AI Before Automation? Start With the Workflow Instead

Many automation conversations now start with the same question: where can we use AI?

It is an understandable question. AI tools are widely available and capable. But as a starting point, it tends to produce pilots that look impressive and change little. A more useful first question is: where is work creating unnecessary friction?

Why technology-first automation disappoints

When automation starts from a technology, the search is for problems the technology can address. That reverses the logic. The result is often a solution applied to a process that was not well understood, automating steps that should not exist, or handling the straightforward cases while the difficult exceptions remain manual.

Automation applied to a poor process tends to make the poor process faster. It can also make it harder to change, because the workaround is now embedded in technology.

Find the friction

Start by looking for where work is slow, repetitive, error-prone or dependent on chasing. Useful signals include:

• the same information keyed into more than one system;
• approvals that wait for someone to notice an email;
• documents that are read and re-typed;
• reports assembled by hand every week or month;
• queues where work waits for information from another team;
• frequent rework because inputs were incomplete.

These are the places where intervention is most likely to matter to the people doing the work.

Map the workflow

Before choosing a solution, understand the work. Map the steps, the people, the systems and the decisions involved, then separate the different kinds of activity within it:

• Rules: steps that follow clear, stable logic.
• Judgement: steps that depend on experience, context or discretion.
• Data movement: steps that exist only to move information between systems.
• Approvals: points where someone must authorize the next step.
• Exceptions: cases that fall outside the standard path.
• Content: documents, messages or text that must be read, summarized or produced.
• Decisions: choices that change what happens next.

Each type suits a different intervention. Treating them all as one automation problem is a common source of disappointment.

Choose the right intervention

The options are broader than AI:

• Process simplification: remove steps that add no value before automating anything.
• Workflow automation: route work, notifications and approvals automatically.
• Integration: connect systems so information does not need to be re-entered.
• Rules-based automation: handle steps where the logic is clear and stable.
• Document processing: capture and classify information from documents.
• AI-assisted work: support tasks involving interpretation, summarization or drafting, with people reviewing the output.
• Reporting automation: assemble recurring reports from reliable sources.

Often the most effective solution combines several of these, with AI playing a targeted role rather than the central one.

Not every problem needs AI

If a step follows clear rules, conventional automation is usually cheaper, more predictable and easier to test. If the problem is that two systems do not share information, integration solves it more reliably than a model reading screens. If the process has unnecessary steps, simplification may remove the need for automation entirely.

AI is most useful where work involves variation that rules cannot capture well: unstructured documents, varied requests, summarization or drafting that a person then checks. Those are real opportunities. They are also a subset of the friction in most organizations.

Human oversight matters

Where AI is part of a workflow, design the oversight with the same care as the automation. Decide which outputs need review, who reviews them, how errors are corrected and how the organization will know if quality changes over time. Define who owns the workflow once it is in operation. These questions are easier to answer before deployment than after something goes wrong.

Measure the operating outcome

Judge automation by what changes in the work, not by how much technology it contains. Useful measures include:

• Cycle time: how long the process takes from start to finish.
• Rework: how often work has to be redone.
• Error: how often outputs are wrong or incomplete.
• Handoffs: how many times work changes hands.
• Visibility: whether people can see where work is and what is blocking it.

Establish a baseline before changing anything. Without one, it is difficult to say whether the intervention helped.

Automate the right work

The organizations that get the most from automation are rarely the ones that adopt the most technology. They are the ones that understand their workflows well enough to choose the right intervention for each kind of work, and to leave alone what does not need to change.

Start with the workflow. The technology choice becomes much clearer afterwards.