Winking at machines: teaching AI to read between the corporate lies
AI can read the annual report. It cannot see the boardroom wink.
We have all seen the headlines. Eighty percent of leading companies already use AI in their decision-making. Large language models predict the next industry disruption. And yet most AI-generated innovation strategies break up on the rocks of corporate reality. Why? Because the machines keep missing what anthropologists call webs of meaning: the unwritten rules, the power dynamics and the tribal rituals that decide which ideas live and which die. A recent Johns Hopkins study found that AI models fall short at predicting social interaction, which happens to be the skill you need to get anything done in a real organization. For now, at least.
The thin fiction of corporate rationality.
Companies like telling shareholders they decide on cold, hard data, and AI takes them at their word, reading annual reports and press releases and suggesting "rational" innovation plays. In my experiments this works perhaps 60 to 70 percent of the time. The failures are the interesting part, because they show where the stated strategy collides with the culture. A pharmaceutical company's AI-recommended microbiome platform dies because the R&D chiefs are protecting small-molecule budgets. A retailer's perfectly sensible proposal for augmented-reality fitting rooms is killed by a CFO who has not forgiven the metaverse, and to be fair, who has.
These are not data problems. They are what the philosopher Gilbert Ryle called failures of thick description: the machine sees the twitch of an eyelid, a plant closure say, and misses the wink, which is a CEO's pet project being protected.
AI needs cultural interpreters.
MIT research puts large language models at 44 percent accuracy in predicting innovation adoption when they only have public data, while human-AI hybrids beat pure AI by 31 percent on contextual reasoning. The division of labor follows from that. The machine is good at pattern recognition, speed at scale, prediction and sentiment. The human is good at political nuance, cultural code-breaking, mapping the unspoken rituals and reading the hierarchy.
So the corporate anthropologist's toolkit looks something like this. Meeting semiotics: decoding what is not said at the leadership offsite, where a long silence after a "disruption" proposal is the cultural antibodies at work. Promotion forensics: tracking who gets rewarded for safe bets and who for moonshots, which is the truest indicator of an organization's appetite for innovation. Email latency: measuring how long each department takes to answer an innovation memo, where a 72-hour lag from Legal is a fair proxy for a veto, being Legal's way of saying no without the paperwork.
One pharma company combined AI analysis of its clinical-trial data with ethnographic study of its R&D teams, and found that researchers were avoiding high-risk projects because of an unspoken stigma left by a pipeline disaster in 2018. No model could have seen that in the data.
Three what-ifs for anyone building with AI.
What if the AI analyzed promotion patterns instead of press releases? Would who gets rewarded predict your real appetite for risk better than what you say about it?
What if your innovation dashboard showed cultural antibodies: a heat map of which departments have historically killed disruptive ideas, and why? HR's risk aversion, say, or Legal's fondness for patents.
What if the AI simulated your culture's immune response? Before you pitch the blockchain initiative, test it against personas modeled on your longest-serving middle manager, your most risk-averse board member and the innovation director who burned out in the failed pivot of 2020.
Building a thicker innovation engine means adding the layers the data does not carry on its own. A behavioral layer: does the executive team default to gut feel, Kahneman's System 1, while claiming to be data-driven? Most do, and the data on that is quite good. A cultural layer: where did the "digital transformation" budget actually go, when 73 percent of it went to maintaining legacy systems? A political layer: which functions answer innovation proposals slowest?
The most innovative companies of the next decade will be the ones that best translate their own irrationalities into signals a machine can read, and that has less to do with having the best AI than with being honest. The question is less whether the AI will understand your strategy than whether you are willing to admit what your strategy really is.
What cultural contradiction is your innovation strategy ignoring?