Fourteen Agents, and the Order to Build Them
The full portfolio, what each one needs first, and where to start
The program that switched on all at once
A portfolio of agent types covers almost the entire practical opportunity across the employee lifecycle. The order you build them in decides whether the program survives its first year.
One large employer we studied stood up nine agents in a single quarter. Mobility, benefits, hiring, performance, all live on the same launch day, all announced in the same all-hands. It looked decisive. Six months later the program was frozen. Not because any single agent failed, but because nothing had been sequenced. The knowledge base was thin, so every agent hedged. Managers got three different tools in one week and adopted none. The one agent that actually worked, a quiet policy assistant, was lost in the noise of eight that did not.
That is the mistake this chapter exists to prevent. There is a defensible portfolio of agent types, and there is an order to build them. Get the roster right and you still fail if you build it in the wrong sequence. The portfolio below is fourteen core agent types plus one governance agent, the Compliance Monitor, that some organizations fold into their trust and governance function rather than run as a standalone. Fourteen you build for people. One you build to watch the other fourteen.
The full portfolio, grouped by who it serves
Group agents by the person on the other side of the conversation. It changes how you staff, sponsor, and sequence them.
Read each profile for one thing above all: what it must never do. The boundary defines the agent more than the feature list.
Personal Work Agent
The front door. Handles routine requests, PTO, expenses, letters, and hands off cleanly to specialists. Serves every employee.
Never: approve exceptions or interpret gray-area policy on its own.Policy Guide
Answers "what is our policy on this" with a citation, in plain language, across HR, IT, travel, and conduct. Serves everyone.
Never: invent a policy or answer where the source is silent.Onboarding Guide
Walks a new hire through the first ninety days: paperwork, access, introductions, and the questions they are too new to ask. Serves joiners.
Never: grant system access or complete legal attestations for the person.Benefits Navigator
Explains coverage, life events, and enrollment tradeoffs in the employee's own terms. Serves everyone, heavily at open enrollment.
Never: give tax, legal, or medical advice, or enroll on the person's behalf.Mobility Advisor
Surfaces internal roles, gigs, and mentors matched to skills and stated goals. Serves employees looking to move.
Never: move someone or promise a role a hiring manager has not offered.Capability Coach
Turns a skill gap into a plan: learning, stretch work, and feedback loops tied to real openings. Serves developing employees.
Never: rate a person's competence as an official record.Transitions Assistant
Guides the mechanics of a change already decided: relocation, role switch, return from leave, or exit logistics. Serves people mid-change.
Never: decide whether the transition should happen.Leadership Assistant
The manager's daily aide: prep for one-on-ones, team context, draft messages, and the answer to "how do I handle this." Serves managers.
Never: deliver difficult news to a report on the manager's behalf.Performance Partner
Assembles evidence, prompts for specifics, and drafts fair, balanced reviews. Serves managers running performance cycles.
Never: assign a rating or make the calibration call.Workforce Analyst
Answers people-data questions, attrition, span, cost, skills coverage, without a spreadsheet detour. Serves leaders and HRBPs.
Never: expose individual data beyond the asker's authorization.Engagement Listener
Reads sentiment signals and open text, then routes themes to the right owner. Serves leaders watching team health.
Never: deanonymize a response or attribute a comment to a person.Hiring Assistant
Drafts requisitions, screens against defined criteria, and schedules. Serves recruiters and hiring managers.
Never: reject a candidate or make the hire decision.Compensation Analyst
Models pay scenarios against bands, equity, and market data so a human can decide with clean numbers. Serves comp teams and managers.
Never: set or commit a pay outcome.Employee Relations Assistant
Handles intake, documentation, and routing for sensitive cases: it captures the facts and gets them to the right person, fast. Serves ER and HRBPs.
Never: reach a finding or make a decision. It is intake and paperwork, never judgment.Compliance Monitor
The governance agent. Watches the other agents for policy drift, bias signals, and out-of-bounds actions, and flags them. Serves trust and governance.
Never: serve employees directly or act on the violations it detects.Value against effort, on one map
Plot each agent by how hard it is to build against how much value it returns. Two names sit in the corner you want.
The map is not a ranking. A point low and to the right, high effort for modest value, is not a bad agent; it is a bad first agent. The specialist agents cluster there because they touch regulated decisions and the richest data, which is exactly why they should wait until the ground under them is solid.
