Skip to content

Private Demo Mode

The 2026 People ops Playbook

Why this changed

Most people teams are a help desk in disguise, whatever the org chart says. You hired a head of people to build a culture, design a review cycle, and coach managers, and then you watched that person spend most of every week answering the same forty questions, chasing signatures, and keying changes into payroll before the cutoff.

That was the price of a small team. A company of 150 people usually has one or two people in the function, and the ratio guides from the last decade said that was correct: roughly one people person for every 70 to 100 employees. The ratio assumed that every request needed a human to read it, find the policy, open the system, and make the change. That assumption is now wrong for most of the work.

An agent can read the handbook, the benefits summary, and the payroll calendar. It can open Rippling or Gusto with a scoped login, prepare a change, and put it in front of a person for one click. It can send the I-9 reminder on day two, file the laptop ticket, and check that the new hire actually enrolled in benefits before the 30-day window closes. None of that needs judgment.

This playbook comes from Ines Halvorsen, who runs people operations for 214 people at Kestrel Freight with one other person and a fleet of agents. Three other people leads gave us their methods, their numbers, and the places where they stopped automating on purpose. The question we want to answer for you is simple: which parts of people ops should you hand off now, and which parts must a person always hold?

“I didn’t get into this work to reset passwords and explain the dental plan. Once the agents took the queue, I finally had time for the thing my CEO hired me for, which was managers.”

Sort the work first

Before you buy anything, sort every request your team got last quarter into two piles. One pile is work where the answer exists already, in a policy, a plan document, a system record, or a legal rule. The other pile is work where the answer depends on a person: their history, their feelings, their manager, and what is fair.

Pull 90 days of tickets, Slack messages to the people channel, and emails to your HR inbox. Tag each one. When Dayo Adeyemi did this at Fernway Health, the result was 1,312 requests from 143 employees. Of those, 1,018 (78 percent) had a written answer somewhere. Another 211 were changes to a record, such as an address, a bank account, a tax form, or a title. Only 83 needed a person to decide something.

  • Policy questions: PTO balances, parental leave rules, expense limits, holiday calendars, the remote work stipend.
  • Benefits questions: what the plan covers, how to add a spouse, when open enrollment closes, how the HSA match works.
  • Record changes: address, bank details, W-4 updates, name changes, manager changes, title changes after a promotion is approved.
  • Lifecycle tasks: offer letter assembly, background check start, onboarding checklists, equipment, access, offboarding.
  • Compliance tasks: I-9 timing, state new-hire reports, required training, posting updates, visa and permit dates.
  • Judgment work: performance, leave with medical detail, conflict between people, accommodations, terminations, pay decisions.

The first five lines are where agents go. The last line is where you go. Most teams find that the last line is 5 to 10 percent of the volume and almost all of the risk. That ratio is the whole argument of this playbook.

What stays with a person

Some moments belong to a person, no matter how good the agent is. A performance conversation, a leave request that carries a diagnosis, a complaint about a manager, a termination: in each of these, the employee needs to see that a person heard them, weighed it, and accepts the responsibility for the outcome. An agent can prepare. It cannot be the one who decides.

This is not only about kindness. It is also about legal exposure. A leave request may trigger FMLA or a state program, an accommodation request starts an interactive process, and a complaint may be a protected report. If an agent answers one of these with a policy paragraph, you may have missed a duty you had from the moment the message arrived. So the rule is not “the agent answers carefully.” The rule is “the agent does not answer at all, and it gets a person within the hour.”

  • Performance: reviews, improvement plans, and any message that says someone is not meeting expectations.
  • Leave and health: any request that mentions a medical condition, pregnancy complications, mental health, or a family emergency.
  • Conflict and complaints: harassment, discrimination, retaliation, or any report about a manager or coworker.
  • Accommodations: requests for changes to work because of a disability or a religious practice.
  • Terminations and resignations under pressure: the conversation, the final pay decision, and the severance terms.
  • Pay decisions: raises, equity refreshes, offer exceptions, and any reply to “why do I make less than…”.

For each of these lines, write a trigger list and give it to the agent as a stop rule. When a message matches, the agent replies with one fixed sentence that says a named person will contact them today, and it puts the thread at the top of that person’s queue with the full context. Hana Moriyasu at Quarry Studio measured her escalation time before and after the change: from a median of 19 hours to 41 minutes, because the sensitive messages no longer sat in a general inbox behind benefits questions.

