The Framing That Causes Bad Decisions
Vendors present this as a generational upgrade: rules are the old way, AI is the new way, migrate. That framing costs teams money, because the two approaches fail differently and the difference is what should drive your choice.
Rule-based automation fails visibly. It hits a case you did not define and falls back — the customer sees a bot that did not understand, which is annoying but obvious and cheap.
AI fails invisibly. It produces a confident, well-written, wrong answer. Nobody notices until a customer acts on it. Annoying is cheap; wrong is expensive.
So the question is never "which is more advanced". It is "for this specific message, which failure can I afford".
The Decision Framework
Run each use case through four questions.
1. Is the input structured or open?
Collecting an order number, a pincode, a date, or a choice between three services is structured. Use rules — buttons and list messages, not free text at all. A model adds cost and uncertainty to a problem a dropdown already solved.
"Why was I charged twice" is open. Rules cannot enumerate the phrasings. Use AI.
2. What does a wrong answer cost?
If the worst case is mild confusion — business hours, delivery areas — AI is fine. If the worst case is a refund you did not authorise, a missed medication warning, or a legal commitment, do not let a model generate it freely. Put it behind rules or a human.
3. Does the answer come from a document or a database?
Documents — policies, FAQs, product descriptions — suit AI well, because the task is comprehension and rephrasing.
Databases — order status, stock, balance, appointment slots — must be read, never inferred. Use a rule that queries the system and a template that states the result. This is the single most common failure in AI deployments: a model that sounds authoritative about an order it cannot see.
4. How often does the case occur?
High-volume and repetitive justifies building either. Low-volume and unpredictable is where AI earns its keep, because writing rules for the long tail never finishes. Low-volume and high-stakes should just go to a human — automating it is optimisation theatre.
Side by Side
| Dimension | Rule-based | AI-based |
|---|---|---|
| Handles unanticipated phrasing | No | Yes |
| Behaviour is auditable | Yes — read the branch | Only if you log per conversation |
| Failure mode | Visible fallback | Confident wrong answer |
| Cost per message | Effectively zero | Per-inference, scales with volume |
| Multilingual | One flow per language | One knowledge base |
| Setup effort | High per case, finite | Low per case, ongoing tuning |
| Safe for money and promises | Yes, when scoped | Not without hard constraints |
The Hybrid Pattern Most Teams Land On
Almost every mature deployment ends up with the same shape, and it is worth designing towards it directly rather than discovering it after two rebuilds:
- Rules at the entry point. A greeting with three or four buttons routes the obvious majority instantly — track order, buy something, support, talk to a human. No model needed to know which of four things someone wants.
- Rules for structured collection. Once routed, gather what you need with explicit steps and validation.
- System lookups for facts. Order status, stock, slots — queried and stated in a template.
- AI for the open middle. Anything that did not fit a button goes to a grounded model that answers from your documented material and escalates when unsure.
- AI assist for agents throughout. Draft replies and thread summaries, with the agent as the guardrail. This is the highest return per unit of risk in the whole stack.
- Humans on money, complaints and edge cases. Structurally, not by prompt instruction.
Note that AI appears twice, and the more valuable appearance is the one where a human presses send.
A Worked Example
A D2C brand answering 400 messages a day:
- "Where is my order?" — the largest bucket. Rule: ask for order number, query the system, send real status. No AI. It must be right, and it is a lookup.
- "Is this cream OK for sensitive skin, I get eczema" — open, and medical-adjacent. AI drafts from product documentation; an agent reviews before sending. Never autonomous.
- "Do you ship to Nagpur, and how long" — open phrasing, documented answer, low stakes. Fully autonomous AI.
- "I want a refund, this arrived broken" — money and a complaint. Straight to an agent, with an AI-generated summary attached.
- "What are your hours" — either works. Use a rule; it is free and cannot be wrong.
Roughly 60% of volume handled without a human, and the model never touches money or facts it cannot see.
Frequently Asked Questions
Should I replace my existing flows with AI?
No. Keep flows that work — especially structured collection and system lookups. Add AI where flows currently fall back, which is where you are losing conversations today.
Is rule-based automation obsolete?
No, and the framing is a trap. Buttons and list messages beat free text for structured choices on every metric: accuracy, speed, cost and customer effort. Meta's interactive message components exist because tapping is better than typing.
What is the cheapest way to start with AI?
Agent-assist. Let the model draft and agents send. Immediate benefit, near-zero risk, and it shows you where autonomous answering would be safe. See our guide to AI WhatsApp chatbots.
How do I stop AI inventing policies?
Ground it in your actual policy documents, instruct it to answer only from them, and put refunds and commitments behind human confirmation as a hard constraint rather than a request in the prompt.
Choose Per Message, Not Per Platform
The right answer is almost never all-rules or all-AI. Sort your top twenty incoming message types by "cost of a wrong answer" and "structured or open", and the assignment mostly writes itself. Build both in Tenreply — the flow builder and AI suggestions run in the same inbox.