Teach Your AI How You Make Decisions

Some organizations have used AI agents to genuinely transform how work gets done. Others are stuck running small low-stakes experiments, unable to get their agents to perform consistently at scale.

The divide between these two groups isn’t technological, as you might expect. Most organizations today have access to the same models, the same tools, and roughly the same infrastructure. The divide stems from different approaches to something that most leaders have never had to confront: making judgment explicit. That’s the new bottleneck in AI adoption, and it’s catching most organizations off guard.

For decades, organizations had no need to articulate how their best people made decisions. Expertise was absorbed through mentorship, observation, and experience. New employees watched, listened, and gradually internalized how the organization thought. That model worked when humans were doing the executing.

AI agents have changed the equation. Unlike traditional software, they can operate in ambiguous environments and make decisions in real time. But unlike people, they can’t absorb norms through observation or infer context from organizational culture. They operate based on what is made explicit. Nothing more.

This creates a specific failure mode that is now showing up across industries: Companies deploy customer-facing AI agents without first codifying how their best service reps make their decisions—handle a pricing exception, for example, or a frustrated long-term customer, or a request that sits just outside policy. The agent eventually goes off track, unaligned with the firm’s goals, because no one ever wrote down how that particular decision actually gets made or what inferred context is necessary to make the right decision.

What does it mean to codify judgment? It means translating the tacit decision-making principles of your organization into structured guidance that agents can execute. These principles include such factors as risk tolerance, brand voice, escalation thresholds, quality standards, and the subtle logic of exception handling. Historically, these things lived in the minds of experienced people. For AI to work, they need to live somewhere else too.

What Separates Leaders from Laggards

Organizations that are pulling ahead are building what we call judgment infrastructure, which is what allows expertise to scale.

Three structural shifts are necessary for leaders to create this infrastructure:

1) The business units, HR, and IT govern together.

Defining acceptable risk boundaries, setting performance expectations for agents, managing how they’re onboarded and offboarded—these are organizational questions as much as technical ones. And critically, they can’t be outsourced.

What’s necessary is for business leaders, HR, and IT forge a partnership to govern digital labor. The goal should be to treat agents less like software licenses and more like operational contributors whose behavior must be shaped and continually refined. This mirrors what we described last year, in “Agentic AI is Already Changing the Workplace”: The organizations that succeed in this new company will be those that actively manage the full spectrum of labor—human and digital—as part of a coherent workforce strategy.

ITA Group, a global events, incentive, and recognition company, learned this lesson through an early attempt to build an AI agent for air-travel booking in its events business. The hard part was not building the agent. It was defining what the agent needed to know to be trusted: when to optimize for cost, when traveler experience mattered more, which exceptions were acceptable, and when a human needed to intervene.

That experience exposed a broader issue: Too much judgment was being translated from business experts to technologists. ITA’s response was to change the operating model. The company began giving developers, managers, and knowledge workers the tools and ability to shape the agents acting on their behalf. The COO, Maura McCarthy, working closely with the CIO, Jason Katcher, with support from the CEO and CFO, helped create the leadership alignment to make that shift stick.

“The most valuable lesson we’ve learned,” McCarthy says, “is the importance of pairing our expertise with AI.” The goal was to get agent behavior right in the hands of expert users before scaling it broadly, and to treat that behavior as something that must be continually refined as the work, the business, and the judgment around it evolve.

2) Managers become judgment architects.

This is the most significant shift, and the one most organizations underestimate.

Consider Debbie Riazzi, the director of compliance and labor relations at AWP Safety, the largest field-safety company in North America, with nearly 9,100 employees across 33 states. Riazzi is a one-person department who has built a portfolio of agents, each codifying a different slice of her expertise. One handles medical accommodation requests by pulling the relevant job description, surfacing how comparable requests were resolved, and running through a standardized intake she has refined over years. Another handles the opening moves of every information request the company receives: parsing what’s being asked, routing to the right owner, and drafting the response. The agents save her hundreds of hours a year, but the more important shift is what that time unlocks. “I can go back and show I’ve been doing this consistently,” she says. “That automatically reduces our liability as a company.”

