---
title: Your AI Agents Have Job Descriptions. Your People Don't.
description: "Teams now write precise specs for AI agents and leave human work at \"use judgment.\" Here is how to define human work with the same fields, and why it matters."
image: https://media.babylovegrowth.ai/blog-images/organization-48833/1792742483029_Hand-drawn-role-spec-sketches-framing-the-title.jpeg
---

[Skip to content](https://cognistry.co/edge/give-ai-job-decsriptions#main-content)

[![AI Powered Course Designer for Learning Architects](https://cognistry.co/hs-fs/hubfs/cognistry_primary-1.png?width=250&height=100&name=cognistry_primary-1.png "AI Powered Course Designer for Learning Architects")](https://cognistry.co)

- [Home](https://cognistry.co)
- [Platform](https://cognistry.co/platform)

    - [What is Capability Engineering](https://cognistry.co/capability-engineering-system)
    - [Signal](https://cognistry.co/signal)
    - [Forge](https://cognistry.co/forge)
    - [Sim](https://cognistry.co/sim)
    - [Edge](https://cognistry.co/adoption)
- [Solutions](https://cognistry.co/overview)

    - [Customer Success](https://cognistry.co/customer-success-support)
    - [Learning & Development](https://cognistry.co/learning-development-training)
    - [Manager Enablement](https://cognistry.co/manager-enablement)
    - [Operations](https://cognistry.co/operations-teams)
    - [Product](https://cognistry.co/product-delivery)
    - [Sales Enablement](https://cognistry.co/sales-enablement)
    - [Strategy Transformation](https://cognistry.co/strategy-transformation)
    - [Talent Management](https://cognistry.co/talent-performance)
- [About](https://cognistry.co/about)
- [Resources](https://cognistry.co/resources)

    - [Get the Book!](https://cognistry.co/the-ai-lead-lambert)
    - [Recommended](https://cognistry.co/edge)
- [LOGIN](https://app.cognistry.co)

[Insights](https://cognistry.co/edge) / [Develop People](https://cognistry.co/edge/tag/develop-people)

# Your AI Agents Have Job Descriptions. Your People Don't.

![Brian Lambert, PhD](https://app.hubspot.com/settings/avatar/d1661e3b3c7efd9049a4635f72e4426a)

[Brian Lambert, PhD](https://cognistry.co/edge/author/brian-lambert)

 Oct 25, 2026, 7:30:00 AM ·  20 min read

![Your AI Agents Have Job Descriptions. Your People Don't.](https://media.babylovegrowth.ai/blog-images/organization-48833/1792742483029_Hand-drawn-role-spec-sketches-framing-the-title.jpeg)

#### In this article

1. Loading...

#### Share this article

[Twitter](https://twitter.com/intent/tweet?url=https%3A%2F%2Fcognistry.co%2Fedge%2Fgive-ai-job-decsriptions&text=%3Cspan+id%3D%22hs_cos_wrapper_name%22+class%3D%22hs_cos_wrapper+hs_cos_wrapper_meta_field+hs_cos_wrapper_type_text%22+style%3D%22%22+data-hs-cos-general-type%3D%22meta_field%22+data-hs-cos-type%3D%22text%22+%3EYour+AI+Agents+Have+Job+Descriptions.+Your+People+Don%27t.%3C%2Fspan%3E) [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fcognistry.co%2Fedge%2Fgive-ai-job-decsriptions) [Facebook](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Fcognistry.co%2Fedge%2Fgive-ai-job-decsriptions)

Copy link

#### About the Author

![Brian Lambert, PhD](https://app.hubspot.com/settings/avatar/d1661e3b3c7efd9049a4635f72e4426a)

[Brian Lambert, PhD](https://cognistry.co/edge/author/brian-lambert) brian@digitalcommand.co

[View all posts](https://cognistry.co/edge/author/brian-lambert)

Giving AI a job description is now standard practice. Give an agent a goal, a toolset, permissions, escalation rules and an evaluation suite, and it can be deployed with confidence. The person who supervises that agent usually gets two phrases: “use judgment” and “focus on higher-value work.” That asymmetry, more than model capability, is the operating-model risk we think leaders should worry about next.

