Your Culture is Being Written Down. Who Is Holding the Pen?

Nobody will notice. Her veto was never written down anywhere. Neither is most of what you call your culture — the thousands of business decisions, assumptions and actions your people make every day that determine what work gets done, what behaviors get rewarded, what gets tolerated, whose objection ends a meeting and whose does not. Culture lives in people. It is learned by watching. Nobody has to write it down.

Enter AI. You cannot deploy AI on unwritten things. Someone has to specify what the tool should do. Someone has to define what good output looks like, and what the system may decide without asking a person. Prompts, workflows, thresholds, evaluation criteria are all written statements cementing a decision for your business. What does this all mean for leadership? If more of the culture is embedded in systems, changing the CEO may no longer change the organization as completely as boards assume.

For as long as there have been companies, the way to change culture has been to change people. New leaders arrive with different instincts. Turnover flushes out inherited habits. A successor takes over a culture and renegotiates it. That assumption sits underneath a great deal of corporate governance. It is why boards have believed, correctly, that changing the chief executive changes the company.

But rules that live in systems do not turn over when the people do. A new CEO inherits the org chart, the balance sheet and the strategy. She now also inherits a layer of decision rules — likely written by an LLM learning on new and legacy data without human involvement — that quietly provide the ground rules for the organization’s work. This power to write culture rules is why AI should be a standing part of every board meeting, executive session and committee agenda, not just an IT roadmap or CEO town hall.

For CEOs, this is an operating problem to overcome, and we believe most will. The solution is to be intentional about human influence on this process: understand how decision rules came to be, which should survive, where human decision-making matters most and how the culture is likely to respond.

Who is holding the pen?

The first thing AI does to culture is transcribe it; the second thing it does if you’re not careful is hand the pen to someone else. This demands your attention.

Your values statement was drafted on the executive floor, with care, over several months. With AI, the rules that govern actual decisions are drafted multiple levels below it by whoever is configuring a tool. Often not even by them. A large enterprise buys 30–40 products with AI folded into them, inheriting 30–40 companies’ opinions about when it is acceptable to act without asking. When vendors select a default action, they are writing an operating rule for every customer that doesn’t evaluate and change it. Rules also can change without your knowledge. One enterprise platform, for example, redefined a billable “solution” from resolving a customer issue to examining one. Same tool, same usage, much larger invoice. No customer agreed to the change. The rule announced in a release note.

Because it draws from history, AI can codify an organization’s “unwritten rules” — for good and ill. In one well-known example, Amazon trained a résumé-scoring system on 10 years of its own applications. The system learned to prefer men for technical roles. It downgraded résumés containing the word “women’s” and rewarded verbs more common on men’s CVs. No recruiter had written that rule. The company’s hiring record essentially had written the rule, and the model read it back as policy.

The most common failure is quieter than a bad rule hardening into place; it’s failing to agree on definitions. One company bought a customer-success platform to flag at-risk accounts and deployed it fully. It could not do the job. “At risk” meant four different things across the business, and nobody had taught the system any of them. They had bought the dashboard and skipped the definitions. A default filled the vacuum. The rule that ends up governing you is rarely the one somebody wrote badly. It is the one nobody wrote at all.

What survives transcription, and what doesn’t, in your culture?

Transcription is not neutral. It captures what can be specified and loses what cannot. The parts of your culture that live in documents get scaled. The parts that live in people — judgment, taste, dissent, the instinct that something is wrong before you can say why — don’t survive the transfer. Companies do not become louder versions of themselves under AI. They become lopsided ones.

So it is worth predicting what gets written down with AI. At Spencer Stuart, we use the Culture Alignment Framework to map eight organizational styles: Results, Caring, Order, Learning, Purpose, Safety, Enjoyment and Authority. The framework sorts styles along two dimensions: how people interact and how they respond to change. None is better than another. Each solves certain problems brilliantly and creates others predictably.

