HR Chatbot: IBM’s Bot Handles 94% of Routine Questions. Its First Score Was Minus 35.

An HR chatbot is a chat assistant that answers the routine questions people bring to HR, such as leave rules, payslips, policy wording and onboarding steps, from the firm's own written policies. Anything sensitive goes to a person. Most small firms will never build what IBM built. The useful part is the order in which it worked: routine questions first, training forced and funded, judgment left with people, and a satisfaction score that crashed before it climbed.

Source and label: IBM's figures below are its own, from its AskHR case study and IBM Think writing about its internal deployment. IBM is both operator and vendor here, so treat the numbers as IBM's reported account at IBM's scale, not as a forecast for a twenty-person firm. The other cases are labelled the same way.

What an HR chatbot does well, and what it should not touch

Good at:

• Policy questions that already have a written answer: holiday and PTO rules, sick leave, expenses, benefits
• Balances and status looked up from the HR system, where it is connected
• Documents on request, such as payslips and employment verification letters
• Onboarding checklists and first-week questions: who to ask, where things live, which forms are due

Hands to a person: pay disputes, performance problems, complaints, misconduct, health and family matters, anything with legal weight, and any question where the written answer is missing. Every case below draws that same line.

IBM AskHR: 94% contained, after a minus-35 start

IBM says AskHR automates more than 80 HR tasks and works in a two-tier model: AI handles routine inquiries, and HR advisers handle complex and sensitive ones. It connects to Workday, SAP and Concur, so it can complete transactions such as employee letters, vacation requests and manager transfers, not just answer questions.

IBM reports that in 2024 AskHR handled more than 11.5 million interactions with a 94% containment rate, meaning only about 6% were routed to an HR specialist. It reports a 40% reduction in HR operating costs over four years and a 75% reduction in support tickets since 2016. In 2025, IBM says, AskHR handled more than 16 million user messages. Its HR net promoter score is reported as climbing from minus 35 to plus 74. That first number is the most instructive one in the pile.

The cause, in IBM's telling: in 2018 it shut the HR phone line and email address for all employees, overnight, and AskHR became the only front door. The score fell to minus 35 as people hit a bot where a person used to be. It recovered as the assistant learned more tasks, completed transactions instead of reciting policy, and staff were taught new habits. The licence was the smallest part. Behaviour was the project.

Three more HR chatbots with hard numbers

iFood: 83% of HR requests resolved at the first level. The Brazilian delivery company, with about 8,000 employees, built its own HR assistant, Alli, which runs in Slack and WhatsApp. It receives about 29,000 requests a month and handles onboarding, vacation, record updates and leave. With more than 1,000 active content items its rate of correct answers rose from 32% to 90%. Today 83% of requests are resolved at the first level, with an average rating of 4.8 out of 5. The HR service team went from 11 people to two, without layoffs; the others moved to more strategic HR work. (iFood Benefícios, December 2025, published on iFood's own site; company-reported figures.) The accuracy gain came from content work, not a new model.

Coca-Cola Andina: an HR chatbot on WhatsApp for plant workers. Most of the bottler's Argentine staff work in plants and use phones, not laptops, so its HR assistant, Andi, runs on WhatsApp. Answers change with location, contract type and tenure, because vacation days and benefits do. When Andi can't resolve a request, it opens a ticket for an HR manager with the question and the employee's profile attached. 1,200 of about 3,600 active employees in Argentina use it each month; there have been nearly 5,000 conversations, and it scores 4.1 out of 5 on 976 employee surveys. Weeks before launch it was joking about sick leave until the team fixed a tone setting; one HR owner per region now controls the content. "Content structure mattered more than we expected," says Tomas Colla of its IT team. (Microsoft case study, September 2026; vendor-published, as Andi is built on Microsoft Copilot Studio.)

Microsoft: 31% fewer tickets than its old HR tool. Microsoft reports that, compared with its previous HR tool, people using its Employee Self-Service Agent were 25% more likely to get a correct response and 31% less likely to create a support ticket (Microsoft, April 2025). Its goal of 44% fewer monthly HR tickets by mid-2026 is a target, not a reported result (Microsoft Inside Track, November 2025). Vendor claim: Microsoft sells the agent.

None of these firms is small, and every figure is self-reported or vendor-published. The pattern still holds: written answers first, one content owner, a clear handoff to a person, and a first version that needed fixing.

