Websites kick sand in the face of "agentic AI"

Websites focus on sales, service and information. They were designed to entice humans to view, discover, consider, buy. They are heavy with images, meta data, HTML, https, JavaScript, AJAX etc to entice us in and encourage us to click-around and dwell. APIs were invented for M2M, machine-to-machine, transactions. They are lean, measurable, and secure. “Agentic AI” is still in its infancy. LLMs are still based on next-best-word, statistical, transformer-based Machine Learning engines. They have no internal real-world model. They don’t know up/down, right/left, before/after, time, distance, etc, and especially numbers. There is no such thing as next-best-number. They just mimic based on training data. They are artificial cleverness, not artificial intelligence. “Agentic AI” is not a new AI advance; it’s just an LLM trying to orchestrate other LLMs. They fumble around, potentially creating a dangerous mess. For heavy duty, M2M, use deterministic RPA eg to migrate data from one PMS to another over a few hours. APIs can help if available, but RPA can read “green screens” (using Predictive AI), and move data from field A in system X, split it into a first name and second name, and enter these values into fields B and C in system Y, plus do calculations/validity checking on the fly. An “agentic AI” agent might start okay, but fail when it encounters Mr Stop and starts entering DOB in American date format into UK-formatted arrival date fields, or deleting entire guest profiles to speed up its task, or booking out 100 rooms to the same guest for efficiency. PMS, POS, IBE, Spa etc systems aren’t going anywhere. They are based on solid deterministic programming and relational databases. They already have solid APIs for interconnectivity. Traditional systems may gain Gen AI summary boxes, but the heavy lifting and data movements under the hood will remain under the control of rules-based, deterministic programming to provide reliability and consistency.

The bigger platform players like Oracle, Access, Agilysys are adding an AI layer on top to make functional modules feel more linked, but more importantly to add help features for staff, eg eg how to combine profiles, execute a refund, book adjoining rooms etc. This is likely the main difference users will notice.

The academic advance needed may come with neuro-symbolic AI, mixing rules-based and statistical prediction approaches, but this is hard and isn’t getting funding, although keep an eye on Yann leCun's (ex Meta) new venture. For now, focus on RPA and human-designed process flows.