Every AI use case, mapped to the business.

32 use cases owners are being sold this year, sorted by where they act in a business of 5 to 100 people. Only 3 are proven in production.

Status
Pick a room to see its use cases
How we rate

Four statuses, one rule: rate the evidence, not the pitch.

Each entry gets a status for how proven it is in businesses of 5 to 100 people, and a separate rating for how strong the evidence behind that status is.

Evidence strength

StrongSeveral independent studies or large independent surveys.
ModerateAt least one independent study or credible survey, plus consistent case evidence.
ThinA handful of independent data points.
Vendor claimsOnly what sellers publish about their own products.

How an entry is made

Each entry is drafted from public sources: academic studies, regulators, independent surveys and reputable reporting. Every figure carries a numbered citation you can open. Anything a vendor says about its own product is labelled as a vendor claim.

Review and updates

Every source in this edition was opened and checked against the text in October 2026. Each entry is then signed off by a named operator and re-checked when the evidence moves. Until that sign-off, the entry is marked as a draft at the top of the page.

The full guide

All use cases at a glance

Field GuidePeople & payrollHR policy Q&A assistants
In pilotsPeople & payroll

HR policy Q&A assistants

Also called HR chatbot, employee handbook chatbot, HR helpdesk bot, policy assistant

EvidenceThin
ClassGenerative

How AI chatbots answer staff questions on leave and HR policy, what the evidence shows, what they cost and why wrong answers can bind the employer.

ReviewDraft, awaiting operator review
Drafted27 October 2026
Fact-checkedEvery source opened, October 2026
Key takeaway

An HR Q&A assistant answers staff questions from your own handbook, and is only as reliable as that handbook. In the one academic evaluation we found, of SAP's internal HR chatbot, an expert rated the answers well⁠. But the search step put the right policy article first only 22% of the time, largely because of near-duplicate versions⁠. Grounding answers in documents does not stop errors: legal research tools built this way still gave wrong or unsupported answers more than 17% of the time⁠. Treat answers as guidance.

At a glance

StatusIn pilots
What it doesAnswers employees' questions about leave, benefits, expenses and conduct rules by searching your policy documents and writing a reply, ideally with a link to the source paragraph.
Typical costUS$21 per user per month, billed annually, for an office-suite AI add-on, cut to US$18 for the first year under a promotion running to 31 December 2026⁠. A team chatbot plan is US$25 per user per month, billed annually⁠ (vendor list prices, checked Oct 2026). Dedicated HR helpdesk tools are mostly quote-based.
Data it needsOne current version of each policy, and optionally leave balances and payroll records if it is to answer personal questions.
Setup effortLow for a policy-only assistant; Medium once it touches individual employee records, because access controls then matter.
EvidenceThin. One independent academic evaluation at a large company, analogous studies of workplace AI assistants, and documented public failures; no small-firm studies.

What is an HR policy Q&A assistant?

It is a chatbot that answers the questions an owner or HR person hears every week: how many leave days do I have, can I carry them over, what is the overtime rule, how do I claim travel costs. Staff ask in plain language and get an answer drawn from the company's own documents.

There are two levels. A policy-only assistant reads the handbook, contracts and procedures and answers general questions. A connected assistant also reads individual records, such as leave balances or payslips. The second is far more useful and far riskier, because it handles personal information and must know exactly who is asking.

How does it work?

Almost all of these tools use retrieval. The software searches your documents for the most relevant passages, and a large language model writes an answer from them. Good tools show the source so the employee can check it.

The weak point is the search step. If your folder holds three versions of the leave policy, the assistant may quote the old one. In SAP's internal HR chatbot, built on roughly 50,000 policy articles, the best search method put the right article first only 22% of the time, and the researchers blamed near-duplicate versions⁠.

Office-suite assistants answer only from files the user is already allowed to see⁠. That is a sensible design, but it means an over-shared HR folder becomes an over-shared chatbot.

What it needs to work

The assistant is only as good as the documents behind it. It needs one current version of each policy, written for the law where your staff work, and, if it answers personal questions, records it can open only for the person asking.

  1. One dated, current version of each HR policy (leave, overtime, expenses, disciplinary, grievance) in one folder, with old versions archived out of the assistant's reach.
  2. Policies checked against the law where you employ people: the Basic Conditions of Employment Act in South Africa, statutory rights in the UK, wage-and-hour rules in the US.
  3. A named owner per policy who updates the source folder on the day a rule changes.
  4. Folder permissions reviewed so that payroll, disciplinary and medical files are visible only to people entitled to see them.
  5. For personal questions: leave balances and pay data from the payroll system, matched to each employee's login and refreshed each pay run.
  6. A clear hand-off to a person for anything involving a dispute, dismissal, health or money owed.

Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.

What does the evidence show?

We found no independent study of HR policy assistants in small businesses. The closest evidence is from large companies or from similar tools.

At SAP, researchers from the Technical University of Munich evaluated an HR chatbot for employees. A single domain expert rated the best model's answers 4.41 out of 5 for truthfulness and 4.11 for usability⁠. That is promising, but one rater at one very large company is a thin base.

The best-studied cousin is customer support. In a study of more than 5,000 support agents at one company, an AI assistant raised issues resolved per hour by 14% to 15% on average, depending on the version of the paper⁠. Broader evidence is more sobering: a Danish study covering 11 occupations, including HR professionals, found chatbot users saved 2.8% of their work hours on average, with no measurable effect on earnings or hours worked⁠.

