Rules where they fit.
Agents where they help.

What is the difference between an agent and a workflow?

A CRM workflow follows a predefined sequence of triggers, conditions and actions. An AI agent can decide which permitted step or tool to use next according to the task and the evidence it receives. The distinction concerns how work is directed, not whether the interface includes a chat box.

For example, “create an internal task when an enquiry is accepted” can be a fixed workflow. “Find the relevant conversation and explain what is holding up this opportunity” may require several searches chosen according to earlier results. The second task benefits from flexibility because its path is harder to specify in advance.

Anthropic’s engineering guidance distinguishes predefined code paths from agents that dynamically direct processes and tool use [1]. In practice, products can combine both. A deterministic outer process may validate inputs, run a model-assisted research step, require approval and then apply a tightly defined action.

Use rules when the condition can be stated precisely

Fixed workflows are a good candidate when the relevant facts already have clear fields and the result follows an agreed rule. An accepted enquiry can create a next-action task. A renewal date can place a record into a review queue. An opportunity without an owner can be highlighted for assignment.

The condition must reflect the actual business meaning. “No recorded activity for seven days” is not necessarily “the client has lost interest”. A rule can identify the first condition without asserting the second. Give the resulting alert a label that reflects the evidence and lets the owner interpret it.

Define what happens when the condition stops being true, the record is archived or the rule runs again. Repeated runs should not create an unlimited stack of identical tasks. Record the identity of the business event or enforce an agreed one-open-task rule, according to the system’s supported contract.

Use an AI step when the path is known but the input varies

Some work needs interpretation without an open-ended agent loop. A team might ask a model to summarise an approved conversation, extract a proposed meeting time or classify a message into a reviewed category. The process can still have fixed inputs, a constrained output and clear validation.

Consider an enquiry arriving as free text. The form handler can deterministically validate and save the request. An AI step can suggest a category and explain the relevant sentence. A human or a tested routing rule can then decide ownership. Each step has a separate role, making failures easier to locate.

Keep unknown values possible. A model should be able to say that the message does not state a budget, instead of manufacturing a number to satisfy a required field. Validate outputs against the destination schema and keep the original evidence available for review.

Use an agent when the next useful step depends on discovery

An agent is a candidate when the problem is bounded but the research path varies. Preparing a relationship brief may require finding a company, checking its opportunities, reading a particular thread and investigating a referenced meeting. The result of one read determines whether another is needed.

Give the agent a specific outcome and a restricted set of tools. “Prepare a brief for tomorrow’s review using this account and its recent correspondence” is easier to evaluate than “improve our sales”. Specify whether the task is read-only, may prepare changes or may execute particular authorised actions.

Set a stopping condition. The agent should finish when it has enough evidence for the requested result, ask a precise question when identity is ambiguous, or report the limit it reached. Continued searching is not automatically progress, especially if the tools cannot reach the missing source.

A hybrid example: enquiry research and follow-up

Take a fictional professional-services team. A validated website submission creates an enquiry and a follow-up task using a fixed intake contract. The submission ID prevents a repeated request from creating the same work twice. This initial handoff does not need the model to decide whether required fields exist.

The owner then requests a brief on the prospective client. An agent researches permitted CRM and email context and prepares an evidence-backed summary. If the request contains ambiguity, it asks for clarification or leaves the uncertain detail unresolved. It may propose a draft, but that is not an instruction to send it.

Approval then returns the process to a controlled path: verify the target and current record versions, apply the selected change and save the result. The model helps interpret the context while the application enforces the allowed effect. A failed approval returns a specific conflict for review.

Compare the approaches using the same work

Write a shared acceptance test before comparing solutions. For a follow-up task, success could require the correct contact, an appropriate draft based on the latest conversation and no invented deadline. Run the same representative cases through the existing manual process, a simple workflow and the proposed agent design.

Count review and repair effort as part of the result. A faster first draft may still require more time overall if it attaches the wrong relationship or repeats an outdated commitment. Conversely, an agent that correctly identifies missing context may produce no draft while still doing the right thing.

Record variability across repeated trials. A one-off success is a demonstration, not a dependable rate. Keep the dataset and grading rules consistent when comparing a new configuration with the previous version, and supplement the controlled cases with observed production exceptions.

Write down the decision and its limits

Use the worksheet to record the task, input ambiguity, allowed tools, expected outcome and recovery owner. Choose a fixed workflow, one AI-assisted step or an agent according to the work. Avoid building flexibility that no one needs to use or maintain.

Define an initial operating boundary: which record types, which source accounts, which action types and which cases require a person. Make the fallback explicit. If the model is unavailable, can the team still use the records and complete the handoff manually?

Revisit the choice when evidence changes. A repeated agent investigation may reveal a stable pattern that can become a simpler workflow. A rigid rule that repeatedly misclassifies complex enquiries may need a reviewed interpretation step. Architecture should follow observed work rather than a permanent preference for one label.

Common questions.