Why Policy Guide and Leadership Assistant go first
The argument, not the assertion. Each first pick earns its place for a different reason, and together they cover the two things a young program needs most.
Start with the Policy Guide because it carries the highest question volume at the lowest stakes. Nobody is harmed by a well-cited answer about carryover PTO, and everybody asks. That volume is the point. To answer policy questions well, you have to write down, structure, and reconcile the knowledge that every other agent will later depend on. The Policy Guide forces the organizational context work you were going to need anyway, and it does that work under the safest possible conditions. If it is wrong, it is wrong about something recoverable, and you learn fast.
Pair it with the Leadership Assistant because it buys you the fastest visible win with your most influential internal population. Managers set the weather for adoption. When a manager saves twenty minutes before every one-on-one, they tell other managers, and they stop treating the program as an HR project. The Leadership Assistant is also forgiving: it drafts and prepares, it does not decide, so its early rough edges cost credibility rather than compliance.
One builds the foundation quietly. The other builds belief loudly. A program needs both in its first two quarters, and almost every other agent gets easier once you have them.
Build the agent that forces the knowledge work first, and the agent that earns political air cover second. Everything else is downstream of those two.
What each one needs before it works
Every agent has a floor of company knowledge and system access below which it cannot function. This is the organizational context layer from Chapter 8, made concrete.
Sequence by prerequisite, not by ambition. An agent that needs write access to your system of record and a clean skills taxonomy is a later build than one that needs a well-tagged policy library, no matter how much you want it.
| Agent | Company knowledge it needs | System access it needs |
|---|---|---|
| Policy Guide | Tagged, versioned policy library | Read-only knowledge sources |
| Leadership Assistant | Org structure, manager norms | Read: HCM, calendar, comms drafts |
| Personal Work Agent | Service catalog, escalation rules | Read and write: system of action |
| Onboarding Guide | Role-based onboarding paths | Read: provisioning status, HCM |
| Benefits Navigator | Plan documents, life-event rules | Read: benefits system |
| Mobility Advisor | Skills taxonomy, open roles | Read: talent marketplace, profiles |
| Capability Coach | Skills model, learning catalog | Read: LMS, skills graph |
| Transitions Assistant | Process maps for each change type | Read and write: system of action |
| Performance Partner | Review framework, calibration rules | Read: performance and goals data |
| Workforce Analyst | Metric definitions, access policy | Read: people-data warehouse |
| Engagement Listener | Survey model, anonymity thresholds | Read: survey and sentiment feeds |
| Hiring Assistant | Job architecture, screening criteria | Read and write: ATS |
| Compensation Analyst | Pay bands, equity, market data | Read: comp and market systems |
| Employee Relations | Case types, documentation standards | Write: case system, read HCM |
| Compliance Monitor | Policy rules, bias and audit criteria | Read: every agent's logs |
Which side of the human line each one sits
Where an agent sits against the human line from Chapter 5 is not a detail. For the highest-stakes agents, it is the entire design.
Most of the everyday agents live comfortably on the agent-led side: routine, reversible, low cost if wrong. The growth agents advise and prepare, then hand a decision to a person. But three agents sit deliberately on the human side and never cross it, no matter how capable they become. The Employee Relations Assistant does intake, documentation, and routing, and never reaches a finding. The Compensation Analyst models scenarios and never sets pay. The Performance Partner assembles evidence and never assigns a rating.
These three do everything except the one thing a person must own. That single restraint, agent everywhere in support, human at the decision, is this entire book in miniature.
If you design those three to make the call "just this once," you have not built a faster HR function. You have built a liability with a chat window. Keep them in support, and they become the safest high-value agents you run.
- Pick the two starters and name a single accountable owner for each. No committees.
- Audit your policy content: what is written down, versioned, and tagged, and what lives only in people's heads.
- Interview ten managers about the twenty minutes before a one-on-one. Design the Leadership Assistant around what you hear.
- Draft the "must never do" line for both agents and get legal to sign it before a single prompt ships.
- Set one adoption metric per agent and a date to read it. Kill or continue on evidence, not enthusiasm.
- Turning everything on at once. The wide launch feels bold and produces a frozen program six months later. Adoption is serial, not parallel.
- Starting with the exciting agent. Hiring and comp agents demand the deepest data and touch regulated decisions. Build them first and you learn your hardest lessons where they cost the most.