This article shows the two documents side by side, names the gap, counts its cost, and offers a way to close it using the same fields you already use for agents.

## TL;DR

- Organizations specify AI agents in detail (goal, tools, permissions, escalation, evaluation) but describe human oversight in vague terms.
- We call this the Human Contribution Gap. It is our observation from the research and our work, not a proven finding.
- The visible costs are rubber-stamp review, blurred accountability and rework.
- The fix: write human work with the same fields: decision owned, context, criteria, and who accepts.
- Human work that isn’t specified can’t be priced or verified. That is a finance problem as much as an HR one.

## Table of Contents

- [Two documents, one workflow](https://cognistry.co/edge/give-ai-job-decsriptions#two-documents-one-workflow)
- [Naming the Human Contribution Gap](https://cognistry.co/edge/give-ai-job-decsriptions#naming-the-human-contribution-gap)
- [Why “human in the loop” is not a role](https://cognistry.co/edge/give-ai-job-decsriptions#why-human-in-the-loop-is-not-a-role)
- [What the gap costs](https://cognistry.co/edge/give-ai-job-decsriptions#what-the-gap-costs)
- [The CFO hook: unspecified work can’t be priced](https://cognistry.co/edge/give-ai-job-decsriptions#the-cfo-hook-unspecified-work-cant-be-priced)
- [The fix: write human work with the same fields](https://cognistry.co/edge/give-ai-job-decsriptions#the-fix-write-human-work-with-the-same-fields)
- [A note on legal hygiene](https://cognistry.co/edge/give-ai-job-decsriptions#a-note-on-legal-hygiene)
- [Designing oversight into the workflow](https://cognistry.co/edge/give-ai-job-decsriptions#designing-oversight-into-the-workflow)
- [Our perspective](https://cognistry.co/edge/give-ai-job-decsriptions#our-perspective)
- [Where Cognistry fits](https://cognistry.co/edge/give-ai-job-decsriptions#where-cognistry-fits)
- [Monday: one question for leaders](https://cognistry.co/edge/give-ai-job-decsriptions#monday-one-question-for-leaders)
- [FAQ](https://cognistry.co/edge/give-ai-job-decsriptions#faq)
- [Sources](https://cognistry.co/edge/give-ai-job-decsriptions#sources)

Cognistry

Specify the Human Work Around AI

 

Cognistry uses organizational evidence to diagnose required capabilities, determine whether learning is appropriate, and measure outcomes back to the business.

[Explore Cognistry](https://cognistry.co)

## Two documents, one workflow

Here is an agent spec for an invoice-exception workflow, followed by the human role that sits next to it.

| Field | AI agent spec | Human role description |
| --- | --- | --- |
| Goal | Resolve invoice mismatches under $5,000 | Ensure vendor payments are accurate and defensible |
| Tools | ERP read access, vendor email, ticketing | “Use available systems” |
| Permissions | Can draft adjustments, cannot post them | “Exercise appropriate authority” |
| Escalation | Escalate when confidence is low or vendor is flagged | “Escalate issues as needed” |
| Evaluation | Weekly accuracy test against a labeled set | “Meets expectations” in annual review |

The left column could be handed to an engineer and built. The right column could not be handed to anyone and acted on. Yet the person on the right is the one who decides what happens when the agent is wrong.

This example is illustrative, not a customer case. But if you have written an agent spec recently, you have probably seen the same shape in your own documents.

## Naming the Human Contribution Gap

We call this mismatch the **Human Contribution Gap**: the distance between how precisely an organization defines what an agent does and how loosely it defines what the human next to it contributes.

A caution first. The Gap is a proposed construct. We offer it as an observation that fits what we see in the research and in enterprise work, not as an established, measured finding. Treat it as a lens you can test against your own workflows.

