Most discussions about AI agents start from the wrong premise: that an agent should be intelligent, reason autonomously, and make decisions. This is precisely what makes most agents unfit for real business use.
In this pricing system, the agent does not decide anything. It does not “think”. It does not optimize.
The decision is made entirely by a deterministic pricing model trained on historical data and market context, as described in parts one to three of this series. Given the same inputs, the model will always produce the same output.
The agent’s role is deliberately limited:
- to collect missing inputs,
- to call the decision engine,
- and to explain the result in clear business terms.
In other words, the agent is dumb by design, and that is a feature, not a limitation.
Humans remain in the loop not to compete with the system’s intelligence, but to validate it. In practice, this means accepting recommendations far more often than rejecting them, precisely because the system is predictable, auditable, and grounded in data rather than intuition.
Only after establishing this principle does it make sense to talk about real-world usage.
Every New Year’s Eve, my family and I spend a week in Lagos, Algarve. During my most recent booking, a short conversation with the hotel manager — now repeated for the fifth year in a row — made something very clear: she had no real visibility into how prices were set.
“Booking decides,” she said.
That moment is what ultimately motivated this series.
I’ll now show how to implement this agent in a Canvas App. This is intentionally simple. If it feels underwhelming, that’s the point.
You may create a “ChatGPT-like” experience in Power Apps by combining a flexible-height gallery with a text input control and a submit button. With some UI refinement, the experience can look professional and be embedded in a web screen or mobile device.
Here’s the overall look and feel of the chatbot:
There is very little training required for users to interact with the chatbot — and that is one of its greatest strengths. You simply chat with it, and it works out of the box.
It has been designed to collect all the data required to feed the pricing model, which then responds with a recommendation or asks for additional information when inputs are incomplete.
Here’s the final recommendation:
We recommend setting the price for the sea view rooms at €100, which is competitively lower than the current competitor price of €150. At this price point, we anticipate an occupancy rate of 86.22%, resulting in an expected revenue per room of €86.22. This pricing strategy is designed to maximize occupancy during the low season by offering an attractive price that encourages bookings, while still maintaining a healthy revenue per room. By positioning ourselves below the competitor’s rate, we aim to capture a larger market share and optimize room utilization.
Naturally, the recommendation is based on the trained model, using the hotel data described in part 2 of this series.
Whenever market conditions change, or when hotel policies are adjusted, new data must be collected and the model fine-tuned. Its accuracy must also be evaluated continuously over time.
The agent is predictable, obedient, and intentionally limited by its conversational scope. One of its most valuable features is its ability to explain why a given price has been recommended.
Even when a recommendation feels counter-intuitive, the explanation provides a foundation for understanding the underlying drivers and for fostering informed discussion within the management team.
Because the agent never decides, every recommendation is fully traceable: inputs, model version, price grid tested, and final outcome. This traceability is what makes the system viable for real businesses — not just demos.
I did speak again with the hotel manager. The hotel was almost empty. The property is pleasant, the staff polite and kind. Yet the pricing issue remains evident.
With better pricing intelligence, they could significantly improve both customer loyalty and overall profitability.
Want a tailored pricing strategy for your Company?
Contact Swell AI Lab to schedule a quick diagnostic.
