Redesign the Whole Workflow—Don’t Just Add an AI Step

A practical guide for small-business owners and operations leaders

Putting AI into one step of a process can make that step faster while making the overall process worse. A generated draft still needs review. A summary may omit the detail the next employee needs. An automated handoff may send an exception to a queue nobody checks.

The useful unit of change is the whole workflow: inputs, decisions, handoffs, exceptions, and outcomes. Before adding a model, decide what work it will change and what new work it will create.

Draw the future processLink to this section

Start with the current-state map from the first article. Then sketch a target-state process with four explicit lanes: AI, people, software/rules, and customer or supplier. For every handoff, specify what information moves, who is accountable, and what happens when the system is uncertain or unavailable.

Give AI a bounded job. It might extract fields from a form, summarize a case, draft a response from approved policy, or suggest a category. Do not confuse a plausible output with an authorized decision. If a decision has financial, safety, legal, or employment consequences, define the responsible human and the review standard before testing.

Design exceptions as first-class paths, not afterthoughts. Examples include missing information, contradictory records, a low-quality scan, a request outside policy, or a tool outage. A good workflow routes those cases to a named owner with enough context to resolve them. A “human in the loop” label is not enough if no one has time or authority to intervene.

Example: a customer-support inboxLink to this section

Suppose a small repair business receives service questions by email. The existing process is: read the message, search policy and job history, decide whether it is urgent, draft a reply, and create a work order.

A narrow first design could let AI summarize the email and retrieve the relevant approved policy paragraph. It can draft a reply, but a staff member checks the source, urgency, and customer-specific details before sending or creating a work order. Messages involving safety, billing disputes, or unclear policy go to a specialist. If the policy source is missing or conflicting, the system should say so and route the case—not improvise.

This changes the employee’s job too. They may spend less time composing routine replies and more time checking evidence, handling exceptions, and improving policy documentation. That work requires training and should be counted when measuring whether the process improved.

Test both the ordinary case and the messy caseLink to this section

Before rollout, assemble representative examples: common cases, rare but important exceptions, incomplete records, ambiguous wording, and adversarial or irrelevant content. Have people who know the workflow score the output for correctness and usefulness. Measure downstream impact, not just whether the model produced text.

Research supports careful task-level testing. In a study of customer-support agents, AI assistance improved average issues resolved per hour, with larger gains among less-experienced workers; effects differed across workers. A separate experiment with consultants found tasks where AI improved performance and other tasks where it hurt. These studies show why workflow configuration and task fit matter; neither guarantees the same result in your business.

Treat redesign as a learning loopLink to this section

Run the new path on a bounded share of work or in a supervised “shadow” mode. Compare the old and new process, record corrections and exceptions, and ask staff where the design creates friction. Update the workflow and test again. Do not silently let the model’s own outputs become the new “correct” examples without review.

The outcome should be a clearer process, not merely a new AI node. Once the roles and handoffs are explicit, you can choose the right technical pattern for that job.

Sources and notesLink to this section

• Brynjolfsson, Li, and Raymond, “Generative AI at Work”, NBER Working Paper 31161; published in Quarterly Journal of Economics (2025). Single customer-support setting; the NBER working-paper abstract reports 14% average productivity lift and heterogeneous effects.
• Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, HBS AI Institute, 2023. A study of 758 consultants found task-dependent effects; do not generalize its effect sizes.