The lens is useful because the evidence points in the same direction from several sides:

- Researchers studying how domain experts shift from doing tasks to overseeing AI output note that [undetected errors can have serious consequences](https://doi.org/10.1145/3742413.3789100), which makes the oversight role itself a design question rather than an afterthought.
- A working paper on [putting AI agents on the org chart](https://www.emmawiles.com/storage/ai_employee.pdf) examines delegation and accountability when firms assign agents designated roles and responsibilities.
- Work on [episodic oversight in generative AI workflows](https://link.springer.com/article/10.1007/s12525-026-00915-x), drawing on 15 interviews in two German tech firms, looks at how workers preserve human agency once model output enters real deliverables.

None of these papers uses our label. What they share is the question underneath it: what, exactly, is the human supposed to do?

## Why “human in the loop” is not a role

Most organizations answer that question with a slogan. “Keep a human in the loop” feels like a control. It is a placement, not a job.

Research on oversight design argues that effective oversight is not a static final check. It has to be built into the workflow, with humans and systems working as partners. The same research warns about automation bias, where people lose the ability to spot system errors after relying on the system. The ACM [co-design workshop study on human oversight](https://doi.org/10.1145/3742413.3789100) is a good starting point if you want the detail.

A slogan cannot tell a reviewer which decisions are theirs, what good looks like, or when to stop trusting the output. A role description can. Our own view on this is set out in [human in the loop as the missing capability in AI leadership](https://cognistry.co/edge/human-in-the-loop-leadership).

## What the gap costs

The cost of vague human work shows up in three places. You will recognize them.

### Rubber-stamp oversight

When the reviewer’s job is “check the output,” the easiest way to do that job is to approve. Volume rises, attention falls, and review turns into a click.

The oversight research is blunt that checking every output is both ineffective and frustrating. It points to episodic oversight instead: stepping in at specific moments such as drafting, refining and reviewing, and designing the work so the person feels their input adds something specific. Without that design, the human becomes a passive bottleneck.

### Blurred accountability

The org-chart paper above reports that [23% of surveyed managers](https://www.emmawiles.com/storage/ai_employee.pdf) said their organizations now list AI agents on organizational charts. Once an agent has a box and a role, it is easy for a manager to treat the agent as the owner of the result. The researchers’ framing of a “hot potato” effect is that managers may catch fewer errors themselves and shift accountability to the system.

When a bad output reaches a customer, who owned the decision? If the human role says “use judgment,” the honest answer is nobody in particular.

### Rework

Practitioners writing AI-governance role postings emphasize governance artifacts, decision pathways and oversight mechanisms over raw technical output. Look at a posting such as this [AI risk management and evaluation role](https://joblume.totalh.net/job/ai-risk-management-evaluation-sme), which centers on implementing a risk framework and documenting the work. Where those pathways are missing at go-live, teams discover them through mistakes: outputs get checked twice by different people, or by no one, and fixes land after the damage.

We are not offering a percentage for any of this. We have no customer study that supports one, and we would rather say so than invent a number.

## The CFO hook: unspecified work can’t be priced

Here is the sentence we would put in front of a finance leader: **human work that isn’t specified can’t be priced or verified.**

Consider what finance can do with an agent. It has a unit cost, an error rate, a volume and an evaluation. You can compare it with the process it replaced.

Now consider the human beside it. If their contribution is “use judgment,” finance cannot say what the organization is paying for, cannot tell whether the review is catching errors, and cannot decide whether adding a second reviewer or removing the first is the better call. The agent has a business case. The human has a salary line.

This matters at the top of the market too. AI talent is expensive: one [2026 salary analysis](https://cadence.withremote.ai/blog/ai-engineer-salary) puts median total compensation for applied AI engineering in the US at roughly $215,000. Whatever you pay the people who oversee and accept AI output, you should be able to say what decision they own.

