Turning Policy into Something Agents Can Use
Building the organizational context layer, and why most pilots fail without it
An engineer in your Munich office is expecting a child. She opens the agent you were so proud to launch and asks how much parental leave she is entitled to. The agent answers in a second, in full sentences, with a number. The number is wrong. It came from a policy PDF written before the German rules changed and before your company added an enhanced top-up at two years of tenure. She plans her year around a confident, wrong answer. When payroll corrects it four months later, she does not feel corrected. She feels lied to.
Here is the claim this chapter is built on: an agent is only ever as good as the company knowledge underneath it. The model, the interface, the demo that impressed the board, all sit on top of your policies and decision rules. If those are stale or written for a human to interpret, the agent will do what software does. It answers anyway, fluently, and is wrong. A confident wrong answer about someone's entitlement is worse than no answer, because it gets believed and acted on.
Get someone's parental leave wrong once, with confidence, and you have not lost a support ticket. You have lost that person's trust for a year.
What sits under a good answer
The organizational context layer is the curated, machine-readable body of company knowledge an agent reads before it responds. It is not one document. It is six kinds of thing, each with an owner.
Policy
What the company permits, requires, and prohibits. Parental leave duration, eligibility, notice periods, the enhanced tiers you actually offer.
Competency frameworks
The skills, levels, and behaviors that define each role, so the agent can reason about growth, not just recite a job title.
Job architecture
How roles, families, grades, and reporting lines fit together. Without it, an agent cannot tell a manager from an individual contributor.
Decision rules
The if-then logic a human applies today in their head. Who approves what, under which conditions, with which exceptions.
Escalation paths
When the agent must stop and route to a person, and to which person. The boundary that keeps it below the human line.
Entitlement logic
The rules that resolve what a specific person gets, given their country, tenure, and contract type. Our worked example lives here.
A document is not a rule
The most common objection from an impatient sponsor is that this work is already done. It is all on the intranet: a parental leave policy, a benefits handbook, a wiki. Point the agent at it and let it read.
This misunderstands what those documents are. A policy document is written for a human being who will apply judgment. It says things like "eligible employees may be entitled to enhanced leave subject to local statutory minimums." A person reads that, knows they are in Germany, remembers the two-year top-up, and lands on the right answer. The document assumes a reader who fills the gaps.
A rule is different. It is written to be applied, not interpreted. It states the country, the tenure threshold, the contract types it covers, the exact number of weeks, and what happens at every boundary. Most HR policy is the first thing wearing the costume of the second: prose that looks authoritative but leaves the real decision to a human who is no longer in the loop once an agent takes over.
Look for the words "may," "subject to," "at the discretion of," and "where applicable." Each one is a decision the author quietly handed to a human. An agent cannot exercise discretion it was never given.
From a stale PDF to an answer you can trust
The same parental leave policy can exist in four states. Each rung is more work than the last, and only the top rung produces answers you would stake someone's year on.
Documents
What it looks like: a PDF handbook and a few wiki pages, written in prose, last touched who knows when. What an agent can do: retrieve passages and summarize them. What it still gets wrong: nearly everything specific. It quotes the German section to a French employee and cannot tell a contractor from a full-time hire.
Structured content
What it looks like: the same policy broken into tagged sections by country and topic. What an agent can do: route to the right country and quote the right paragraph. What it still gets wrong: the math. It finds the German section but cannot compute weeks for someone at twenty-two months on a fixed-term contract, because the tiers are described, not defined.
Machine-readable rules
What it looks like: entitlement as explicit logic. Country, tenure threshold, contract type, weeks, and every boundary condition, in a form the agent evaluates rather than reads. What an agent can do: compute the correct answer for a named person. What it still gets wrong: it goes stale silently. When the German statutory minimum changes, nothing tells the rule, and it stays confidently wrong.
Rules with named owners and review cycles
What it looks like: the same logic, plus a named owner in the benefits team and a scheduled review tied to legislative change. What an agent can do: give an answer that stays correct, with a clear line of accountability when it does not. What it still gets wrong: less, and never silently. Someone owns the gap between the rule and reality, and closing it is their job.
Most companies live on rung one and believe they are higher. The distance between rungs three and four is the part teams skip, and it matters most. Rung three gives you a correct answer today. Rung four keeps it correct after you have stopped looking. Knowledge that nobody owns decays quietly, until an agent surfaces the decay to the one person who trusted it.
Follow one answer down and back up
Watch a single question travel down through the context layer and back up. At each stage the agent consults one knowledge object. Miss or stale any one of them and the answer breaks in a specific, traceable way.
The branching happens inside that fourth step. The entitlement rule is a small decision tree, and every fork is a knowledge object that has to be right.
Take the Germany, at-least-two-years path. If the enhanced top-up rule is missing, the agent silently drops to the statutory minimum and under-quotes by weeks. If the rule exists but is stale, it quotes last year's number with full confidence. Either way the person acts on it. The fix is not a smarter model. It is one owned, dated rule at that fork.
Do the knowledge work first
This is the argument that survives an impatient sponsor. Trust in the agent is built or lost on the accuracy of its first hard answer, and accuracy is a property of the knowledge, not the model.
You do the knowledge work before the agent work because the alternative is a launch that teaches your workforce the agent cannot be trusted, a lesson that is expensive and slow to unlearn. Curating one entitlement rule to rung four is unglamorous. It is also the difference between a tool people rely on and a demo people quietly stop opening.
- Pick one high-stakes entitlement. Parental leave is a good first choice: it varies by country, tenure, and contract, and getting it wrong has real human cost.
- Find every version that exists. The PDF, the wiki page, the answer your benefits lead gives from memory. Note where they disagree. They will.
- Write it as explicit logic. Country, tenure threshold, contract type, weeks, and every boundary. Where the source says "may" or "subject to," resolve it into an actual rule.
- Name one owner and set a review date. A real person in benefits, and a trigger tied to legislative change, not a calendar reminder nobody honors.
- Test it against ten real cases before an agent ever reads it, including the awkward ones: the fixed-term contractor, the person at twenty-two months, the recent transfer.
- You point the agent at the intranet and call it done. Retrieval over prose gives you fluent, plausible, unowned answers. It is rung one wearing rung four's confidence, and it will under-quote someone's entitlement without ever flagging that it did.
- You build the rules and never assign owners. Rung three feels like victory. Then a statute changes, no one updates the logic, and six months later the agent is confidently quoting a number that stopped being true. Unowned knowledge does not stay still. It rots.