Talk is cheap - but supply chains don't talk

I spent this weekend reading Steven Pinker's When Everyone Knows That Everyone Knows. One idea struck a cord; specifically the idea of "cheap talk." In game theory, cheap talk is communication that costs nothing to send, where no one wagers anything by speaking. In a so-called coordination game, where two actors have aligned incentives, cheap talk is enough. The two friends who want to meet at one of two restaurants just need to send a text, and the problem is solved.

Supply chains are supposed to be such a coordination game. A customer benefits when its upstream supplier knows what is coming. The supplier benefits when it can attune capacity, raw materials, and inventory to actual demand instead of a guess. Both end up with higher service levels at lower cost. So why, after thirty years of EDI portals, cross-company ERP integrations, and weekly forecast meetings, do most of the supply chains still operate on bad information?

I have two hypotheses.

Two restaurants, one dinner

The classical coordination game example are two friends wanting to go for dinner. There are two restaurants in town. If they both pick the same one, they have a good evening. If they pick differently, they have a sad one. Their incentives are fully aligned: neither of them prefers one restaurant strongly enough to care which one wins, as long as they end up together. A single text resolves the problem. The signal costs nothing, the speaker has no reason to lie, and the listener has no reason to discount what they hear.

Written as a payoff matrix, the structure looks like this:

The interesting feature is that there is no dominant strategy. Both players want the same outcome but cannot reach it on their own. The shared signal is the entire solution. This is what cheap talk is built for. In a pure coordination game, communication is enough.

Now translate this to a frozen-vegetables producer and its customer, a private-label retailer ordering ready meals six weeks out. If the retailer shares an honest view of expected sales, the producer can plan production runs, lock in raw material purchases, and avoid both stockouts and write-offs. The retailer in turn gets better service at a more stable cost base. Both win.

In the textbook, both then pick up the phone. In practice, they almost never do.

Where supply chains break the model

The information that actually matters rarely reaches the supplier. EDI handles the easy parts: stock on hand, sell-through, replenishment orders. What it does not handle is the part that drives short-term variation. A retailer that ran a stock clearance does not flag that sales spiked because of a promotion, not because of organic demand. A retailer planning to delist a SKU at the next season change does not communicate this to the supplier still building safety stock. A retailer that has a competing supplier on the bench almost never mentions it.

That last point is where the first hypothesis lives.

Hypothesis one: it is not really a coordination game

The assumption behind cheap talk is that incentives are aligned. In a lot of supplier-customer relationships, they are not, or at least not as cleanly as the textbook assumes. The customer has an interest in keeping options open. The cheapest way to do that is to make non-committal estimates to several suppliers at once.

Imagine a ready-meal manufacturer with two possible producers of a frozen vegetable mix. One is significantly cheaper but has long lead times. The other is more expensive but flexible. The optimal strategy for the manufacturer, if there is no cost to lying, is to tell both suppliers it intends to buy from them. The expensive flexible supplier holds inventory ready to be called off; the cheap supplier prepares a long-lead-time production run. The manufacturer picks whichever option turns out to fit best, and walks away from the other without paying for the optionality it just bought.

Written as a payoff matrix, the picture changes shape entirely:

In contrast to the restaurant case, the customer now does have a dominant strategy: always claim they need it. The downside of overclaiming is zero for them and -50 for the supplier. The supplier's rational response is to discount everything they hear. This is not coordination. It is a prisoner's dilemma in business clothing.

Where suppliers and customers face misaligned incentives, the standard model of cheap talk collapses. Information becomes strategic, and strategic information is rarely honest.

The good news, if you have read Axelrod's The Evolution of Cooperation, is that this game is almost never played one-shot. The same partners trade week after week, season after season. Strategies that consistently cooperate and punish defection tend to outperform purely selfish ones over time. Tit-for-tat is the canonical example: cooperate until cheated, then punish, then forgive. In supply chain practice, this looks like a supplier quietly reducing the weight it gives to a chronically over-promising customer in its forecasts, and gradually restoring trust when behaviour improves.

This hypothesis is real, and it explains some of what we see. I do not think it is the main story.

Hypothesis two: cheap talk is not cheap

The more honest hypothesis is that most customer-supplier relationships are coordination games, both sides know it, and the parties involved would happily share information if they could. The problem is that the relevant information is genuinely difficult and expensive to produce.

Think about what a supplier actually needs to know to be helpful. Not stock levels; those are easy, EDI solved that. What the supplier needs is the texture behind the numbers. Was last month's sales bump structural or promotional? Is the buyer planning to delist next quarter? Is there a weather forecast that changes the demand profile for ice cream or umbrellas? Did a competing supplier just go down, putting more volume on the line? Is the assortment shifting? Are price tests planned?

This information exists, but it lives in heads, in Slack threads, in planning meetings, in promotional calendars, in email exchanges with the marketing team. It almost never lives in the ERP. Even when an organisation knows perfectly well what is going on internally, surfacing it in a form that is shareable, accurate, and timely is a serious amount of work. It is the kind of work that gets postponed until the supplier explicitly asks, by which time it is too late.

This reframes the problem in a useful way. If suppliers and customers are not sharing information, the first instinct of most consultants is to install a governance layer: weekly forecast reviews, joint business planning sessions, S&OP integrations. These help at the margin, but they treat the symptom rather than the cause.

Especially because these usually take place at the supplier - without direct customer input at the table. The closest thing to a customer is a salesperson, who is likely to have imperfect information. The result is more planner time, more meetings, more reconciliation between fragmented systems.

What the supply chain literature has been quietly waiting for, I think, is a way to make expensive talk cheap.

Making expensive talk cheap

This is where I become moderately optimistic about generative AI, despite my generally tempered view of the hype. If you connect a language model to the systems where the relevant information already lives (the planner's mailbox, the promotional calendar, the meeting notes, the assortment plan, the competitor monitoring tool) you can synthesize the contextual information that EDI never captured. You can flag a delisting before the supplier builds stock for it. You can warn that a promotion is driving an apparent demand bump. You can summarize the operational state of the customer in a way the supplier can actually act on.

This is, in essence, EDI 2.0, with the things that mattered all along: intent, context, plans. The technical pieces have existed in fragments for years. What was missing was the cheap translator between unstructured internal noise and structured external signal. That translator is now plausible, and at a unit cost that is for the first time small enough to make the math work.

There are reasons to be cautious. These models can make mistakes. Even more importantly: they could potentially share information that you would rather not have shared. Building the pipes is not trivial, even if the model itself is. And the prisoner's dilemma part of the problem does not disappear just because the coordination part gets easier; cheap talk is only as honest as the speaker. I am as fallible as the next person on whether this lands in five years or fifteen.

The direction matters more than the timing. For thirty years, supply chains have been told to coordinate better and have mostly failed, not because they did not intend to, but because the cost of doing so quietly outstripped the benefits.

Talk was never cheap. It might finally become so.

Further reading

• Steven Pinker, When Everyone Knows That Everyone Knows: Common Knowledge and the Mysteries of Money, Power, and Everyday Life (Scribner, 2025). The source of the cheap-talk framing in this piece and a wider treatment of why public common knowledge changes behaviour even when private knowledge does not.
• Robert Axelrod, The Evolution of Cooperation (Basic Books, 1984). The canonical study of how cooperative strategies (notably tit-for-tat) win out over selfish ones in repeated games. Still one of the most useful books I know for thinking about long-running commercial relationships.