Most of the conversation about AI agents is about the agent: how well it reasons, how many steps it can take, how much it costs to run. Much less is said about what the agent is reasoning over. For software that writes code or summarizes documents, that is fine. The inputs are already digital.
It stops being fine the moment an agent has to act in the physical economy: book the room, reroute the truck, schedule the crew, decide whether the site is safe to open. There, the limiting factor is not intelligence. It is whether the facts are reachable by a machine at all.
Agents are being handed real decisions
The forecasts are large. They are forecasts, and should be read as such.
- Gartner predicts at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024
- The same release expects 33% of enterprise software applications to include agentic AI by 2028
- McKinsey estimates agentic commerce could orchestrate up to $1 trillion in US retail revenue by 2030, and $3 to $5 trillion globally
Gartner's release carries a second number that gets quoted less often: it expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Both halves can be true. Demand for agents that act is real. So is the rate at which they fail once they leave the demo.
Opinion: a large share of those failures will be information failures, not reasoning failures. An agent that acts on a wrong, stale or missing fact makes a confident bad decision. No model upgrade fixes that.
Commerce already hit this wall
Online retail is the most digitized part of the economy, and even there the first thing the industry had to build was a way for machines to read it. In September 2025, Stripe and OpenAI released the Agentic Commerce Protocol alongside Instant Checkout in ChatGPT. Stripe's announcement is direct about the requirement: "businesses must expose their products, pricing, and checkout in a way agents can use."
Read that as an admission. Websites built for human shoppers were not enough. For agents to buy, merchants have to publish a second, machine-facing version of themselves.
That works for a product catalog. A product has an identifier, a price and a stock count, and one owner who controls all three. The physical economy has none of those conveniences.
Why the physical world is harder than a catalog
It changes by the minute. Road conditions, wait times, air quality, availability and grid status are not listings. They are measurements, and they expire.
Nobody owns the whole answer. One decision about one place can depend on a public agency, a private operator, a sensor network and a commercial feed, each with different access rules.
Most of it was never published for software. It sits in pages written for human eyes, in operational systems with no outside interface, or nowhere.
We have measured each of these. The samples are small and we report them with their limits, but the pattern has been consistent.
What we found when we looked
Uneven by field. We audited 44 hotels across 110 pages, asking field by field whether a machine could read the answer. Hospitality is a useful test because it is commercially motivated to be found.
- Address 100% · Phone 90.9%
- Check-in time 9.1% · Pets 2.3%
- Parking fee 0% · Breakfast 0%
The facts that identify a business are machine-readable. The facts an agent needs to complete a transaction mostly are not. One sample, one domain, but it shows where the gap sits: not at the level of the source, at the level of the individual fact.
Unstable at the source. We probed 32 real-world information sources across 17 verticals, from marine and flood data to transit and EV charging. 15 were open with no credential. 4 needed a free key. 5 were private. 8 were unreachable or had changed shape since they were documented. A quarter of sources quietly breaking is not an edge case. For any team integrating these one by one, it is a permanent maintenance cost.
Perishable. We replayed 15,318 paired transit observations with controlled delays. As the information aged, error rose. The direction was consistent in this replay, and the size of the effect depends on the feed and the field. An agent does not only need a source that has the answer. It needs to know how old the answer is.
The agent is ready to act.
The world is not ready to be read.
What this means if you run something
This section is opinion.
If you are building agents, your ceiling is your information supply, and it deserves the same attention as model choice. Every real-world source you wire in directly is a dependency that can change format, go dark or go stale without telling you. Ask of every fact your agent acts on: where did it come from, how old is it, and what happens when that source fails?
If you operate physical assets, sensors, fleets, venues, networks, your information is about to get a second customer. For decades its consumer was a person on a website or an operator on a dashboard. Machines will ask for it continuously, and the organizations that carried the cost of producing it should decide the terms on which it is consumed.
If you allocate capital, the pattern is familiar. When a new kind of consumer arrives, the layer that connects it to fragmented supply becomes infrastructure. Payments needed it. Commerce is building it now. Real-world information does not have it yet.
Where PlaceRouter fits
PlaceRouter is building that layer for real-world information: discover the sources that can answer a need, evaluate them, route each field to the best source for that place and moment, verify the answer, and learn from the outcome. One integration for the application, instead of thousands.
We test the thesis before we sell it. We publish our results whether or not they support the thesis, including one where our own routing lost. We think that is the only credible way to build infrastructure other people will depend on.
AI can reason.
PlaceRouter connects it to the world.
PlaceRouter figures come from our research page, reported with method and limits. Third-party figures are forecasts from the linked sources: Gartner, June 2025; McKinsey, The agentic commerce opportunity; Stripe, September 2025. If your systems depend on real-world information, or produce it, we would like to talk.