaping an answer for a particular buyer, or waiting for someone authorized to confirm it? Those are different problems. A better source library, a clear response template or a changed approval step may be enough. AI is worth considering where it can help with a specific part of that sequence without obscuring responsibility.

02Make information checkable

The next question is what the work needs to know. A model can help draft or organize material, but it may lack the current specifications, policies and decisions on which an answer depends. Retrieval can bring relevant documents and their sources into view for a reviewer. It cannot establish that a document is current, that the retrieved passage answers the question or that the resulting draft is accurate.

Information quality is part of the design. Which source is authoritative? Who keeps it current? Can the person checking an answer open the underlying material and spot a gap? If the available records disagree, the useful response may be to surface that conflict rather than to produce a confident sentence. Showing a source makes checking possible; it does not replace the check.

Showing a source makes checking possible; it does not replace the check.

03Give review real space

Keep the person responsible for the decision in the workflow. Imagine a draft reply assembled from approved product documents. The reviewer needs enough time to read it, access to the original sources and a clear way to correct or reject it. A review step added only in name can become a bottleneck or a rubber stamp. Its cost belongs in the evaluation alongside any time saved drafting.

Start with the task. Give the reviewer a way to verify the answer. That is a more useful brief than asking for an agent to handle everything. For low-risk, reversible work, a first draft may be helpful. For a consequential answer, a human may need to do much more than approve a sentence. The level of review should follow the stakes, not the novelty of the tool.

04Measure the change

Measure the work before and after a small trial. Record a baseline that fits the task: time spent, corrections, missed information, handoffs or the effort of review. Then compare the changed process against that baseline, including maintenance of sources and the cost of mistakes. A faster first draft is not a win if checking it takes longer or errors become harder to notice.

Expand only when the evidence supports it. Sometimes clearer instructions, a template or better-organized information will do the job with less overhead. If AI does help, define where it begins and ends, keep ownership visible and revisit the measure as the work changes. The objective is not more automation. It is work that people can do better, with a result they can explain and sustain.

Questions to take into the work

  • Which step is actually slowing or weakening the work?
  • What information is authoritative, and who keeps it current?
  • Who has the time and context to correct or reject an output?
  • What baseline would show whether the changed process helps?

Connects toAI, Data & Automation Digital Products & Commerce

Next essay

04Search / AI · 3 min read

AI search is reading your website like a customer

People now ask an assistant who to choose. The assistant reads what a careful customer would read, faster and across more sources, and it does not fill in the gaps for you.

Read
All perspectives

First Light™

Bring us the question
you’re working through.

A short conversation and a preliminary read on your strengths, your gaps and where we’d start. Our team reviews every read before anyone follows up.