Remove the slowest step without removing the review

A purchasing analyst at a network of clinics compares supply quotes every week. She opens a chat, explains the criteria (price, delivery time, contract length, certifications), pastes one quote and asks for an assessment. She repeats this for each of the six to eight quotes, copies every answer into a spreadsheet by hand and opens another chat to draft the recommendation email. It takes up her Wednesday afternoon. Sometimes she forgets a criterion in one of the chats, and the assessments are not comparable.

She uses Claude all the time. The workflow is still slow and unreliable.

Efficiency and effectiveness

To optimize a workflow has two sides, efficiency and effectiveness: that it takes less time and that it comes out better. Almost always both problems share the same root. In the analyst’s case, repeating the criteria in each chat costs her time and makes her drop a criterion now and then.

Where to look

Before changing anything, measure or at least estimate how long each step takes and where errors appear. The typical places where time is lost:

Symptom	Redundant step	Fix
Explaining the same context in every chat	Repeating instructions	A Project with the criteria in the instructions
One request per case, one after another	Processing one by one what can be compared together	Assessing all cases in a single request, in a table
Copying by hand what Claude wrote in prose	Retyping	Asking for structured output: a table with the spreadsheet’s columns
Each result comes out in a different format	Reformatting	A fixed output template
A Claude check to catch errors born in a manual step	Controlling an avoidable error	Removing the manual step that creates the error

A redundant step is one that adds no value: someone does it because the workflow was built that way, not because it is needed.

An example with numbers

The legal team at a property management company receives about 40 lease contracts a week. Their workflow: Claude summarizes each contract in prose, a paralegal retypes the key data into the tracking spreadsheet (parties, rent, term, renewal), Claude checks the entry against the contract and a lawyer reviews and approves. They measured for a month: retyping takes 70% of the time, and almost all the errors the lawyers find are transcription errors.

The measurement points to a single step. If Claude extracts the data straight into the spreadsheet’s columns, the retyping disappears, and with it most of the errors. Adding a second check after the retyping attacks the symptom; removing the step attacks the cause.

What does not get optimized

Human review that adds value is not a redundant step. The lawyer who approves each contract and the manager who reviews each recommendation before it goes out stay. The fact that Claude “already checks” does not replace whoever answers for the result.

Nor do you optimize by switching models: a more capable model still repeats the same steps.

The typical traps

• Removing the lawyer’s or manager’s review because Claude already checks.
• Adding another Claude check instead of removing the step that creates errors.
• Moving to the biggest model without changing the workflow.
• Splitting the work into more separate chats “for accuracy.”
• Optimizing the step that is visible, not the one the measurement points to.

On the exam

If the scenario describes a slow workflow with manual steps, look for the option that removes the step that takes the most time or where errors originate (repeating context, going one by one, copying by hand) and keeps the human review. The clues are usually in the numbers: which step takes the most time and where the errors come from. Rule out the options that remove the review, the ones that add checks on an avoidable step and the ones that only switch models.