AI Agents as Digital Employees: Why They Need a Job Manual, Not Just APIs

At a morning operations meeting, an AI agent automatically pulled ERP data, calculated a scrap-rate chart, and posted it to the leadership group. The chart showed Plant 3's casting scrap rate up 1.2 percentage points year-over-year. The chairman asked for an explanation. The team discovered the agent had compared two incompatible denominators: Plant 3 used "total weight of castings out of furnace" while Plant 1 used "number of charged pieces." Worse, Plant 3 had a furnace change on Wednesday; that batch should have been excluded per standard practice, but the agent didn't know.

The team's response was to give the agent a "job manual" — a codified set of business rules and context. The author argues that most current "agents" are merely API-calling interns: they can fetch data and call tools but lack identity, persistent context, and authority boundaries. True agents need three independences: independent identity (a role in the process), independent context (carrying the role's background data and history), and independent authority (making judgments, submissions, and rejections within defined limits).

However, independence without guardrails is dangerous. The morning-meeting agent had independence — it pulled data, computed, and published — but it crossed a business-common-sense boundary: it didn't know the denominator definitions or the furnace-change exclusion rule. The author compares this to hiring a junior accountant and letting them approve a 5 million payment or set depreciation policies without a job description, approval limits, segregation of duties, or audit logs.

The real business guardrails are not prompt engineering, ReAct frameworks, or permission whitelists (those are technical reins). They are two things: business semantics and business ontology . In plain language, onboarding a warehouse supervisor means giving them:

A job dictionary defining terms: what "casting," "work-in-process," "tooling" vs. "mold" mean, and how "complete set" differs from other plants.

A responsibility manual clarifying scope: receiving, shipping, and cycle counts are theirs; scrap write-offs over 5,000 yuan require the warehouse manager.

A set of tacit business rules not in SOPs: "materials arriving before 8 a.m. get priority inspection," "no inventory moves on month-end closing day," "customer batch numbers must match quality certificates."

An object-relationship ledger mapping which part belongs to which BOM level, which work order links to which sales order, which prepayment offsets which supplier payable, which equipment spares belong to which division.

Together these form the "business common-sense operating system" for the digital employee.

The article contrasts two agent responses to the same query: "Pull Plant 3's casting scrap rate for last week."

Without business semantics/ontology: "Plant 3 casting scrap rate last week: 3.2%." It called an API, returned a number, but doesn't know what the number means — denominator, exclusions, comparability. With business semantics/ontology: "Plant 3, Oct 13–19, casting scrap rate per our plant definition (denominator = total cast weight, excluding Oct 15 furnace-change anomaly batch) is 2.87%. Week-over-week up 0.4 pp, driven by Line A third shift, sand-hole defects rising from 18% to 31%. Suggest quality adjust sand ratio."

The second response acts like a competent quality engineer: it knows the definition, the exception, the trend, and the next investigative step. Only then is the agent a trustworthy, independent business role in the process.

Many enterprises buy models, compute, and agent frameworks but find agents unusable because they equipped "hands and feet" (APIs, tool calls, RPA) but no "brain and job" (business semantics and ontology). The brain is business semantics; the job is business ontology.

The final insight: business ontology isn't something you buy today. It's the 20 years of transactions you entered into ERP — customer master, vendor master, material codes, BOM levels, routings, cost centers, profit centers, plant codes, storage locations, batch numbers, serial numbers. You thought you were entering data; you were actually brick-by-brick building the company's accumulated lessons, disputes, and rules into the system. Companies that cleaned master data, aligned BOMs, and locked cost-center definitions 20 years ago can onboard agents instantly. Those that cut corners on ERP now face a tangled mess no model can untangle.

Therefore, stop asking "When do we deploy agents?" Ask instead: "Is our business ledger clean?" If it's clean, the agent is a digital employee with a job manual. If it's messy, you're not hiring an employee — you're inviting the next embarrassing incident at the morning meeting.