AI agent or automation: where to start?

Start with the work that needs to get done

A team receives requests by email, checks documents and logs each request in another system. Before choosing a tool, it pays to map out that routine: where the information comes from, which decisions are made and how a correct result is recognized. For a first project, that map is usually more useful than a list of AI models.

In the distinction Anthropic draws, a workflow orchestrates models and tools through paths defined in code, while an agent dynamically decides which steps and tools to use. The difference is who controls the path. More autonomy also means weighing cost, response time and the risk of errors compounding.

Three ways to handle the same routine

Consider a hypothetical example: triaging incoming requests. If every message already contains standardized fields, an automation can check that they are filled in and route the request to the right queue. The rules are explicit, and exceptions have a known destination.

If requests arrive as free text, an AI step can suggest a category, extract information and prepare a summary. The system can still require mandatory fields before logging anything. This design pairs interpretation with a controlled process.

If the task requires working out what is missing and consulting different sources for each case, an agent is worth trying. In the pilot, it can present a draft along with the lookups it ran. The team assesses whether that flexibility is useful before expanding the agent’s permissions.

Set the operating boundaries

For the first version, we recommend writing a one-page operating contract. Separate what the system can look up, what it can suggest and what it can actually change. Name a person responsible for reviewing exceptions, and a channel through which work returns to the team when it cannot proceed.

• Input: what information is needed to start handling a request?
• Output: which fields and evidence make a request ready for review?
• Action: which changes need a person’s approval?
• Stop: what should happen when data is missing, the target system fails or the request is out of scope?

Compare it with the current routine

Set aside representative examples and record how the team handles them today. In the comparison, account for review time, requests sent back and corrections needed. A convincing answer that demands heavy review may deliver little operational value. Factor in maintenance, too: someone will need to keep categories, rules and access up to date.

Follow one case from start to finish. Did the summary help the person who received the request? Did the category prevent misrouting? Was the record usable in the target system? These questions keep the evaluation close to the real work and help pinpoint the next adjustment.

The first step forward can be small: organizing intake and preparing consistent drafts for the team. Define in advance the result that would justify a second phase. That way, the automation grows in step with a demonstrated need, with clear responsibilities and room to learn during the pilot.