The three-question scorecard
Before we add AI to any step in a client's flow, it has to pass three questions:
- Is the input unstructured? Free text, voice, documents, a customer's rambling message. If the input is already structured (a form field, a date, a stage) a rule will beat a model.
- Is a small mistake cheap? A slightly off summary costs nothing. A wrong appointment time or a wrong price costs a customer.
- Will someone see the output before it matters? Drafts a human sends are safe. Messages a bot sends unsupervised need much tighter guardrails.
Two or three yeses and AI is a good bet. One or none and we build a rule, even if the client asked for AI.
Where AI earns its place
- Qualifying inbound leads. Reading a messy enquiry and tagging service, urgency and fit. Saves the team the first five minutes of every conversation.
- Drafting replies. A suggested response in the conversation view that a human edits and sends. Faster than typing, safer than sending blind.
- Summarising calls and threads. Two-line summaries on the contact record so nobody re-reads forty messages before a follow-up.
- Booking conversations with a narrow brief and firm hand-off rules, as covered in the bot article.
- Classifying documents that arrive as PDFs and photos: invoices, forms, IDs.
Where a rule wins
- Anything with a price. Quotes, discounts, payment terms. Look them up, don't generate them.
- Scheduling logic. Availability, buffers, time zones. The calendar already knows; the model will guess.
- Permissions and routing. Who sees what, which rep gets which lead. Deterministic rules you can audit.
- Compliance messages. Opt-out handling, consent, legal notices. Exact wording, every time.
- Anything the customer will hold you to. If they can screenshot it and quote it back, it should come from a template.
Ask: "If this went wrong, would I explain it to the customer as a typo or as a breach of trust?" Typo-level risk is fine for AI. Trust-level risk is a rule.
The cost nobody budgets for
Model usage is cheap. Review time is not. Every AI output that a person has to read and approve costs a minute or two, and a busy account generates hundreds a week. When we scope AI features we estimate the review load first and often discover that a well-designed template plus a rule gets 90 percent of the value with none of the review.
The other hidden cost is drift. Prompts that worked in March quietly stop matching how the business talks in September. Someone has to own the prompt the way someone owns the pipeline.
Where to start
Pick one job the team already finds tedious, where the output is reviewed before it reaches a customer. Lead qualification or reply drafting are the usual first wins. Run it for a month, measure the minutes saved and the corrections needed, and only then move to something customer-facing. A single AI feature that the team trusts is worth more than five they quietly switch off.
Key takeaways
- Use AI where the input is messy language and the cost of a small mistake is low.
- Use a rule where the input is structured and the answer must be exact, like pricing, scheduling logic and permissions.
- Reading, summarising, drafting and classifying are the high-return jobs. Deciding is not.
- Budget for review time. Every AI output someone has to check costs minutes, and minutes add up.
- Start with one visible win, usually lead qualification or reply drafting, before touching anything customer-facing.