“The agent made us faster on the hard cases too, which surprised me. It doesn’t touch them. It just finds them and puts them in front of me before anything else.”

The five-step framework

Hand the work over in a fixed order, one lane at a time. Each step makes the next one safer, because each one gives the agent better source material and gives you a record of how it behaves. Teams that skipped to payroll changes in the first month usually rolled back within a quarter.

  1. 01Write the answers down. Put your handbook, benefits summaries, PTO rules, payroll calendar, and expense policy in one place the agent can read. Fix the contradictions you find; there are always some.
  2. 02Start with questions only. Let the agent answer policy and benefits questions in Slack and email for four weeks, read-only, with a link to the source in every answer. Read a sample of 50 answers each week.
  3. 03Add lifecycle checklists. Give the agent onboarding and offboarding: it sends the documents, opens the tickets, chases the signatures, and reports what is late. Nothing it does here changes pay.
  4. 04Add record changes with approval. The agent prepares address, bank, tax, and title changes in the HRIS and payroll system, and a person approves each one. Track the approval rate and the corrections.
  5. 05Expand approval-free changes slowly. After 60 days with no corrections on a change type, let the agent make that type without approval, but keep bank details and pay amounts behind a person forever.

The second step is where most teams learn the most. Tomás Achterberg found that 14 of the first 200 agent answers at Lumen Ledger were wrong, and in 12 of those cases the handbook itself was wrong or out of date: an old parental leave length in the Netherlands, a stipend amount from 2024, a holiday that the company had dropped.

“The agent is a very honest mirror. If your policy is vague, it will be vague. If your handbook says two different things, it will say both. Clean the source and the answers get clean.”

A decision rule for the last step: approval-free only when a mistake is cheap to find and cheap to reverse. A wrong address on an employee record is found at the next mailing and fixed in a minute. A wrong bank account sends a paycheck to a stranger, and you may not get that money back. Those two changes look the same in the system. They do not carry the same risk, so they do not get the same rule.

Onboarding and offboarding

Onboarding is the best first job for an agent, because it is long, it has many small deadlines, and every missed step is visible to a new hire who is still deciding if they chose well. A typical onboarding at a 200-person company touches eight systems and has 30 to 45 tasks across the people team, IT, the manager, and payroll.

At Kestrel Freight, Ines gave one agent the full onboarding run for each new hire, from signed offer to day 30. It sends the documents from Rippling, checks that the I-9 section one is done before the first day, books the section two review, files the laptop request, adds the person to the correct Slack channels, sends the manager a first-week plan template, and checks benefits enrollment on days 7, 20, and 27. Before the change, Kestrel missed at least one step for 31 percent of new hires. In the two quarters after, that number was 4 percent, and each miss was a manager task the agent had reminded them about twice.

Offboarding is the same job in reverse, and the risk is larger. The agent runs the checklist the moment a departure is confirmed: it schedules access removal for the last day, starts the equipment return, prepares the final pay summary for a person to review against state timing rules, sends COBRA information through the provider, and confirms each step closed. A person still holds the exit conversation, decides any severance, and reads the final pay before it runs.

“Our audit last year found nine accounts still open for people who had left. This year it found none. That isn’t because I got more careful. It’s because the agent never gets tired of checking.”

Payroll and benefits

Payroll is the place where you should go slowest, because a mistake reaches a person’s bank account and their trust in the company at the same time. That does not mean you keep agents out. It means the agent prepares and checks, and a person releases. Most of the payroll errors at small companies come from late inputs, not from bad math, and agents are excellent at chasing inputs.

Here is the worked example from Fernway Health, which runs semimonthly payroll in Gusto for 143 people in 19 states. Three business days before each cutoff, the agent compares the payroll draft to the HRIS: new hires, exits, title and pay changes approved since the last run, unpaid leave, and commission files from the sales team. In the first six runs it found 23 differences. Nineteen were real gaps, such as a raise approved in the review tool but never entered in payroll, and four were timing differences it correctly flagged as expected. Dayo now spends about 40 minutes per run in review, down from about 4 hours.

Benefits questions are the opposite case: high volume, low risk, and well documented. At Fernway, benefits questions were 36 percent of all requests and peaked at 140 in the two weeks of open enrollment. The agent answers from the plan summaries and the carrier documents, links to the page it read, and sends any claim dispute or coverage denial to a person, because those often hold health details. During last open enrollment, the agent answered 412 questions, and employees rated 94 percent of the answers as helpful.

