The number was going down. That was the problem.
It sat in a spreadsheet with the company's other key performance indicators, neatly labeled and presented to me with some pride. Cost per hire was falling.
Almost everything around it was getting worse.
Quality was inconsistent. Sales performance was uneven. Customer service problems were multiplying. Experienced employees were leaving, and their replacements were struggling.
The spreadsheet said at least one part of the company was becoming more efficient. The business itself suggested otherwise.
A perfectly respectable number
Cost per hire is not a foolish thing to measure. For a large recruiting organization, it can reveal useful information about the efficiency of hiring. But this was a smaller company, and hiring efficiency was not its real problem.
Its real problem was that it had to keep hiring.
The company knew what it cost to put a new person in a seat. It did not meaningfully track why the previous person had left, what knowledge had left with them, how long a replacement took to become capable, or what mistakes occurred while that person tried to reconstruct the role.
The metric was accurate. The story it appeared to tell was false.
Across different industries and organizations, I have seen versions of this same pattern. A company chooses a measure because it represents something important. The measure becomes a target. Then people begin optimizing the number rather than the outcome it was meant to represent.
Goodhart's law warns us about this failure. Once a measure becomes a target, it can lose its value as a measure.
I think of it more simply as “make the number go up” or “make the number go down” syndrome.
Shorten customer calls. Increase sales activity. Reduce labor costs. Close more tickets. Raise production volume.
Each instruction sounds sensible in isolation. Each can also reward behavior that damages the larger system. A service team can reduce call times by rushing customers away. A sales team can make more calls to poorly qualified prospects. A company can lower its hiring cost while repeatedly replacing people it never prepared or persuaded to stay.
The number improves. The business gets worse.
The problems were connected
The company's difficulties in quality, sales, and customer service initially appeared separate. They shared an underlying mechanism.
Experienced employees were leaving. Their knowledge was leaving with them. New employees entered roles with little structured onboarding, few durable reference materials, and inconsistent support.
They were not failing because they did not care. Many were well-intentioned people doing the best they could with the information and experience available to them.
The system expected inexperienced employees to reproduce experienced judgment without giving them a reliable way to acquire it.
We began looking beyond the hiring transaction and toward the full system around it: recruitment, onboarding, training, performance, development, retention, and the preservation of operating knowledge.
We documented roles and created durable reference materials so critical knowledge did not live entirely inside one person's head. We built onboarding plans that gave employees tools, context, and continuing support. People needed to understand not only what to do, but why the work mattered and how their part fit into the larger organization.
Compensation remained constrained by the realities of a smaller business, but the model became more transparent and recognized excellence and innovation alongside tenure.
We also identified something the company had never formally considered an employee benefit: autonomy.
Employees received a small amount of discretionary time to explore recurring problems and pursue ideas. Coaching and leadership review gave those ideas somewhere to go. Some were useful. Many were not. The immediate value was not a stream of guaranteed innovations. It was agency.
Over time, useful ideas did emerge. More importantly, employees began to see the organization as something they could help shape rather than simply a place where they completed assigned tasks.
Procedures should preserve judgment, not replace it
This is not an argument for documenting every movement or prescribing one management formula to every business. These interventions belonged to this company and its circumstances.
The more general lesson was that knowledge, context, and agency were not moving through the organization effectively. The procedures mattered because they made experience transferable. The onboarding mattered because it made context visible. The autonomy mattered because no procedure can anticipate every condition in which judgment will be needed.
A durable procedure is therefore not simply a list of steps. It should help someone understand the intended outcome, the evidence behind the work, the common exceptions, the boundaries of their authority, and when to ask for help.
That was useful for a new employee. It is also useful for artificial intelligence.
Put AI in the new employee's chair
Years later, I see the same organizational problem returning through a new technology.
Companies want AI to understand their businesses. Yet much of what their businesses know has never been made understandable.
Critical context lives in experienced employees' heads. Official procedures describe an idealized version of work while exceptions survive in email, conversation, and habit. Measures reward visible activity without preserving the reasoning that connects activity to an outcome.
An AI entering that environment faces a version of the new hire's problem. It can read the procedure without knowing when the procedure does not apply. It can calculate the KPI without understanding what the KPI has displaced. It can reproduce an answer without possessing the context that made the answer valid.
The work that makes an organization intelligible to its people is also the work that makes it intelligible to AI.
But there is an important warning inside that promise. AI does not automatically correct the system it enters. It can operationalize the organization's misunderstandings as readily as its knowledge.
If the metric is wrong, AI can help the wrong number move faster. If the procedure omits the real exception, AI can apply the incomplete procedure more consistently. If nobody has defined where judgment belongs, confident output can conceal the absence of authority.
AI readiness is therefore not only a technology problem. It is an organizational clarity problem.
What are we actually trying to accomplish?
Metrics remain necessary. Without them, leaders can mistake anecdotes for patterns, activity for progress, and confidence for competence.
But a metric should be treated as evidence, not reality itself.
Before turning a number into a target, leaders should ask what behavior it will encourage, what information it leaves out, and whether the number could improve while the outcome deteriorates.
Before asking AI to scale a process, they should ask whether the process captures the knowledge, context, exceptions, authority, and feedback required to produce a sound result.
The most dangerous metrics are not necessarily inaccurate. They are often precise measurements of something that matters less than we have decided it does.
Sometimes the numbers show us what we need to know. Sometimes they show us only what we wanted to see.