![The CFO hook: unspecified work can't be priced — overview diagram](https://media.babylovegrowth.ai/blog-images/organization-48833/1792742525588_The-CFO-hook-unspecified-work-can-t-be-priced-overview-diagram.jpeg)

## The fix: write human work with the same fields

You do not need a new framework. Reuse the structure of an agent spec and apply it to the person. Four fields do most of the work.

### 1. Decision owned

Name the decisions this person makes that the agent does not. Be concrete: “approves or rejects vendor adjustments above $2,500,” not “ensures quality.”

Also name the decisions they cannot delegate to the agent. The research on cognitive degradation from long automation use recommends explicitly defined no-delegate zones and risk-tiered boundaries. Those belong in the role, in writing.

### 2. Context

What does the person need to see to decide well? The agent’s inputs, its confidence, the vendor history, the policy in force. If the reviewer sees only the final output, they are being asked to judge without evidence.

Context is also where design and training meet. Before anyone builds a course on “AI oversight,” ask what information the person actually has at the moment of decision. Often the gap is in the environment, not in the person.

### 3. Criteria

How does the person know the output is acceptable? Write the test: what must be true, what is a reject, what is a send-back. Give the three or four situations where the agent is known to fail.

This is also where you define when to intervene, when to trust, and how to document exceptions. Research on governance-style roles describes exactly those protocols as what lets an organization remain in control of the AI’s operational impact. See the [Electronic Markets protocol paper](https://link.springer.com/article/10.1007/s12525-026-00915-x) for one structured approach to intervening at specific episodes.

### 4. Who accepts

Every agent output ends up accepted by someone. Name that person, and name who accepts *their* decisions when the stakes are high. This single field turns blurred accountability into a chain you can audit.

### A filled-in example

| Field | Human role: invoice exception reviewer |
| --- | --- |
| Decision owned | Post, reject or escalate each adjustment the agent drafts; no adjustments to flagged vendors without a second approver |
| Context | Agent’s draft, its stated confidence, vendor payment history, current payment policy |
| Criteria | Matches contract terms; amount reconciles to receipt; no unexplained pattern across the vendor’s last invoices |
| Who accepts | Accounts payable manager for the reviewer’s decisions; controller for escalations |
| Escalation | Any mismatch tied to a contract dispute goes to the controller the same day |

It takes about a page. It can be priced, reviewed and improved.

## A note on legal hygiene

If you turn these fields into formal job descriptions, involve HR and counsel. SHRM’s guidance is that AI tools are useful for drafting job descriptions but [cannot replace human judgment](https://www.shrm.org/topics-tools/employment-law-compliance/viewpoint-ai-tools-cant-replace-authentic-thinking-crafting-job-descriptions) for legal and operational compliance. Essential functions under the ADA should be defined by job-related necessity, and the same guidance advises against assigning percentages to tasks, which can create credibility issues in employment disputes.

So keep the decision-rights spec as an operating document that sits beside the formal job description, and review both at least annually. Employers in North Carolina, South Carolina, Florida, Georgia, Virginia and Maryland should confirm state-specific requirements with employment counsel before changing formal position documents.

## Designing oversight into the workflow

A role spec only works if the workflow gives the person a real chance to do the job. That means the reviewer sees the right context at the right moment, the volume is humanly reviewable, and exceptions have a place to be recorded.

This is where governance and workflow design meet. If you want a practical grounding in how organizations structure controls around models, this overview of [AI model governance practices](https://forgeprivate.ai/blog/ai-model-governance) is worth a read. For the people side, we cover [operational AI governance with a diagnosis-first start](https://cognistry.co/edge/operational-ai-governance) and [why decision capability underpins AI and human collaboration](https://cognistry.co/edge/decision-capability-core-ai-human-co-work).

## Our perspective

We see the same pattern from the training side. When an agent rollout goes wrong, the first request is often “build us a course on working with AI.” Before you build that course, ask a different question: what capability does this work actually require, and is learning the right enablement play at all?

