Every team has an inbox full of questions it has answered before. Leads ask about pricing. Customers ask about onboarding. Candidates ask about documents. The answers live in one person’s head, old threads and long guides.
When budgets tighten and the volume keeps growing, the reflex is simple: put a chatbot on it. Doing more with less.
But automating a conversation is not the same as improving it. In a recent project, we built an AI copilot for a team that runs most of its conversations on WhatsApp. The lesson was clear: start with the knowledge, not the bot.
Humans decide. AI drafts.
A fully automated bot breaks the one thing that makes a service work: the personal relationship. It also adds risk. One wrong answer can cost a customer, a deal or weeks of delay.
So we designed a copilot, not an autopilot. For every incoming message, the AI reads text, voice notes and screenshots, and proposes reply drafts in the team’s own tone of voice. A person picks one, edits it and sends it. Nothing goes out automatically.
The copilot runs alongside the tools the team already uses. No new number, no new app, no new habit for customers. AI and automation should be connected to real action, not added as another silo.
The knowledge is the product
The model writes the sentences. Curated knowledge decides whether they are right.
Every draft is grounded in a reviewed knowledge base. When someone writes their own answer, the copilot proposes it as a new entry, approved or dismissed with one tap. The knowledge grows from real conversations, but nothing enters it unreviewed.
We deliberately avoided “self-learning”. Answers are often specific to one person, rules change and mistakes spread fast. Trusted data beats more data, every time.
Three pitfalls to avoid
1. Rules without memory
An “always mention” rule once made the copilot repeat something the team had already explained. Rules now apply once per conversation, and the AI always knows today’s date.
2. Silent failures
When one link in the chain fails, messages can disappear without anyone noticing. Store first, process second, and log every failure loudly.
3. Compliance as an afterthought
Processor agreements, a defined retention period and a privacy notice that tells customers AI helps prepare replies. Build it in from day one, not after launch.
Considering it? Start here
Before you choose a tool, answer three questions:
- Which questions does your team answer this week that someone already answered last week?
- Where does the knowledge behind those answers live today?
- Where must a person always stay in control?
The goal isn’t fewer conversations. It’s better ones, with more time for the people behind them.
Want to know how ready your data and processes are? Take the AI Readiness Quiz.
Which conversations in your team would you never hand to a bot?