The most useful decision in an automation project is what the agent must not do. Write the boundaries first and the rest of the design gets simpler. The steps below produce a one-page specification to review with your support, legal and marketing leads before anything goes live.

Step 1: Choose tasks by consequence

List your most common conversations, then ask two things about each. How often does it happen? What happens if the answer is wrong?

  • Good first tasks are frequent and easy to correct: opening hours, branch locations, order status, collecting details before a person takes over.
  • Keep with people for now anything where a wrong answer costs money, health or trust: refund disputes, complaints, credit decisions, medical or legal questions, and customers in distress.

Generative systems add a specific risk. NIST's generative AI profile uses the term confabulation for confidently stated but false content. A rigid menu fails by not understanding; a language model can fail by sounding sure and being wrong. Tasks that need exact prices or policy terms should draw on content you control, and you should test them against those facts.

Limit what the agent can do as well as what it can say. OWASP's guidance on prompt injection says it is unclear whether fool-proof prevention exists, and it recommends minimum access plus human approval for privileged operations. Reading an order status is one thing. Issuing a refund because a chat message asked for one is another. Start with read-only access.

Step 2: Tell people what they are talking to

Say it is automated at the start. Nielsen Norman Group's 2026 work on site chatbots reports that users prefer upfront disclosure and advises, at a minimum, a persistent indicator. A line such as "I'm this shop's automated assistant. I can check orders and opening hours, or get you a person" sets the scope and the exit in one message.

Rules are moving the same way:

  • The EU AI Act's Article 50 duty to inform people they are interacting with AI has applied since 2 August 2026. It falls on providers, and the European Commission says providers outside the EU are covered if the system's output is used in the EU. If you serve EU customers, check how it reaches your setup.
  • Meta's Messenger and Instagram policy requires disclosure where law demands it: at the start of a thread, after a significant lapse of time, or when a chat moves from a person to automation. It recommends disclosure even where law does not require it.

That last trigger is easy to miss. If a person hands a conversation back to the bot, say so.

Explain data use inside the conversation too, because Nielsen Norman Group notes that users do not read privacy policies. If you ask for a phone number, say why. In Kenya, where you rely on consent, the General Regulations require you to mention automated decision-making where relevant, and section 35 of the Data Protection Act limits decisions based solely on automated processing that significantly affect someone. An agent that declines a loan is a very different thing from one that shares opening hours.

Step 3: Write the stop conditions

A stop condition is the point where the agent stops trying and hands over. These are starting points to tune against your own transcripts:

  1. The customer asks for a person, in any wording or language. Nielsen Norman Group's guidance is never to gatekeep this request.
  2. The agent has failed to understand twice on the same point.
  3. The customer keeps rephrasing the same question or shows frustration.
  4. The topic is on your people-only list.
  5. The agent lacks the source or system access it needs for an accurate answer.
  6. A safety concern, a threat or a sign of a vulnerable customer appears.

Balance matters. The same guidance warns that escalating too quickly can undermine trust in the bot, so the agent can try one honest repair first, such as offering options, unless the customer has asked for a person.

On WhatsApp this is also a policy matter. The WhatsApp Business Messaging Policy allows automation within the 24-hour window but requires prompt, clear and direct escalation paths, listing options such as in-chat transfer to a human agent, a phone number, email and a support form.

Step 4: Design the handover

A handover fails when the customer has to start again. Microsoft's Copilot Studio documentation shows what a good one can carry: the conversation history, the variables collected and an optional private message to the agent. Whatever platform you use, pass:

  • The transcript so far.
  • A two-line summary of what the customer wants and what has been tried.
  • Details collected, such as order number and whether identity was confirmed.
  • The reason for the handover, taken from your stop conditions.

Tell the customer who is next and how long it should take. Outside staffed hours, say so honestly and take a message. Once a person has taken the conversation, the bot should stay out of it. The shared inbox guide covers ownership from there.

Step 5: Review transcripts on a schedule

Each week, read a sample from four groups: conversations the agent closed, conversations it handed over, conversations abandoned midway, and any that drew a complaint. Mark each one:

  • Was the answer correct, checked against your source?
  • Did it stop too late, or too early?
  • Did the agent promise anything?
  • Did the person taking over have to ask for something the agent already had?

Fix causes in order: wrong content first, then missing stop conditions, then wording. Keep a set of test conversations and rerun them after any change to prompts, content or the model. Transcripts are personal data, so limit access. Kenya's General Regulations require a retention schedule.

A note on WhatsApp and AI vendors

Meta's terms restrict AI providers from using the WhatsApp Business Platform where AI is the primary functionality on offer, as Meta determines, while its pricing documentation recognises service messages powered by third-party AI. If an AI vendor acts as your solution provider, the terms bar use of platform data to train AI models, apart from fine-tuning a model for your exclusive use. Read the current terms with your vendor before you build.

Where Jadility fits

Jadility's AI and agents capability covers conversational flows and agents with a clear route to human support, and the shared team inbox gives your people one place to reply. If you have a task list and stop conditions drafted, talk to us. The IVR handoff guide applies the same ideas to voice.

Sources

Share