An HR chatbot for a 20–50 person firm

At 20 to 50 people there is usually no HR department, just an owner, an office manager and a handbook. You don't need IBM's integrations. An assistant built on your handbook, leave rules and onboarding checklist covers most of what people ask.

1. Handbook Q&A. "What's the policy on…" answered from one current handbook, quoting the section it used. If it can't find a section, it says so.
2. PTO and leave. The rules, how to request time off and who approves it. Show balances only if your payroll or HR tool provides them safely; otherwise, point to where people can check.
3. Onboarding. The first-week checklist, who to ask for what, where files and logins live, and which forms are due by when.

What to map first. For two weeks, log every HR question that reaches the owner or office manager: who asked, what they asked, how it was answered and where the answer lives. A handful of questions usually make up most of the list. Map that one process, from question to source document to answer to handoff, before you pick a tool. That is what a process map is for, and the free first step is a one-page sketch of it. Once the questions are mapped, routing a leave request to its approver is a plain workflow automation job, not a chatbot one.

When a person must stay in the loop. Pay, performance, complaints, harassment or discrimination, health and family leave decisions, terminations, and anything that commits the firm legally. Also any answer the bot can't tie to a written source. Make it say so plainly and hand off with the conversation attached, as Andi does. Write that rule down where staff can see it; our AI policy template has a section for it. Limit what personal data the assistant can see to what the three jobs above need.

AI agents for small business covers which assistants we would set up first. The same split shows up on the customer side, where AI customer service agents finish about half the questions and the return depends on how the other half is handled.

How to run the minus-35 phase without IBM's budget

Don't copy IBM's scale or its 40% saving; copy the sequence, and expect the first score to be bad. You don't need a change programme. You need one owner, one hour a week, and complaints treated as training material. In the first month, read twenty real exchanges a week and sort every failure: answer missing, answer wrong, or a question that should never have stopped at the bot. Fix the document, not the person who asked, and tell the team what changed. That loop is what turned iFood's 32% into 90% and IBM's minus 35 into plus 74.

Decide in advance where the saved time goes, or it disappears into meetings and inbox. MIT's research found that most organisations saw no measurable P&L impact within about six months; the MIT article covers why. Our readiness assessment builds the directed use into the plan.

Ask Ada which three questions your team answers every single week. That is where your HR chatbot starts, and she will build it into your free report at /audit/. Licences alone will get you IBM's minus 35 without IBM's training budget.

Frequently asked questions

What is an HR chatbot?

An HR chatbot is a chat assistant that answers employees' routine HR questions, such as leave rules, payslips, policies and onboarding steps, from the firm's own written policies and HR systems. It hands sensitive or unclear matters to a person.

What are the best HR chatbot use cases?

Handbook and policy questions, PTO and leave rules, onboarding checklists, document requests such as payslips and verification letters, and routing people to the right form.

Do HR chatbots actually work?

For routine questions, the reported numbers are strong: IBM reports 94% of AskHR interactions contained in 2024, and iFood reports 83% of requests resolved at the first level. Both are company-reported and took years of content work; IBM's first satisfaction score was minus 35.

Does a small business need an HR chatbot?

Only if the same HR questions keep reaching the owner or office manager. Log them for two weeks. If a handful make up most of the list and the answers are written down, an assistant built on the handbook, leave rules and onboarding checklist can take them.

Will an HR chatbot replace HR staff?

It replaces the repeat questions, not the judgment. iFood's HR service team went from 11 people to two without layoffs; the others moved to other HR work. Pay, performance, complaints and anything with legal weight should stay with a person.

Sources: IBM, AskHR case study, IBM Transformation: HR and IBM Think, Embracing the future of HR by becoming an AI-first enterprise (IBM-reported figures; IBM is operator and vendor); Fortune coverage of IBM HR leadership comments, July 2024; iFood Benefícios / Think Work, 5 December 2025 (company-reported); Microsoft, Coca-Cola Andina case study, 8 September 2026 (vendor-published); Microsoft Tech Community, 17 April 2025 and Microsoft Inside Track, 13 November 2025 (vendor's own deployment). Updated October 24, 2026 with a plain definition of an HR chatbot, three more cases with hard numbers, a section for 20–50 person firms and an FAQ.