Start with the workflow, not the model
Write down the repeated task in normal language. For example: enquiries arrive, someone reads them, checks service fit, asks for missing information, drafts a reply, and updates a tracker. That workflow is easier to improve than a vague request for an AI chatbot.
Check the data and documents
AI can only help if the information exists somewhere useful. Gather policies, price lists, PDFs, FAQs, spreadsheets, CRM exports, inbox examples, and previous proposals. If the source material is messy, the first sprint may need to clean the knowledge base.
Decide what success means
Good measures are practical: minutes saved per enquiry, fewer missed follow-ups, faster quote preparation, clearer management reporting, better response quality, or fewer repetitive admin steps.
Know where human approval belongs
For most businesses, AI should draft, summarise, search, classify, or recommend. A human should approve anything involving money, legal wording, safety, client promises, or sensitive decisions.
Choose a sensible first build
A good first AI implementation might be a private assistant, document Q&A tool, lead triage helper, internal dashboard, or automated report. It does not need to be a giant platform.
Quick checklist
- One repeated workflow has been chosen
- Example inputs and outputs are available
- A human owner is responsible for approval
- Success can be measured in time, accuracy, response speed, or revenue support
- Privacy and access rules are understood