Process before technology: where AI agents should start

Most businesses ask "which AI tool should we use?" first. It is the wrong first question. The tool depends on the work, and most businesses have never written the work down end to end.

Choose the right first process

Agentic systems fit best where three things are true:

• High volume. The same kind of work arrives again and again, so improvements add up.
• Many steps. The work passes through several hands before it is done.
• Several systems. People copy information between email, the CRM, documents and spreadsheets.

Good candidates include lead-to-revenue, client onboarding, order-to-cash, procure-to-pay, service and case resolution, and reporting. These are examples, not limits. The processes that carry the business are usually the ones that strain first as it grows.

Map it end to end

A useful map does not need special software. For each step, write down:

1. Who does it, and in which system.
2. What information they need, and where it comes from.
3. Where the work waits for the next person.
4. Which decisions are made, and what rules they follow.
5. What goes wrong, and how often.

The waits are usually the surprise. In most manual processes, the work itself takes minutes and the waiting between people takes days. As volume grows, so does the queue.

Find the constraint

Once the map exists, look for where volume, cycle time and errors break down as the business grows. That is the constraint, and it is where an agentic system creates the most leverage. Improving a step that is not the constraint makes the map look better without making the business faster.

Rebuild, do not patch

Here is the shift that matters. Patching means keeping every step and adding AI to some of them. Rebuilding means asking: if agents can read, sort, check and update, what should this process look like now?

Often the answer is that intake, checks and updates go to agents, and people keep the decisions that need judgment. A process that used to pass through four people with a wait at every handoff becomes one where the routine work flows straight through and a person approves the step that carries risk.

Write the business case

Finish with a short, prioritized roadmap: which process first, what will change, and which numbers it has to move. Cycle time, throughput, exception rate and cost per transaction are good places to start. We cover how to set those up in measuring an AI system in business terms.

That is what the first step of every build here produces. It is also the part most teams skip, and the reason so many end up with tools that never change how the business runs.