Find the work that repeats.
Start with a specific task: categorising incoming enquiries, preparing a draft response or moving approved information between systems. Document the trigger, required input and expected output. If the task has no stable process or reliable information, clarify those foundations before adding AI.
Choose rules before intelligence.
A deterministic rule is often the right answer for routing by postcode, sending an acknowledgement or checking a required field. AI can help with unstructured text, summarisation and drafting. We scope each step according to what it needs, so the workflow remains explainable and maintainable.
- Enquiry classification and routing
- Draft responses for human approval
- Website knowledge assistants
- CRM and internal tool integrations
Design the exception path.
A useful automation must know when to stop. Low-confidence answers, missing records and failed integrations should move to a human queue. Define logs, retry behaviour, access permissions and a manual fallback. Test with realistic edge cases before connecting a production workflow.
A controlled path to rollout.
Discovery maps the existing process and data access. A narrow pilot tests whether the proposed workflow saves useful effort. Review output quality and failure cases, then expand only after the pilot meets agreed criteria. Third-party subscriptions and usage costs are separate scope decisions.
Keep customer information purposeful.
Use only the information a task needs. Avoid sending sensitive records to a model without an appropriate review of permissions and provider arrangements. Customer-facing assistants should explain their role and provide a direct contact route rather than invent answers about prices or project commitments.
Questions worth asking.
Do we need an AI chatbot?
Not necessarily. A clear form or simple routing rule may solve the problem more reliably. The workflow determines the tool, including whether AI is needed at all.
Can automation use our current systems?
Potentially. Integration depends on the systems’ APIs, permissions and data quality. These are checked before committing to a production scope.