Section 1
The three reasons that hold up
Latency: a quality inspection at line speed, a safety interlock, a vehicle system. If a decision must land in tens of milliseconds, a network round trip has already spent the budget. Connectivity: a rural clinic, a mine, a ship, a delivery van on a bad route, a factory whose link drops twice a week. If the process must continue while the connection is down, the intelligence has to be local. Data residency and privacy: footage, patient records, or personal data that regulation or contract says must not leave the premises. Processing locally and transmitting only a count or a flag is often the difference between a project being permitted and not. If none of the three applies, use the cloud. Scoping this like any other product decision is covered in [AI Automation in Product Development: From Idea to Launch](/blog/ai-automation-in-product-development-from-idea-to-launch).
Section 2
What you are actually taking on
The model is the easy part. The commitment is a fleet. Every deployed device is a small computer in a place you do not control, in dust or heat or a cupboard, needing power, updates, monitoring, and eventually replacement. Fifty sites means fifty of those. When one behaves oddly you cannot open a log in a browser, because the log is on the device and the device is four hundred kilometres away. Edge projects that succeed treat device management as the primary engineering problem and the model as a component. Before committing, answer three questions concretely. How do you update a model on every device without visiting them. How do you know a device has stopped working. What happens to the work when one fails.
Section 3
Design for local first, cloud for the rest
The workable pattern splits the job by clock speed. Fast, narrow decisions run locally: detect, count, classify, alert, act. Slow, wide analysis runs centrally: retraining, cross-site comparison, dashboards, trend reporting. Devices then send summaries rather than raw data, which cuts bandwidth cost dramatically and is often what makes the privacy position defensible. Two design rules save most of the pain. Devices must work offline and reconcile when the link returns, so buffer locally and never assume a connection. And treat the model as versioned software with staged rollout, because updating every device at once is how a whole estate fails simultaneously. The cost framing behind these choices is in [The ROI of AI Automation: What Founders Need to Know](/blog/the-roi-of-ai-automation-what-founders-need-to-know).
Section 4
Pilot in a way that tells you something
One site is not a pilot. It is a demonstration, and it will succeed because the founder is standing next to it. Run three sites that differ in the ways your estate differs: newest and oldest, best connected and worst, busiest and quietest. Run for a full operating cycle, including any seasonal or shift variation. Measure two things that get skipped. How much staff time each site spent on the system, since that decides whether a rollout is tolerable at scale. And how it behaved during the week something went wrong, because that week represents the next three years.
Section 5
Physical devices, physical risk
Governance at the edge has concerns the cloud version does not. A device can be stolen, opened, or tampered with. Anything stored on it should be encrypted, and credentials should be per device and revocable. Be deliberate about what the device retains. A camera that stores nothing and reports a count is a very different conversation with staff and regulators than one that keeps footage, and the cheaper option is usually the safer one. The forgotten control is decommissioning. Devices are removed, replaced, and sent for repair, each carrying data and access. Keep an inventory with a named owner, and a written process for wiping a unit before it leaves your control.
Section 6
What the research says
Founders evaluating edge and agentic deployments should weigh genuine momentum against documented hype. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls, and estimates that of the thousands of vendors claiming agentic capability, only about 130 offer the real thing, a pattern it calls agent washing (Gartner, 2025). The economics, however, are moving in founders' favor: inference costs for GPT-3.5-level performance fell roughly 280-fold between late 2022 and late 2024 (Stanford HAI, 2025), which is precisely what makes running models closer to the point of work, on devices, in branches, at the edge, financially plausible for smaller firms. The failure data still applies. More than 80 percent of AI projects fail, twice the rate of comparable IT projects, most often because the problem was misunderstood before the technology was chosen (RAND, 2024), and Gartner separately predicts 60 percent of AI projects will be abandoned through 2026 if not supported by AI-ready data (Gartner, 2025). With only 39 percent of organizations reporting enterprise-level EBIT impact from AI (McKinsey, 2025), the founder playbook is unchanged: pilot edge automation where latency, privacy, or connectivity genuinely justify it, not because the architecture is novel.
Section 7
The numbers that decide a rollout
Start with the operational measure the deployment exists to move: detection time, downtime avoided, waste reduced, throughput, or hours of manual checking removed. If that number is not moving at three sites, more sites will not help. Then track the fleet metrics that determine whether scale is survivable. Percentage of devices reporting healthy. Mean time to detect a failed unit. Success rate of remote model updates. Support hours per site per month. Bandwidth and cost per device. Watch that support hours figure closely. Multiply it by your target number of sites and see whether the result is a job you are willing to fund. Many edge projects are technically sound and operationally unaffordable, and the arithmetic shows it long before the rollout does. Explaining that trade-off to a board without losing the room is its own skill, discussed in [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation).