Giving Every Employee a Personal Work Agent
What it knows, what it does, and why it is not a chatbot
The assistant that forgets you
Most of what gets sold to employees as AI is a stranger who resets every morning.
You ask the company chatbot how much time off you have left. It gives you a link to the policy. You ask again next week and it has no idea you ever asked, no idea who you are, no idea you have a trip booked. It answers inside a single window and forgets the moment you close it. It is a search box that talks.
A Personal Work Agent is the opposite of that. It is a persistent, secure, context-aware agent assigned to one person, one that knows their role, their team, their goals, and their entitlements, and can reach the company knowledge it is allowed to touch to actually get something done on their behalf. It holds context. It completes multi-step work across systems. And it carries something forward from last week, so the second conversation starts where the first one ended. That difference, session memory versus a working relationship, is the whole chapter.
It knows only what it is allowed to know
Useful and safe are not in tension. The boundary is what makes the agent trustworthy enough to be useful.
An agent that knows nothing about you is a search box. An agent that knows everything about you is a liability. The design problem is drawing the line between the two: hold the context that lets the agent help you get real work done, and never hold the context it has no business touching in the first place. That line is not an afterthought you bolt on later. It is the feature.
A Personal Work Agent works across four domains of context. It knows the individual: role, tenure, skills, goals, PTO balance, entitlements. It knows their team and manager: reporting line, team objectives, who approves what. It knows company policy and job architecture: the leave rules, the pay bands, the definition of every role. And it reads the outside environment: which internal roles are open, which projects need people, what the labor market is doing to a given skill.
Four domains of context, one strict rule for each: hold what helps, surface only to the person it belongs to.
The individual
Role, tenure, skills, stated goals, PTO balance, benefits and entitlements. The working memory of one person's professional life inside the company.
Their team and manager
Reporting line, team objectives, approval paths. Enough to route work correctly, not enough to read a colleague's private record.
Policy and job architecture
Leave rules, pay bands, the skills and responsibilities that define each role. The rulebook, applied to this person's actual situation.
The outside environment
Open internal roles, projects short on people, how the market values a given skill. The context a person would never assemble on their own.
It holds your data to serve you. It never holds a coworker's private record, your manager's notes on other people, or information above your entitlements. Your agent answers to you: your manager's agent cannot read your salary history, and yours cannot read theirs. When the agent acts across a boundary, such as notifying your manager, it does so on your instruction and shows you exactly what it sent. Governance for the wider program is Chapter 10; the point here is narrower and non-negotiable, an agent that cannot be trusted with the boundary is not one you should deploy.
From a sentence to finished work
Answering is the floor, not the ceiling.
A Personal Work Agent operates on a spectrum. At the simple end it answers a question with the specific answer, not a link. One step up, it drafts: a message, a request, a plan. Further along, it completes structured work, filling in the forms and fields a task requires. Then it coordinates multi-step work across systems, moving a request through the systems of record and action that HR runs on. It recommends learning tied to a real goal. And at the far end it surfaces opportunities the person would never have found alone, an open role, a project, a mentor, that match where they are trying to go.
- Answer. The direct answer to a real question, pulled from policy and the person's own record, not a document to go read.
- Draft. A message, a plan, a request written and ready for the person to approve or edit.
- Complete. Forms filled, fields populated, the routine work a task demands done end to end.
- Coordinate. Multi-step work moved across systems, from request to approval to record, without the person chasing it.
- Surface. A role, a project, or a path the person did not know existed and would not have found alone.
Four requests, four ordinary Mondays
The same agent, four sentences a real employee might say, and what actually happens next.
Read these as the person would say them, out loud, on a Tuesday. Each sentence looks like a question. It is not. Each is a piece of multi-step work spread across systems, policy, and people, compressed into a single spoken line. Watch what the agent retrieves, what it drafts, where it routes the work, and the exact moment it stops and hands the rest to a human.
I want to be a senior engineer within two years.
My father was admitted to the hospital last night.
This one is different, and it is the reason the rest of the book exists. The agent does not console the person. It does not say it is sorry. It does not handle the situation. It does exactly one thing, and it does it well: it clears every piece of administrative friction standing between the person and the moment they can stop thinking about paperwork, so the human conversation, with a manager, with family, happens sooner and better.
Help me get ready for my performance review.
I want to move to a different team.
Standing one up for your own people
The good news is you can put a real version in front of people fast. The bad news is a shallow one erodes trust just as fast.
- Pick one population, say engineering, and one person's real questions to design against, not a generic FAQ.
- Wire the agent to two sources first: the individual's own record and the policy and job architecture. Prove it can answer with specifics, not links.
- Write the context boundary down before you build: what it holds, what it never holds, who can see what. Get security to sign the one-pager.
- Ship the "answer" and "draft" levels only. Hold "complete" and "coordinate" until the boundary is proven in the two safe domains.
- Run the four requests above as your acceptance test. If request two makes the agent try to comfort someone, send it back.
- You ship a chatbot and call it an agent. It answers in a session and forgets, and people stop asking within a week.
- You skip the boundary to move faster, and the first time an agent surfaces something it should not have, the whole program loses the room.
- You let it reach for warmth. An agent that offers sympathy on a family emergency does not feel helpful; it feels hollow, and it undercuts the human it was supposed to free up.
- You over-scope. It tries to complete high-stakes work before it has earned trust on the simple answers, and one bad automated action costs you a year of goodwill.
The Personal Work Agent is the unit the rest of this book builds on. Get one right for one person and you have the pattern for the whole company. The hardest part is not the technology. It is the discipline in request two: holding the line, under pressure to look helpful, between the work the agent should take off a person's plate and the human moment it should never touch. That line has a name, and the next chapter is about how to draw it.