Nathan Mapp, who serves as the controller at a global venture-capital and applied-technology firm, has taken this further. Over more than a dozen years in finance, Mapp developed a deep body of expertise and codified it into a series of markdown files that his agents, built on Claude and Claude Code, can reference in real time. A team of two now covers ground that would previously have required 10. In every task that those agents handle, Mapp’s judgment is applied consistently, as if a top-tier accountant were paying attention to every detail, including the ones that would otherwise fall to a more junior team member.

What Riazzi and Mapp have done is codify judgment at work. Managers are now focused on operationalizing expertise, in both human and digital form. That’s a fundamentally different skill set, and one that most organizations haven’t yet developed or rewarded.

3) The “thought-doer” becomes the most valuable employee.

The traditional divide between strategic thinkers and operational doers is collapsing. High-performing employees increasingly are both—they reason strategically and operationalize their thinking through agents. They design workflows, encode judgment, build and iterate on systems, and continuously move up the value chain.

Ramp, a leading financial platform used by 30,000 companies, has bet the company on this profile. Every employee gets access to ChatGPT Enterprise, Notion, and Perplexity, and is trained during onboarding to build their own AI tools rather than act as a “button pusher” on systems someone else created. Individual employees are equipped to codify their own expertise and deploy it through agents. This is the profile we’ve started calling “the thought-doer”: someone who doesn’t just use AI as a tool but shapes how AI executes on their behalf. Organizations that cultivate this profile at scale will learn faster and adapt faster than those still optimizing for one skill or the other.

Where to Start

Most organizations approach this challenge the wrong way. They ask experienced people to write down what they know. That rarely works. Experts are notoriously poor at articulating tacit knowledge in the abstract, because they know far more than they can say when asked directly to document it.

A more effective method: Don’t ask them to document their judgment. Create conditions where it surfaces naturally.

Convene a small panel of experienced practitioners in the same role. Bring in a skilled moderator and walk the group through a series of realistic scenarios and actual edge cases faced by the organization. Where the panel agrees quickly, you have a clear policy. Where they disagree, you have judgment worth capturing. The transcript of that conversation becomes your first draft of codified judgment.

A claims team at an insurance firm might surface more nuance about risk tolerance, customer empathy, and escalation logic in a single two-hour session than years of documented procedures ever captured, because debate externalizes reasoning in a way that documentation never does. That transcript is needed as a critical context layer for all future agentic deployments.

A New Differentiator

When judgment is successfully codified, something strategically significant happens: Expertise becomes portable. Best practices are no longer locked inside your most senior people. Institutional knowledge can be deployed across functions, geographies, and products at scale. The organizations that figure out how to encode tacit knowledge will see structural advantage: faster decisions, more consistent quality, greater capacity for innovation, and a self-reinforcing ability to learn and improve.

ITA Group’s trajectory shows what this compounding looks like in practice. The first six to seven months were slow because the company was learning how to translate judgment and domain expertise from expert employees’ heads into the context files agents use. But once that operating model began to take hold, the pace changed, especially in software development. With some initial coaching, developers began using agents not just to generate code, but to move faster from idea to working prototype. Much of the work became self-directed, timelines shrank from months to weeks, and the habit of using agents to rethink the work itself began spreading into other functions.

The lesson is that codified judgment compounds. The first use cases are slow because the organization is learning how to make expertise explicit. The next ones move faster because the organization has built the trust, governance, and operating rhythm to repeat the pattern.

This is the natural evolution of the hybrid-workforce strategy we outlined in our earlier article about how agentic AI is changing the workplace. Mapping tasks and integrating human-AI teams is the necessary first step. But it’s not enough. The organizations that build human-AI teams quickly discover that the quality of those teams depends entirely on the quality of the guidance they give their agents. Deployment is table stakes. Judgment infrastructure is the competitive moat and the strategic differentiator for a frontier firm.

. . .

The first phase of AI adoption was about who had access to the best models. That phase is largely over, because access has been commoditized. The next phase will be defined by who has done the harder work of encoding how they actually think and work.

Most organizations haven’t started doing that work yet. The ones that do will shape the future of work in their industries. The ones that don’t will find themselves in the same position as the firms that hesitated to build a workforce strategy in the first place: structurally disadvantaged in ways that compound over time.