Redesign the workflow before automating it

The fastest automation can still accelerate a broken handoff. That sentence has ended more than one project kickoff, because most teams arrive wanting to automate the workflow they have, and the workflow they have is usually a fossil record of old tools, departed employees, and workarounds that made sense in some forgotten quarter. Automating it preserves the dysfunction and speeds it up.

Redesign comes first. The sequence below is unglamorous, mostly involves talking to people and drawing boxes, and routinely produces more value than the automation that follows it.

Work backward from the outcome

Begin with the outcome the customer or stakeholder actually needs, the treatment plan understood, the closing completed, the claim resolved, and work backward through every handoff that produces it. Walking the workflow backward exposes what walking it forward hides: steps that exist because a previous tool required them, approvals that no longer approve anything, and information collected twice because two systems refuse to talk.

As you map, separate the work that creates value from the work caused by missing information, unclear ownership, or duplicate tools. The second category is usually larger than anyone expects, and it is the real automation target, not because AI should do that work faster, but because redesign should make much of it disappear entirely.

Make the source of truth explicit

Most broken handoffs are information problems wearing a process costume. The coordinator re-asks the customer because the intake form's answers live in a tool the coordinator does not open. The follow-up is missed because three people each assumed another team's spreadsheet was current.

For every important fact in the workflow, such as the customer's status, the price quoted, or the date promised, name the one system where it lives and the one role that keeps it current. Until the source of truth is explicit, any automation layered on top is guessing, and an AI agent grounded in three contradictory systems does not resolve the contradiction. It launders it into confident prose.

Define decision rights, review points, and exceptions

Before assigning any step to software, decide who is allowed to decide. Which approvals are real controls, and which are rubber stamps that only add delay? What are the recurring exceptions, the rush order, the distressed caller, the incomplete file, and what should actually happen for each? Exceptions are where workflows quietly fail today and where automations loudly fail tomorrow, so they deserve design attention proportional to their frequency, not their glamour.

This is also where human judgment gets protected on purpose. Steps involving professional interpretation, relationship repair, or consequential trade-offs are marked as human work, so that no later phase of the project tries to automate them by accident.

Only then choose the executor

With the redesigned workflow on the table, each remaining step gets the same four-way question: does this belong with a person, with conventional deterministic software, with an AI model, or with an agent that can take actions?

The honest answers are boring. Calculations, routing, record updates, and anything where identical inputs must produce identical outputs belong in deterministic software. It is cheaper, testable, and never hallucinates. Language-heavy work with genuine ambiguity, such as summarizing a messy inquiry, drafting a first response, or extracting facts from unstructured documents, is where a model earns a role. Agents, with their broader authority, are reserved for jobs bounded enough to supervise. And people keep the judgment calls. AI is one option in the lineup, chosen where it wins on merit, not the premise of the exercise.

Measure the workflow, not the automation

Finally, instrument the redesigned workflow with a small set of operating indicators: cycle time from request to outcome, rework and error rates, missed or late follow-ups, capacity returned to judgment-rich work, and customer experience at the moments that matter. Set the baseline before anything launches, or the after will have nothing honest to be compared against.

Notice that none of these measures mention AI. That is deliberate. The business did not need an automation; it needed the outcome to arrive faster, more reliably, and with less wasted effort. Measuring the workflow keeps every later decision, including the decision to remove an underperforming automation, anchored to the thing that was always the point.