Skip to main content
Back to Field Notes

AI in the Real World Private draft · 3 min read

AI Won’t Fix a Broken Workflow. Start Here Instead.

A practical way to decide where AI belongs in a service operation.

Start with the outcome

Pick one workflow you can name without describing a tool. A quote starts with a customer request and ends with something the customer can accept. Reconciliation starts with records that need to agree and ends with discrepancies resolved or assigned to someone who can resolve them. Writing down those boundaries makes it easier to see where the work actually begins and where responsibility ends.

Ask the people doing the work to walk through a recent case. Follow the information across handoffs. What arrived incomplete? What had to be checked twice? Which decisions depended on knowledge someone carried in their head? The official process may be useful as a starting point, but the walkthrough tells you which steps people actually rely on.

Find the bottleneck before choosing the tool

Separate time spent doing the task from time spent waiting. If a request sits because nobody owns the approval, a quicker draft still enters the same queue. If staff repeatedly search for complete information, the next step might be a better intake rule, a more dependable source, or a focused assist that collects the relevant materials.

Write the friction down in a sentence that names the task and the consequence. A description such as “quote approvals wait when the request is missing information” gives the team something to test. “We need AI for quoting” leaves too many decisions unresolved.

Illustrative workflow comparison
Where it waits
  1. Request
  2. Missing details
  3. Approval waits
One change to test
  1. Clear intake
  2. Named reviewer
  3. Decision

Test one change with a named reviewer

Choose a narrow task with clear inputs and an output a person can check. Define what the tool can read, what it may produce, and which cases it must hand back. Assign a person to approve consequential decisions and give them enough context to spot an error.

Before changing the workflow, record its current behavior using measures relevant to the job. For a quote, those could include time to approval, how often the approver needs to ask for missing details, and how much work corrections take. Measure the same things in the pilot. Include review time and failures as well as speed.

Make the next decision explicit

At the end of the pilot, decide whether the change helped the customer, the team, or both. If the apparent savings came from shifting work to a reviewer, account for that. If an intake fix resolved the issue on its own, keep it. If the pilot helped on routine cases but struggled with exceptions, limit it to the cases the team can handle confidently.

Start this week by picking one recurring workflow, sitting with the person who does it, and marking the first place the work waits, repeats, or changes hands. That is a better place to start a test than a list of tools.