Across more than 230 companies studied, Results ranked as a dominant style in 89% and Caring in 63%. The pattern held across industries, regions and company sizes. Statistically, if you run a large company, you run some version of a Results-and-Caring culture. That pairing carries a familiar tension. Results pushes people to optimize and compete. Caring pushes them to preserve harmony.

AI breaks the truce. It gives the Results side an accelerant. It gives the Caring side a way around friction, because a hard judgment can now be deferred to what the model says. The result is a company that moves faster and confronts less.

A company’s dominant style traditionally describes what a company is like. Read it now as a forecast of what AI models will capture and codify in your culture.

EXHIBIT 1: CULTURE STYLES AND AI

The cost of leaving transcription to AI

You could reasonably object that none of this is permanent. Software can be changed, and the cautionary cases you’ve seen in the public realm were all reversed. Amazon scrapped the recruiting engine. Klarna was rehiring some of the 700 service agents replaced by an AI assistant within about a year, with its CEO conceding that cost had become too dominant a factor. After coming under fire, Australia’s Robodebt reversed its decision to automate welfare debt recovery, raising roughly 500,000 debts and collecting some $751 million from about 380,000 people. The scheme ended in a federal court finding of a massive failure of public administration, and a royal commission ruled it unlawful.

Look at what each of those reversals required. An investigative news story. A collapse in service quality public enough that a founder recanted at cost to his own standing. A royal commission. Not one of those rules was reversed because somebody reviewed it. They were reversed because they became scandals.

The unexamined mechanism is the one that will affect your business. At one company, the definitions of the pipeline stage in the customer relationship management system were written years ago by an employee who has long since left. They still govern how the entire go-to-market organization talks about the business. Not because they are expensive or difficult to change. The definitions survive because nobody’s number depends on changing them. That points to the hopeful conclusion: if the chief revenue officer believed that changing those definitions would help him hit his number, they would change tomorrow.

That is the real lock, and it is not technical cost. It is the absence of an owner whose objective requires the rule to be different. The remedy is ownership.

The remedy: three things to do this quarter

Culture change through meetings, metrics and promotion cycles still works. It is also still slow, taking months and sometimes years before behavior moves. Defining the decision rules that AI is codifying cannot wait that long. CEOs and boards must understand the cultural risks of AI adoption for their organization. This starts with three narrow and unglamorous steps that CEOs should ensure begin now.

• Put a name on the key areas where AI is encoding decisions for each system in your organization. This means a person, not a committee, and they should be senior enough that the choice registers as policy, but close enough to the work to know what the current setting actually is. For each system, look at procurement, configuration and defaults, prompts and system instructions, autonomy thresholds and evaluation criteria.
• Write a “do not automate” list and test it. Every organization can produce a long list of friction it considers sacred, and much of it is merely friction. The test: each entry names an owner and the specific failure it prevents. Anything that fails the test is not a control. It is a preference.
• Change what gets rewarded, because that is what gives a rule an owner. Ask who was promoted last year for catching a mistake the system made. The silence is the finding. The rules will stay exactly as written until somebody’s advancement depends on reading them.

One structural point matters more than it sounds. Autonomy thresholds for anything customer-facing, regulatory or decided about a person should be reported to the board, not approved by it. Approval turns directors into a bottleneck they are not equipped to be. Reporting keeps the decision where it belongs and still puts it in front of the people who can ask why.

The best chance you will get

This is the opportunity of a lifetime to lay the foundation for best aspects of your culture to live in perpetuity while evolving for the future. Against the risk that transcription by AI-run systems will erode humans’ influence on culture, leaders can take the reins by evaluating the culture they have, considering the culture they want and configuring systems to support that culture by writing the rules themselves. You cannot configure a system on aspirations. You have to write down what actually happens. Somebody finally has to say it out loud in a room.

Do this well and the analyst does not have to slow-walk anything. She says it in the meeting, gets promoted for it, and the rule that ends up in the file is the better one.