A customer request arrives. Someone reads it, checks the account, chooses a team and copies the details into another system. You want less of that manual work. Before asking for an AI agent, describe the decision the system needs to make.
Can the next step be written as a rule? Does the system need to interpret a message before following a known route? Or must it decide what to investigate next as new information appears?
Those requirements lead to different scopes. Start with the simplest option that meets the agreed standard on representative requests. Here is how to compare them through one hypothetical example, rather than choosing from technology labels.
One process, three different requirements
Imagine a service business that receives customer requests. The team wants each request to reach the right person with enough context to act. The examples below illustrate possible designs; they are not a customer case study or measured Madency results.
| What arrives | First option to test | What a person checks |
|---|---|---|
| A form with a selected department and a valid account ID | Rules that validate the fields and create the assigned task | Missing details and exceptions |
| A free-text email whose topic determines the department | An AI classification step followed by a defined route | Ambiguous or mixed requests and incorrect classifications |
| A request that requires choosing the next source to investigate based on earlier findings | A bounded agent, compared with a fixed investigation sequence | Evidence, unresolved questions and the proposed next action |
Each row adds a requirement. A useful first implementation may cover only the first or second row, while a person handles investigations.
When rules are enough
Suppose a form already asks the customer to select Billing, Support or Sales. The account ID is present, and the business has a clear assignment rule. A workflow can validate the fields, create a task and record the result.
The important design questions concern exceptions: what happens when the account does not exist, a field is missing or the destination system is unavailable? Assign an owner for those cases and make the failed step visible.
Adding a model to choose a department would introduce another decision to check, even though the customer already supplied that information. Start by testing whether the rule-based route removes the copying work you wanted to eliminate.
When one AI step helps
Now the same request arrives as an email: “We changed our subscription and the reports stopped working.” There is no department field. A model could suggest a category from an approved list, such as Account access or Reporting issue.
The surrounding process can remain defined: read the message, propose a category, check that the output uses allowed values, then create a task or send it for review. A valid category is not proof that the classification is correct. Mixed requests and unfamiliar wording still need evaluation.
Define when the system should defer to a person. Avoid treating a model's own confidence score as a guarantee. During a pilot, review its decisions against examples the team has already classified, including messages that belong to more than one department.
Here, AI interprets the input. The business still determines the route after that interpretation. This can be a useful scope without giving the model control over the full process.
When an agent is worth testing
Consider a broader requirement: prepare an evidence-based investigation before assigning the request. What the system finds in the account history determines whether it should inspect a recent change, consult a relevant support document or ask for missing information.
Write down those choices first. If a short, fixed sequence covers the useful cases, test that sequence. A bounded agent becomes a candidate when choosing the next step dynamically is a material part of the job and simpler routes leave too much of that work to people.
Specify its permitted sources and tools, a stopping condition and the point where responsibility returns to a person. For this example, the first version could gather evidence and propose a route. Changing the subscription or sending a customer response would be separate capabilities. Our article on permissions and approvals explains how to define those action boundaries.
An agent therefore needs an evaluation of the investigation it performs, including unnecessary steps, missed evidence and cases where it should stop.
Compare the work that remains
Before approving a larger build, agree on a small, representative set of permitted or anonymized requests. Include ordinary cases, incomplete messages and exceptions. Ask the team to define what a useful result looks like.
Compare the candidate designs on the same examples:
- Did the request reach the correct team with accurate context?
- Which cases needed a person, and was that handoff clear?
- How much time went into review, correction and investigating failures?
- What ongoing model, integration and support work would the design require?
Decide what is acceptable before looking at a polished demonstration. A route that looks impressive but leaves the team correcting every task has not yet demonstrated that it removes work. Our readiness checklist can help identify missing inputs before a pilot.
Define the first build around one decision
Useful scope statements describe the work: “Route complete forms using approved rules,” or “Suggest an email category for review.” They also identify what remains manual and what evidence would justify expanding the system.
At Madency, the proposed starting point is a review of that process: inputs, decisions, exceptions, integrations and deployment requirements, including whether it needs to run on your infrastructure. Discovery is $500 and is credited toward a subsequent Build; it defines scope rather than promising a completed integration for that price. See how we work.
Bring one recurring process and the point where the next step becomes unclear to a fit call. We can assess whether Discovery is the right next step and which implementation is worth testing first.
Technical background
Anthropic's Building effective agents distinguishes predefined workflows from systems in which a model selects its next steps. The n8n Production AI Playbook describes AI steps alongside rule-based orchestration. The customer-request comparison above is Madency's hypothetical application of these principles.
