AI and automation

Starting with the workflow: a practical test for worthwhile AI opportunities

A disciplined way to separate valuable AI automation from demonstrations that never become dependable business processes.

01Context02Decision03Control04Outcome

The fastest way to create an impressive AI demonstration is to start with the tool. The better way to create a durable business result is to start with the work.

A worthwhile AI opportunity has a recognizable workflow, a useful output, an owner, and a result the organization can measure. Without those elements, experimentation may still teach something—but it should not be confused with operational transformation.

01 · Practical guidance

Look for friction that repeats

Good candidates often involve recurring effort: gathering information from several places, classifying requests, drafting a first response, comparing documents, routing exceptions, or assembling the same type of summary again and again.

Repetition alone is not enough. The workflow must also matter. Saving seconds on an infrequent task rarely justifies integration, testing, security review, and ongoing ownership. Focus on work where delays, manual touches, inconsistency, or rework create a visible operating cost.

02 · Practical guidance

Use six tests before choosing the technology

  • Frequency: Does the work occur often enough for improvement to matter?
  • Inputs: Can the required information be identified, accessed, and trusted?
  • Output: Is there a clear definition of a useful result?
  • Boundaries: Can the automation operate within a limited, understandable scope?
  • Oversight: Is there a sensible point for human review, escalation, or approval?
  • Measurement: Can the organization compare cycle time, effort, quality, backlog, or another outcome before and after the change?

03 · Practical guidance

Know when not to automate

Avoid beginning with workflows whose source information is unreliable, whose exceptions are poorly understood, or whose mistakes could create material harm without detection. A process that changes every time it is performed may need clarification and redesign before it needs AI.

Be especially careful when an automated action could expose sensitive information, communicate externally, change a system of record, approve a transaction, or make an important decision. In those cases, narrow permissions, logging, testing, and explicit human control are part of the solution—not optional additions.

04 · Practical guidance

Pilot the smallest useful workflow

Map the current process first. Record the inputs, decisions, handoffs, systems, exceptions, and baseline effort. Then select one bounded step where AI or conventional automation can produce a useful output without pretending to solve the entire process.

Test normal cases and edge cases. Confirm who reviews the output, how errors are corrected, what data is retained, and what happens when a dependency is unavailable. Measure the result against the original baseline before expanding the scope.

For example, an internal request process might begin with assisted classification and a draft summary for a human owner. That creates a measurable improvement without allowing the system to make commitments or modify critical records on its own.

05 · Practical guidance

Treat adoption as part of the system

An automation is not successful because it runs. It succeeds when people understand when to use it, trust it appropriately, recognize its limits, and know what to do when it fails. Process ownership, training, feedback, and support are therefore part of the technical design.

The practical takeaway

The best AI opportunities are not defined by how advanced the model appears. They are defined by a bounded workflow, responsible control, and a measurable improvement in real work.

Scott Craig leads SMBiSYS, a boutique technology consultancy connecting fractional IT leadership with Microsoft cloud, infrastructure, business systems, and practical AI automation.

Start a conversation

Bring your technology decision into focus.

Begin with fit, priorities, and the smallest useful next step.

Contact SMBiSYS