AI policy assistant vs asking the owner or HR

Compared onAI policy assistantAsking the owner or HR
Speed of answerInstant, at any hourWhen the person is free
AccuracyGood on simple questions if documents are clean; errors sound confident⁠Depends on the person; applies judgement
ConsistencySame answer every time, including the same wrong oneVaries by who is asked and when
Cost modelPer-user monthly subscription⁠Management time and interruptions
Best forFrequent, rule-based questions with a clear written answerDisputes, exceptions, health and discipline

Status and evidence

StatusIn pilots
Evidence strengthThin

Rated pilots because the one academic evaluation of an HR policy assistant we found comes from a very large company with a single expert rater⁠, and the customer-support evidence is an analogy rather than proof⁠. Evidence is thin: we found no small-firm study, and document-grounded systems still make errors that a business can be held to⁠.

Where does it go wrong?

  • It quotes the old policy. Duplicated and outdated documents were the main cause of wrong search results in the one HR deployment studied⁠.
  • It answers confidently beyond the documents. Legal research tools built on curated databases still gave wrong or unsupported answers more than 17% of the time⁠.
  • It gets employment law wrong. New York City's business chatbot, launched to draw on more than 2,000 official web pages, told employers they could take a cut of workers' tips⁠. The city switched it off in 2026; the mayor said it was costing around half a million dollars⁠.
  • You are held to what it says. A Canadian tribunal found Air Canada liable for its chatbot's wrong answer and rejected the argument that the bot was a separate entity⁠.
  • It shows someone else's data. If payroll or disciplinary files are shared too widely, the assistant can surface them to anyone with access⁠.

Numbers it should move

Response timeHow long a customer waits for a first reply.
Staff cost ratioPayroll as a share of sales.

Ask the demo business

Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.

Fictional data
How fast do we answer customer queries?
2.2 hoursmedian first reply in September

Half of September's 180 customer queries got a first reply within 2.2 hours, but 3 waited more than a working day. December was the slowest month, at 4.7 hours.

OctNovDecJanFebMarAprMayJunJulAugSep2.2 h
How this was worked out

Median hours from a customer query arriving to the first reply, per month, across email and WhatsApp.

Answered in 2 msScanned 1 959 rowsFrom customer queries
Want this on your own numbers, in the AI tools you already use?Book a demo

Where the data usually lives

Google WorkspaceMicrosoft 365SimplePayPaySpaceEmail inboxesWhatsApp Business

Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.

Frequently asked questions

How much does an HR chatbot cost for a small business?

About US$21 to US$25 per user per month for a general-purpose assistant. An office-suite AI add-on that can read your policy files lists at US$21, billed annually, cut to US$18 for the first year under a promotion running to 31 December 2026⁠. A team chatbot plan lists at US$25 per user per month, billed annually⁠. Both were checked in October 2026; dedicated HR helpdesk products mostly quote on request. The hidden cost is cleaning up, and then maintaining, the policies it reads.

Do I need IT support to set up an HR chatbot?

Not for a policy-only assistant: you point an office-suite or team chatbot at a clean folder of policies. You do need someone to check folder permissions first, because these assistants show users whatever files they can already open⁠. Connecting leave balances or payslips is a bigger job that usually needs your payroll provider's integration or outside help.

Am I liable if the HR chatbot gives an employee the wrong answer?

Possibly. In Moffatt v Air Canada (2024), a Canadian tribunal found the airline liable for negligent misrepresentation after its website chatbot misstated a fare policy, and rejected the idea that the bot was a separate entity⁠. Similar reasoning could apply when staff rely on your bot about leave or pay. Label answers as guidance, link to the source policy and send disputes to a person. This is not legal advice.

What do POPIA and UK GDPR require for an HR chatbot?

A policy-only bot handles little personal information. Once it reads leave balances, payslips or disciplinary notes, POPIA requires reasonable technical and organisational safeguards against unauthorised access, and a written contract obliging any supplier that processes the data for you to keep it secure⁠. In the UK, the ICO's employment records guidance covers what to tell workers, how long to keep records and how to keep them secure⁠. This is not legal advice.

Will an HR chatbot save me much time?

Expect modest gains. Customer-support agents with an AI assistant resolved 14% to 15% more issues per hour⁠, but a large Danish study across 11 occupations, including HR, found average time savings of 2.8% of work hours⁠. For a small firm the realistic benefit is fewer interruptions on routine questions, not a smaller HR budget.

More in people & payroll

AI CV screeningIn pilotsEvidence: ModerateTimesheet and payroll anomaly checksIn pilotsEvidence: ModerateAI onboarding assistantsUnprovenEvidence: Thin

Sources

  1. 1
  2. 2
    Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools
    Stanford Law School / Journal of Empirical Legal Studies, 2025Study
  3. 3
    Generative AI at Work
    arXiv (Brynjolfsson, Li and Raymond), 2024Study
  4. 4
    Large Language Models, Small Labor Market Effects
    Becker Friedman Institute, University of Chicago, 2025Study
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
    ChatGPT Pricing
    OpenAI, 2026Vendor
  11. 11
    Protection of Personal Information Act 4 of 2013
    South African Government, 2013Government
  12. 12
    Employment practices and data protection: keeping employment records
    Information Commissioner's Office (UK), 2025Government

This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.

Ask your own business. Get the true number in seconds.

Veld connects your accounting, sales and operations data to the AI tools you already use, and keeps it running, so every answer comes from your numbers.

Book a demo