International hires

Your first hire in a new country is a small project, with large consequences. You must choose between a contractor agreement, an employer of record, or your own entity, and each choice changes the tax, the benefits, the notice period, and the misclassification risk. An agent should not make that choice, but it can do nearly all the research and paperwork around it.

Lumen Ledger has 88 people in 11 countries, and 37 of them are employed through an employer of record. Tomás uses an agent to prepare a country brief for each new hire before the offer: statutory benefits, the usual market extras, the notice period, probation rules, the thirteenth-month salary where it applies, and the total employer cost at the proposed salary. It reads the provider’s country guide and the local government pages, and it marks each line with its source. A person checks the brief and makes the offer. The brief used to take Tomás a day. It now takes him 25 minutes to review.

After the hire, the agent keeps the calendar: visa and permit expiry dates, probation end dates, local holiday calendars for payroll, and the dates when a contractor’s hours or length of service start to look like employment. In March it flagged a contractor in Portugal who had worked full time for 14 months under one manager. Lumen converted that person to an employee through its provider before it became a problem.

  • RipplingPayroll, HR, IT, and spend management softwareHR, IT, and finance apps on one platform
  • DeelGlobal payroll, hiring, and HR compliance for distributed teamsHire, manage, and pay anyone, anywhere
  • GustoPayroll, benefits, and HR software for small businessesPayroll, benefits, and everything else
“The provider handles the payroll in each country. What I needed was something that watched all eleven countries at once and told me what was coming in the next 60 days. That is the agent’s job now.”

Rank the stack

Your HRIS and payroll system is the record that the agent reads from and writes to, so choose it for a clean API and clear permissions, not for its own chatbot. The agent needs to read employee records, prepare changes, and leave every change for a person to approve. A system that hides its data behind exports will slow every step in this playbook.

HRIS and payroll for an agent-run people teamEach of our four contributors scored the tools they have run in production from 1 to 10 on API coverage, permission scopes, approval flows, and report quality. The score is the mean of their votes.
  1. 01Rippling · Best for us companies of 100 to 1,000 people that want hr, it, and payroll in one recordThe widest API and the most granular permissions of the group. Device and app access sit in the same record as employment, so one agent can run a full onboarding or offboarding.8.8
  2. 02Deel · Best for teams that hire in many countries through an employer of record or contractorsThe best country coverage and clear per-country guides an agent can read. Its HRIS features are thinner than Rippling’s for US-only teams.8.3
  3. 03Gusto · Best for us companies under 200 people with simple payrollSimple, readable payroll reports and a fast approval step. The API covers less of the HR record, so expect some changes to stay manual.7.9
  4. 04HiBob · Best for mid-size companies with offices in the us and europe that pay through a separate payroll providerA strong employee record and good workflows. Payroll sits in a different system, which gives the agent one more place to reconcile.7.2
  5. 05Justworks · Best for small us teams that want a peo to carry benefits and complianceGood benefits for a small team, but the narrowest API here. The agent can read more than it can prepare, so plan on more manual entry.6.6

So what do you do?

Start with the audit, not the tool. Ninety days of requests, tagged into the six lines above, will tell you how much of your team’s week is answers that already exist and how much is real judgment. For most teams, the first number is above 75 percent. That number is your case to your CEO, and it is your plan for the next two quarters.

  1. 01This week: export 90 days of requests and tag them. Count the share that has a written answer.
  2. 02Within two weeks: put the handbook, benefits summaries, and payroll calendar in one readable place, and fix every contradiction you find.
  3. 03Within a month: write your stop list and your escalation sentence, then turn on question answering, read-only, with sources.
  4. 04Within a quarter: hand over onboarding and offboarding checklists, then the payroll diff with a person on every release.
  5. 05Every month: read 50 agent answers, review the stop list, and decide which change types have earned approval-free status.

Then use the time. The goal is not a smaller people team. The goal is a people team that spends its week on managers, on reviews that are fair, on the leave request that needs a careful call, and on the exit conversation that someone will remember for years. Ines put it best when we asked what changed for her.

“I used to measure my week in tickets closed. Now I measure it in conversations I had with managers. Last month that was 38. The year before, I couldn’t have told you, because there weren’t enough to count.”