When Josh Tyson and Robb Wilson were writing Age of Invisible Machines, they noticed a condition that had nothing to do with the model. A company that wants to run on agents has to make a long-term investment in knowledge and in the design of its service. For-profit firms answer to stakeholders who want the gain now, so the investment gets delayed or dressed up as a pilot. Credit unions might be the exception. The shareholders are the members. The value has somewhere to land.
Mona Leung, former CFO of Alliant Credit Union, came to the sector from banking and does not romanticize it. What she trusts is the alignment. Customers and owners are the same people, so the profit conversation and the service conversation face the same direction.
Robb reaches for the flywheel Jeff Bezos drew: a better product brings customers, volume lowers cost, lower cost improves the product. Mona’s amendment is small and total. A credit union runs the same wheel, except the profit circles back inside the same population. Robb doesn’t think the commercial version is a wheel at all. It is a facade of one: a gerbil spinning while someone at the rim counts the money as it flies off.
A for-profit ships early and imperfectly because investors are owed a return. A credit union can ask whether members want the thing, and sometimes the honest answer is something plainer and easier to use. The prudence is tied to the balance sheet. A credit union can’t raise outside capital; it spends what it earns. Merging is one of the few ways to get capital to build.
Josh’s idea for that capital: credit unions don’t compete with one another and already share technology. If one gets knowledge management right, the common portion could move. Another credit union could offer a product it had never staffed, because the verified knowledge of how it works already sits in the pool. Mona thinks the structure is half-built in the regional corporates and CUSOs that serve credit unions. AI is the catalyst that makes the sharing urgent.
Robb also likes credit unions for their pace: AI doesn’t throttle itself, and they are better at slowing a launch than winning a sprint. Mona sharpens it. What AI lacks is social context, and slowing it down doesn’t teach that. Robb takes the point. The slowness he means is organizational: don’t release an untested build. Innovation should go faster, because better simulations are how you earn a careful release.
Robb calls knowledge long-term memory, and context the reduction of it to what one task requires. The fashionable failure is the reverse: every agent builds its own context with nothing canonical underneath. Wrong knowledge is worse than no knowledge. A shared canonical layer of policy, products, and regulation would let each institution pull what it needs and add its own context. That layer has to hold judgment, not just facts. A birthday is information; knowing which product fits that moment in someone’s life is wisdom, and it has to live in one place.
Mona adds the return path: if the big brain never hears whether the answer worked, it can’t get better. Robb agrees: human memory is rewritten every time it is retrieved, and a system that never records its own use is not yet a memory.
It sounds enormous, but NASA managed critical knowledge with people, before AI. Roger Forsgren’s method: find the role that depends most on knowledge, treat what it must know like a training curriculum, then move to the next role. The unfinished part was feedback. Two teams, one in metric and one in imperial, were enough to cause a disaster. Mona draws the line hard: better context doesn’t make judgment a function of that context. People say AI will replace humans. Humans what?
Robb’s answer is responsibility. The person who signs has a paycheck and can be sued. AI has no job to lose. Mona finishes the thought: no guilty conscience, either. The human’s work is reading the decision and owning it, and the fear is the skim.
Owning it depends on speaking up; at NASA, Josh notes, silence could be fatal. Mona makes the incentive concrete. People who depend on a paycheck choose the path safest for themselves and stay quiet. The design problem is an incentive that lets feedback in without punishing it, while still holding the speaker responsible. If an agent floods Robb with decisions that put his job on the line, why wouldn’t he refuse them all? Mona knows the script. Need more information. Send it back. It is how a corporation declines to decide.
Josh names a second refusal. The knowledge that matters most is in people’s heads, and nobody will hand it to a model they believe will replace them. A credit union’s answer, that members come because there are people there, has to be believed before any cuts. Mona widens the question. An employee spends into a community. Remove the job, and where does that capital go? If the people you serve live where your employees live, “we saved the headcount” is not yet a result.
A client story of Robb’s turns the same corner: AI built to predict churn found a wide band of customers the company had been content to lose. Automation created work, and employees felt betrayed. Mona’s distinction is the one to keep: not more work, different work, and keeping a member costs less than winning one. Robb’s rule: start from the knowledge and let the data show where the opportunity is, rather than automating whatever is easy.
All of this protects something easy to automate away. After a career in the industry, Mona is still intimidated by banks. In a credit union she can talk like a person. That ease is the asset. Robb’s bank, he says, would prefer he not come in at all. They call it rationing by friction: the deposit slip, the scolding about the ATM. A credit union that copies the tactic in software will have spent its structural advantage on a queue.
Josh proposes that one credit union go deep on knowledge management and another on service design, since so much of the member journey is shared. Robb treats the blueprint as another shape of knowledge, an objective an agent can be given: facilitate this journey well. Mona has lived under the other objective: get them off the phone in under a minute and a half. She has the joke ready. New York–style banking. What do you want?
The joke is the failure mode of the whole hour. An agent will pursue the objective it is given. If the owners want the call to die quickly, the model becomes a teller offended that you came in. If the owners are the members, the objective can be the journey, the simpler product, the person who still answers, the knowledge that got better because someone reported that it failed. The long investment is hard for a company whose value flies off the wheel. It is the ordinary work of a company whose value stays.
Companion essay for the The Shareholder Is the Member ideation—Mona Leung on Invisible Machines S7E20. Read the full transcript. Also in this cluster: Trust Is the Currency, Knowledge Is the Engine (Federico Cohen Freue, Mastercard). Related: AI Activates Service Design (S7E19).