Start With the Workflow, Not the AI

A practical guide for small-business owners and operations leaders

When a new AI tool appears, it is tempting to start with a demo: “Could this write our replies?” or “Can it process these forms?” A better first question is: What part of our work needs to improve, for whom, and how will we know?

AI is not a business outcome. Faster drafting, fewer clicks, or more generated content may help—but only if the whole process gets better. A quick draft that takes longer to check, introduces errors, or creates extra follow-up can make the business slower even when the model itself responds in seconds.

Map the work before shopping for a toolLink to this section

Pick a recurring workflow with a visible pain point: answering customer questions, preparing quotes, processing invoices, or assembling a weekly report. Follow one real case from start to finish. Record who does each step, which systems they use, what information they need, where work waits, and what happens when an exception appears.

Talk to the people who perform the work and the people who receive its results. A manager may see “slow invoice processing”; staff may reveal that half the time goes to resolving mismatched purchase orders or chasing missing approvals. Automating invoice entry would not fix that bottleneck.

Write down a baseline before making changes. Depending on the workflow, that might include end-to-end completion time, error or rework rate, backlog, cost per completed case, customer complaints, or staff time spent on repetitive steps. Keep the measurement proportionate: a small business can begin with a representative sample and a simple spreadsheet.

Ask whether AI is the right interventionLink to this section

Once the process is visible, consider several options. Could a confusing form be simplified? Could a rule-based script handle a stable, predictable step? Could staff use a search tool or a template? Is AI useful for interpreting messy language, summarizing documents, or drafting a response? Sometimes the best first investment is better data or a clearer policy—not a model.

Research on knowledge work cautions against treating AI as universally helpful. In a field experiment with 758 consultants, AI assistance improved performance on some tasks but not others; the researchers described an uneven “jagged frontier.” The important lesson is not to copy the study’s percentages into your business plan. It is to test the actual task you care about, because superficially similar work can demand different judgment.

A simple one-page starting briefLink to this section

Before evaluating vendors, fill in these blanks:

• Outcome: What should improve for the customer, employee, or business?
• Current workflow: What are the steps, handoffs, waits, and common exceptions?
• Baseline: How do we measure current speed, quality, cost, and workload?
• Candidate task: What specific step might AI assist with—and what should remain human-owned?
• Risk: What would a bad result cost? Can it be caught and reversed?
• Alternative: Would process cleanup, ordinary automation, or training solve the problem more simply?
• Owner: Who is accountable for the workflow, not just the software?

This is enough to turn a vague “we should use AI” idea into a testable business question. It also helps you say no to an attractive demo that does not address a real constraint.

The next step is to compare possible workflows and choose a first experiment that is valuable, measurable, and safe to learn from.

Sources and notesLink to this section

• Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, Harvard Business School AI Institute, 2023. Experiment with 758 consultants; useful evidence of uneven task effects, not a universal business forecast.
• OECD, The Adoption of Artificial Intelligence in Firms, 2025. Survey and interview research on firm adoption and barriers.