A guide by SevenLab
An AI agent is not automatically the best way to automate a process. Some tasks need fixed rules, others need help interpreting language, and some should stay with a person. Use this guide to scope an agent or workflow project with SevenLab and to compare custom development with simpler automation.
Choose the simplest automation that fits the work
Map the workflow before choosing technology. Stable rules and structured fields often suit conventional automation or an existing application. AI becomes worth evaluating when requests vary in wording, documents need interpretation or routing cannot be captured reliably in a short rule set. An agent adds further complexity when it selects tools or actions rather than only suggesting an answer.
Illustrative scenario: a small distributor receives orders by email. Fixed rules could route known senders, while AI might help classify free-text requests and suggest order fields. A conventional workflow can then validate those fields. Before replacing an existing RPA process, check whether the problem is genuinely interpretation or simply a brittle screen-based connection.
- Draw the workflow from trigger to completed business record.
- Mark which steps need rules, interpretation or human judgement.
- Test a simpler automation option before adding agent autonomy.
Define what an agent may do without approval
Describe an agent through its permitted actions, available information and stopping conditions. Drafting an answer is different from sending it, and preparing an order is different from committing a transaction. More autonomy increases the importance of permission boundaries, duplicate prevention, action records and a reliable route to a human. These are requirements to agree, not features to assume.
Illustrative scenario: a back-office team wants a digital worker to process incoming administration. A sensible first scope might allow classification and draft preparation while requiring approval for changes to financial records. The pilot should include missing attachments, contradictory instructions and requests that try to override its rules, not only clean examples chosen for a demonstration.
- Write an allowlist of actions and prohibited changes.
- Specify approval triggers, escalation owners and stopping rules.
- Include duplicate, malicious and incomplete requests in testing.
Separate support assistance from customer commitments
Customer support automation spans email classification, ticket routing, answers from manuals and draft quotations. Each task has a different risk profile. An incorrect category may be recoverable, while an incorrect technical instruction or commercial promise can cause serious harm. Evaluate source quality and escalation rules before extending a chatbot into order handling or sales follow-up.
Illustrative scenario: a technical helpdesk wants an assistant to suggest answers from manuals. It should check equipment versions, show the source and escalate uncertain or safety-related questions. Aegir Marine is a relevant SevenLab case to investigate for technical support through the cases page. That reference is not evidence that the same workflow, integrations or results will apply to another helpdesk.
- Separate internal suggestions from messages sent to customers.
- Require version checks and sources for technical answers.
- Reserve pricing commitments, order changes and sensitive advice for authorised review.
Budget for operation, exceptions and change
A workflow is only useful if the team can identify failures and recover work. Ask how failed tasks would be queued, how a person could take over and how model or prompt changes would be tested. Assess dependencies on CRM, email and order systems individually. Required access and available interfaces must be verified before any integration is included in the delivery scope.
Compare cost using the complete process: implementation, external usage, monitoring, human review and maintenance. Start with the existing pricing page and ask for assumptions tied to your workload. Evaluate handling effort, correction burden and unresolved work against a baseline. A high count of automated steps is not a useful success measure if staff spend more time repairing the outcome.
- Name the owner of failed tasks and the manual fallback.
- Agree regression tests before changing models, prompts or workflow rules.
- Estimate total operating effort and review the existing pricing information.