Sometimes the answer is a course. Often it is a missing decision right, an unclear escalation path, or a review screen that hides the context. No amount of training fixes a role that nobody defined. That is an environment problem, not a person problem.

![Our perspective — overview diagram](https://media.babylovegrowth.ai/blog-images/organization-48833/1792742562926_Our-perspective-overview-diagram.jpeg)

So we start with diagnosis. We look at the organization’s own material: strategy, frontline friction, quality findings and what the experts in the room know. From that we decide what the human has to decide, and only then whether to build practice for it. Practice means decision environments where people exercise judgment on realistic cases, not slides that test recall.

The Human Contribution Gap is our name for what we keep finding at the start. Use it if it helps. Challenge it if it doesn’t fit your data.

> *— Brian*

## Where Cognistry fits

If you have drafted the four fields for a role and found that people cannot yet do what the role demands, that is where the [Cognistry Platform](https://cognistry.co/platform) comes in. It diagnoses the capability the work requires, grounds design in your own evidence, and structures the response, including [practice simulations](https://cognistry.co/sim) where people rehearse oversight decisions before go-live. Outcomes are measured back to the business, not to course completion.

## Monday: one question for leaders

Pick one workflow where an agent already does real work. Then ask the person next to it:

**“Which decision here is yours, and how would we know you made it well?”**

If the answer is fast and specific, you have a role. If it is a pause, a shrug or “I review everything,” you have found the gap. Write the four fields before the next release.

## FAQ

### What does it mean to give AI a job description?

It means documenting an agent’s goal, tools, permissions, escalation rules and evaluation criteria in one place, the way you would specify a role. It makes the agent’s scope reviewable and testable. The same discipline should extend to the humans who supervise it.

### Is the Human Contribution Gap an established finding?

No. It is a proposed construct and our own observation. Related research on oversight, delegation and accountability supports the underlying concern, but none of it measures the gap under that name.

### Can AI write the human job descriptions too?

It can draft them. SHRM’s guidance is that AI tools [cannot replace human judgment](https://www.shrm.org/topics-tools/employment-law-compliance/viewpoint-ai-tools-cant-replace-authentic-thinking-crafting-job-descriptions) on legal and operational compliance, so have HR and counsel review before anything becomes official.

### Does every AI output need human review?

Probably not. Research on oversight suggests checking everything is ineffective and frustrating. Risk-tier your workflows, name the no-delegate zones, and place human review at specific episodes where it adds value.

### Is training the answer to weak human oversight?

Sometimes, but diagnose first. If the cause is an unclear decision right or missing context, training will not fix it. Define the role, then decide whether practice is needed.

## Sources

- [Design Considerations for Human Oversight of AI (ACM)](https://doi.org/10.1145/3742413.3789100)
- [Episodic oversight in generative AI workflows (Electronic Markets)](https://link.springer.com/article/10.1007/s12525-026-00915-x)
- [Putting AI on the Org Chart: Evidence on Delegation and Accountability](https://www.emmawiles.com/storage/ai_employee.pdf)
- [AI Risk Management & Evaluation SME posting](https://joblume.totalh.net/job/ai-risk-management-evaluation-sme)
- [AI engineer salary in 2026 (Cadence)](https://cadence.withremote.ai/blog/ai-engineer-salary)
- [SHRM: AI tools can’t replace authentic thinking in crafting job descriptions](https://www.shrm.org/topics-tools/employment-law-compliance/viewpoint-ai-tools-cant-replace-authentic-thinking-crafting-job-descriptions)
- [AI Model Governance: A Practical Guide for Enterprises](https://forgeprivate.ai/blog/ai-model-governance)

