Signaling Theory Explains Why AI Vendors Overpromise and Buyers Underverify

I read the Spence model again last night, not the original 1973 paper because I do not have it in my library, but the game theory textbook that walks through it. The game theory text in my study-hub files lays out the full formal structure of job market signaling. Nature determines a worker's ability. The worker chooses a level of education. Firms observe the education but not the ability, then make wage offers. The crucial assumption is that low-ability workers find signaling more costly than high-ability workers. The marginal cost of education is higher for the worker who gets less benefit from it. That cost differential is what makes the signal work. Without it, everyone would acquire the same level of education regardless of ability, and the signal would tell the market nothing.

The thing that hit me was how cleanly this maps onto the AI vendor market, and how badly the mapping breaks down at exactly the point where it matters most.

Think about what an AI vendor signals to a prospective buyer. Benchmark scores. Demo videos. Case studies. Partnership badges. Gartner quadrant placement. Certification programs. Each of these is a signal in Spence's sense: an observable action that is supposed to communicate unobservable quality. The vendor's actual model quality, its reliability under production loads, its failure modes on edge cases, its data handling practices, these are the things the buyer cares about but cannot directly observe. The signal exists to bridge that gap.

The problem is that in Spence's model, the signal works because it is differentially costly. A college degree is costly for everyone, but it is less costly for high-ability individuals because they can complete it with less effort, less time, less forgone earnings. The cost is what separates the high-ability types from the low-ability types. In the AI vendor market, the signals are not differentially costly. A benchmark score costs a vendor a few engineering hours to produce. A demo video costs a product marketing team a week. A Gartner quadrant position costs money but it costs the same money regardless of product quality. A certification badge from a cloud provider requires filling out forms and running compliance scripts. None of these signals cost meaningfully more for a low-quality vendor than for a high-quality one.

This is exactly the condition that produces a pooling equilibrium in signaling theory. When signals are cheap to produce and hard for the receiver to verify, both high-quality and low-quality types send the same signal, and the receiver cannot distinguish between them. The market does not sort itself. Everyone has the same badge, the same benchmark, the same demo. The buyer stares at two vendors who look identical on paper and has no way to tell which one will hold up under real conditions.

I wrote about how agent washing turns this into a structural problem. Gartner estimated that out of thousands of vendors marketing themselves as agentic AI providers, only about 130 are genuinely agentic. The rest are rebadging existing automation tools with the word "agent" attached. From a signaling perspective, this is what a pooling equilibrium looks like when the cost of sending the "we are an AI agent company" signal approaches zero. There is no technical barrier to claiming the category. There is no certification that verifies it. There is no industry standard definition that gates who can use the label. The signal costs nothing, so everyone sends it, and it tells the buyer nothing.

The buyer side of the equation compounds the problem. In Spence's model, the receivers (employers) are rational and competitive. They update their beliefs correctly based on the signals they observe. In the AI vendor market, the receivers (enterprise procurement teams) are not evaluating signals in a competitive market equilibrium. They are under institutional pressure to adopt AI, they lack the technical depth to evaluate model architecture claims, and they are making six- and seven-figure decisions on technology that is changing faster than their evaluation processes can track. As I noted when discussing agent washing and vendor markets, the procurement problem is that most buyers cannot architecturally distinguish a genuine agentic system from a well-scripted automation. That is not a buyer failure. That is exactly the information asymmetry that signaling theory predicts will produce market failure when signals become cheap.

A separating equilibrium requires a signal that is costly enough that low-quality types choose not to send it. In the job market, education is that signal. In the used car market, Akerlof's lemons problem is solved by warranties and inspection certificates that are too expensive for a seller of a genuine lemon to credibly offer. In the AI vendor market, I am not sure what that signal looks like yet. The closest thing I have seen is the pattern where vendors provide production access for extended evaluation periods. A vendor confident in its system's performance under real enterprise conditions can afford to let a buyer test it for months. A vendor selling a chatbot with an agent label cannot, because the gap between the demo and production performance becomes obvious quickly. Extended access is costly for the vendor in engineering time and support, and it is costlier for a low-quality vendor because the evaluation will reveal the gap. That cost differential is what makes it a credible signal.

The IS research angle here is that signaling theory has been applied in our field before, mostly in the context of IT investments and organizational communication. Li et al. (2023), in their MISQ paper on IT and information security that I have in my study-hub files, use signaling theory to explain how visible IT security investments communicate organizational commitment to stakeholders. The mechanism is the same one Spence described: organizations make observable investments that are costly enough to serve as credible signals of unobservable quality. But the IT security context is different from the AI vendor context in one crucial way. An IT security investment is hard to fake. You either have a SOC, a certified incident response team, and documented security practices, or you do not. The signal is tied to observable organizational capability. In the AI vendor market, the signals are detachable from the underlying capability. You can claim agentic functionality without having it, and the buyer cannot verify the claim without running the system in their environment for weeks.

This detachability is what makes the AI vendor market's signaling problem different from the classic Spence formulation. In the original model, the signal, education, is produced by the same agent whose quality it is meant to communicate, and the cost of producing the signal is what makes it credible. In AI procurement, the signal, benchmarks, demos, certifications, is produced by the vendor but evaluated by a buyer who cannot observe the production process. The buyer cannot tell whether the benchmark was cherry-picked, whether the demo was scripted, or whether the certification reflects genuine capability or a compliance checkbox. The signal is observable. Its credibility is not.

I think the IS field should study this as a market failure problem, not just a vendor ethics problem. Agent washing is not primarily about dishonest vendors. It is about a market structure where the cost of sending quality signals is too low relative to the information they are meant to convey. The fix is not moral exhortation. It is making signals more costly, more verifiable, or both. Industry standards for what counts as agentic AI would raise the cost of claiming the category incorrectly. Third-party evaluation frameworks would make signals more verifiable. Extended trial periods would force vendors to bear the cost of their claims in production, not just in staged demos. Each of these is a mechanism that moves the market closer to a separating equilibrium, where high-quality and low-quality vendors can be distinguished by the signals they send.

I wrote about how technology choices function as signals and how that creates adoption for signaling reasons rather than functional fit. I also wrote about how AI vendor concentration creates resource dependency that locks buyers in. Signaling failure makes procurement worse. Resource dependency makes exit harder. Together, they describe a market where buyers cannot accurately evaluate what they are buying and cannot easily leave after they discover the gap. That is not a good market. That is a market that needs better signals, and IS research is well positioned to design them.