Start with the shape of the task
Moving a website enquiry into a CRM, preparing a research brief and publishing a campaign are different jobs. They do not automatically need the same technical approach. In a repeatable process, you can normally name the trigger, the steps and the expected result. Research may first require deciding which information is missing and where to look for it.
Anthropic distinguishes between predetermined workflows and agents that choose their next steps dynamically. That distinction is useful when discussing a project. It does not mean that every existing marketing process needs an agent. Technical background
Ask three questions before choosing software
Write the task in one sentence: “When an enquiry arrives, create or update a contact and prepare a reply.” Then ask whether the required steps are known, whether the input follows a useful structure, and whether someone can assess the output against clear criteria.
If all three answers are yes, a workflow with a limited AI step may be sufficient. AI could extract the requested service, preferred language and unanswered questions from free text. Established rules can then route the enquiry. An agent becomes more relevant when the research path or tools genuinely need to vary between cases.
Example: Preparing an enquiry for a real response
Imagine a consultancy receiving a request for a team workshop. A sensible example process starts by retaining the original message. It checks for an existing contact, retrieves approved company information and creates a response draft. Missing details, such as the number of participants or delivery format, appear as explicit questions.
The model can suggest wording. It should not invent a price or promise an unavailable date. A responsible person reviews the enquiry alongside the proposed response and decides whether to send it. This is an illustrative workflow, not a claim about a measured customer outcome. It shows why a narrow process can still be valuable.
Where an agent could add something
An agent might research approved public sources to prepare a workshop brief. It needs a precise goal, permitted sources and a limit on time or steps. If company identity remains uncertain, the useful result is a clear uncertainty flag rather than a confident but unreliable profile.
NIST treats ongoing risk assessment and management as part of operating AI systems. Applied to this marketing example, that suggests named ownership, visible failures and regular result reviews. The exact controls should reflect the actual task and consequences. NIST AI RMF
Measure the complete job
Compare more than generation speed. Include input preparation, review, corrections, exception handling and maintenance. A draft that appears in seconds offers little benefit if a colleague then spends substantial time reconstructing missing context. Task frequency matters too: an elaborate integration for an occasional request may be hard to justify.
Before launch, try a complete enquiry, an ambiguous enquiry, an existing contact and a temporarily unavailable CRM. Record what the system does and where a person takes over. Also check that running the same request twice cannot accidentally produce duplicate outbound messages. These tests make the trade-off between flexibility and operating effort concrete.
Choose a bounded starting point
Draw the process with its input, outcome, owner and common exceptions. Our AI automation service covers repeatable handovers and connected marketing workflows. AI agent development addresses tasks that need flexible research or tool selection. In both cases, the point is a process your team can understand, assess and use reliably.
