Prepare your company for AI agents: processes, data and ownership

Before choosing an agent, a company needs three conditions: a written version of the process, data reachable by software and one person authorized to change the process. These tasks can start before a vendor is selected and may expose process problems even if the agent project does not continue.

Write down the process that runs

Agents operate processes, so the process is part of the specification. In many companies, the sequence and exceptions are held by operators while the document describes an older approved version.

For each candidate process, ask the people who perform the work to write one or two pages: the steps, systems and exceptions. Ask when a case does not follow the rule. Examples may include a customer handled differently, an approval that follows two steps in practice or a seasonal variant of the flow.

The first version can be incomplete and corrected against production cases. Starting from a blank page shifts all discovery into paid implementation time. Turning the definition into software is the work described in the operating model as code and one reason company-specific operations remain custom-phase work.

If nobody can produce the two pages, the company has learned that the process is not understood well enough to automate. That finding should pause the build until an owner resolves the missing decisions.

Make the required data reachable

The first process does not require a company-wide data cleanup. It needs access to its sources: a read credential, scheduled export, API or controlled folder. Format and quality problems can then be prioritized by their effect on that process.

A common detour is to make all agent work depend on a multi-year data program. An operation that needs three tables and a folder of PDFs may then wait for infrastructure it does not require. Limit the prerequisite to the sources used by the first process.

List the sources consumed by one process and obtain read access. Transformation can handle inconsistent formats, but a locked system, restricted export or long credential queue blocks the work. Broader cleanup can follow when the first process identifies which fields affect outcomes.

Test access with one dated case. Ask a person outside the owning department to retrieve every input needed to reconstruct it and record what was missing, who granted each credential and how long it took. That result is more useful than an inventory that only lists system names.

Appoint an owner with authority

The third condition is authority.

Agents change work, cross departmental systems and expose differences between the documented process and current practice. Each difference requires a decision. If every decision waits for a monthly committee, the project schedule follows that cadence. Why enterprise AI pilots fail covers this failure mode.

Name one person who can change the process, grant access to systems and staff, and reject scope. Authority matters more than title. Ask who decides when the proposed process conflicts with the sales department’s current practice and within what time. The answer should name a person and an escalation window.

A first 90-day plan

The first quarter can focus on one process and one measurable baseline.

The first month: choose one high-volume process where a caught mistake has a tolerable cost. Ask operators to document it and establish cases per month, human minutes per case and observable errors. The ROI of agents in operations explains the baseline.

The second month: obtain credentials and define controls before execution: what the agent may do, what escalates, caps and limits, and the trace for each action. Retrofitting these controls later changes the architecture. Governing agents by design covers the details.

The third month: run a bounded slice in production, with representative cases and human review. Track how the escalation rate changes as exceptions are encoded. At day ninety, require a running process, an auditable trace, a cost comparison and feedback from its operators.

Those ninety days also provide evidence for whether to buy a tool, build an internal capability or contract the outcome. The decision can use operating records instead of vendor claims alone.

Being ready is not having an AI committee

Committees, strategy decks, bulk tool licenses and an executive title can all exist without one written process, accessible source or authorized owner. Those activities do not demonstrate operating readiness.

For one process, readiness means an operator has written how the work happens, software can reach the required data and a named person can change the process. A company can establish those conditions within a quarter without declaring the entire organization ready.

Questions boards ask me

How long does it take to be ready?

For one process, it can take weeks: document the steps, obtain credentials and assign the owner. Company-wide readiness is not a prerequisite for starting. Expand process by process.

Shouldn’t we train our people on AI first?

General training before a live process is hard to retain. Train the operators of the first process on its controls and escalation paths as it goes live, then extend the material with the rollout.

Do we need to clean our data first?

Clean the data required for one process. Usually that means reachable sources and a person who can explain field meanings. Prioritize further cleanup using errors and costs observed in operation.

What if nobody internal can be the owner?

Then the company cannot currently make the process decisions an agent will surface. Assign authority over one process before spending on implementation, even if broader governance remains unchanged.