Peter Kellis, Founder and CEO, TRAY.
For most of the past decade, the point-of-sale system has been treated as a solved problem—a commoditized terminal to be hollowed out, wrapped in APIs and pushed to the edge of the architecture diagram. Go headless. Let the ordering channels proliferate. Let the register become a dumb endpoint that takes money.
I understand the appeal of that thinking. I also think it is about to be proven wrong, and expensively so.
Here is the tension nobody in restaurant technology has fully reckoned with: We are pouring capital into AI—demand forecasting, voice ordering, dynamic pricing, computer vision, labor optimization—and every one of those systems produces a decision. None of them produces a sandwich. Something has to stand between the prediction and the physical world and make the operation actually behave differently. In a restaurant, that something is the point of sale.
There are three eras, and two of them are behind us.
It helps to see where the category has actually traveled.
In the first era, POS meant a register plus local workflows. Kitchen systems, digital ordering, loyalty and reporting were separate products with separate logins. The structural problem was that every new way to order became another operational exception: another tablet on the counter, another workflow the crew had to remember at 12:15 on a Friday.
The second era brought omnichannel. In-store and digital orders came to share menu and routing logic, and administration moved to the cloud. This was real progress, and most enterprise brands are living in it right now. But it solved integration, not execution. The data got connected while the work at the store stayed fragmented.
The third era is the one that changes the strategic weight of the category. Every order—lane, counter, kiosk, app, phone, third-party marketplace, AI voice agent—flows through a single operational layer that understands channel, timing, kitchen capacity, payment state, device telemetry and failure recovery. This is not a system that records what happened. It is a system that governs what happens next.
That is an intelligence execution layer, and it is a fundamentally more important asset than a register.
Prediction is cheap. Execution is the constraint.
The most common AI failure I encounter in operations is not a bad model. It is a good model with nowhere to land.
A forecast says Tuesday lunch will run 18% above plan. Fine, but what changed? If the answer is that a manager got a dashboard alert and was supposed to act on it during a rush, nothing changed. Multiply that across a few thousand stores and you have a substantial AI investment producing charts instead of throughput.
For that forecast to matter, prep quantities have to shift, station loads have to rebalance, an item nearing depletion has to be suppressed from the digital menu before the guest orders it, and the drive-thru has to be paced against actual kitchen capacity rather than a fixed timer. All of that is execution, and all of it happens in the operational layer that owns the order.
This is why I would argue the POS is becoming more strategically important as AI matures, not less. It is the real-time source of truth about what is happening in the store right now, and it is the only system with the authority to change the outcome.
Intelligence has to be tiered.
The instinct is to centralize AI in the cloud, where the big models live. That instinct breaks the moment latency and connectivity enter the picture.
Think in three tiers instead:
• Cloud intelligence handles what genuinely requires scale and depth: cross-store pattern detection, anomaly identification, menu and offer analysis, natural-language interrogation of enterprise data. These workloads tolerate seconds and benefit from large models.
• Store intelligence handles what has to be aware of local physical state: predictive prep, intelligent item suppression, routing across stations, pacing across channels. These decisions depend on kitchen capacity and timing data that only exists in the building, and they are worthless if they arrive late.
• Device intelligence handles what has to survive independently: order execution, guardrails, crew guidance, capture. This is the tier most architectures neglect.
There is an underappreciated economic argument for pushing work downward. The hardware already deployed in a typical store—terminals and kitchen displays—commonly runs around 10% CPU utilization even at full transactional load. Aggregated across every device in the building, that idle capacity is a meaningful distributed compute substrate the brand has already paid for. Using it beats provisioning, securing and maintaining dedicated edge appliances in thousands of locations.
Offline capability is an architectural position, not a feature.
Vendors list offline mode on a feature grid next to gift cards. It does not belong there.
If your intelligence only functions while connectivity holds, you have not built operational intelligence. You have built an operational dependency with a single point of failure sitting in a strip-mall network closet. Restaurants lose connectivity; the rush does not pause. An execution layer that degrades gracefully, keeps taking orders and keeps making locally informed decisions is expressing a design philosophy about where authority lives.
What does this change about the buying decision?
If the POS is becoming the layer where intelligence meets reality, evaluating it on transaction fees and hardware cost is a category error. Four questions I would put to any vendor:
1. Where does each class of decision execute, and what happens to each tier when the network drops?
2. Can a new ordering channel be turned on as configuration, or does it require an integration project?
3. Does the system expose real-time kitchen and device state, or only end-of-day reporting?
4. Can it be changed without downtime and a technician in a van?
The register is not going away. It is becoming the system your operation cannot run without. Choose it accordingly.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