## Recommended

- [Human Capital Management for an AI World: Start With Capability](https://cognistry.co/edge/human-capital-management-for-an-ai-world)
- [Human in the Loop: The Missing Capability in AI Leadership](https://cognistry.co/edge/human-in-the-loop-leadership)
- [Why decision capability is the foundation of AI human collaboration](https://cognistry.co/edge/decision-capability-core-ai-human-co-work)
- [AI Leadership Development Is Becoming the Most Important Leadership Skill](https://cognistry.co/edge/ai-leadership-development)

[Develop People](https://cognistry.co/edge/tag/develop-people)

## Continue Reading

[![Human Capital Management for an AI World: Start With Capability](https://media.babylovegrowth.ai/blog-images/organization-48833/1792201631886_Sketch-motifs-frame-a-clear-title-area.jpeg)](https://cognistry.co/edge/human-capital-management-for-an-ai-world)

Develop People

### [Human Capital Management for an AI World: Start With Capability](https://cognistry.co/edge/human-capital-management-for-an-ai-world)

 Human capital management in an AI world is not a software upgrade. It is a decision about which...

Brian Lambert, PhD · 20 min read

[![Redefining Roles for the AI Economy: A Diagnosis-First Guide](https://media.babylovegrowth.ai/blog-images/organization-48833/1792386885484_Hand-drawn-AI-role-map-title-card.jpeg)](https://cognistry.co/edge/redefining-roles-for-the-ai-economy)

Develop People

### [Redefining Roles for the AI Economy: A Diagnosis-First Guide](https://cognistry.co/edge/redefining-roles-for-the-ai-economy)

 Redefining roles for the AI economy means working out which tasks AI absorbs, which judgment calls...

Brian Lambert, PhD · 19 min read

[![Defining the AI and Human Operating Model](https://media.babylovegrowth.ai/blog-images/organization-48833/1792633362877_Hand-drawn-operating-model-title-card.jpeg)](https://cognistry.co/edge/defining-the-ai-and-human-operating-model)

Organization

### [Defining the AI and Human Operating Model](https://cognistry.co/edge/defining-the-ai-and-human-operating-model)

 An AI and human operating model states how people, AI agents, data and processes work together to...

Brian Lambert, PhD · 20 min read

[![Cognistry\_Icon\_White](https://cognistry.co/hs-fs/hubfs/Logos/Cognistry_Icon_White.png?width=150&height=60&name=Cognistry_Icon_White.png "Cognistry_Icon_White")](https://cognistry.co)

- [Cognistry](https://cognistry.co)
- [Resources](https://cognistry.co/resources)
- [Sign In](https://app.cognistry.co)
- [Privacy](https://cognistry.co/privacy)
- [Terms](https://cognistry.co/terms)

© 2026 Cognistry. Capability by Design.

```json
{
  "@context" : "https://schema.org",
  "@type" : "BreadcrumbList",
  "itemListElement" : [ {
    "@type" : "ListItem",
    "item" : "https://cognistry.co/edge",
    "name" : "Insights",
    "position" : 1
  }, {
    "@type" : "ListItem",
    "item" : "https://cognistry.co/edge/tag/develop-people",
    "name" : "Develop People",
    "position" : 2
  }, {
    "@type" : "ListItem",
    "name" : "Your AI Agents Have Job Descriptions. Your People Don't.",
    "position" : 3
  } ]
}
```

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Brian Lambert, PhD",
    "url" : "https://cognistry.co/edge/author/brian-lambert"
  },
  "dateModified" : "2026-10-26T17:11:32.781Z",
  "datePublished" : "2026-10-25T11:30:00.000Z",
  "headline" : "Your AI Agents Have Job Descriptions. Your People Don't.",
  "image" : [ "https://media.babylovegrowth.ai/blog-images/organization-48833/1792742483029_Hand-drawn-role-spec-sketches-framing-the-title.jpeg" ],
  "mainEntityOfPage" : {
    "@id" : "https://cognistry.co/edge/give-ai-job-decsriptions",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://cognistry.co/hubfs/cognistry_primary-1.png"
    },
    "name" : "Cognistry Digital Command"
  }
}
```