AI tools for HR teams: where software stops and people decide
For talent-acquisition, HR-operations and people leaders who already run an ATS, an HR service desk or a review cycle, and are being sold AI inside it.
No automation scores. Each product is read step by step: what the software does itself, what it only proposes, what a person must still do or decide — and what nobody has shown yet.
What this page concludes
- A rank is not a rejection.Every recruiting row scores or interviews. Only a rule at Route removes anyone, and two rows document one.
- An answer is not a write.A knowledge layer answers. A service agent can file the request and write the change to your HRIS, with someone’s permissions.
- A draft is not a rating.Drafts, summaries and coaching are software. Every rating, pay or promotion decision documented here is a person’s.
- Not verified is not no.Where the vendor’s own documentation did not establish a step, the page says so instead of guessing.
- Outside its role is not a gap.Analytics that never touches a review is not missing a feature; the decision lives elsewhere.
How to read the boundary
Two lanes and one line. Software above it, people below it.
The line is the same on every row. What moves is where each step lands, and the steps with a red dot are the ones that change something for a person.
- Software acts: Software acts. Ranks applicants.
- Software assists: Software assists. Manager decides; software records it.
- Person decides: Person decides. Team makes the offer.
- Not verified: Not verified. No first-party source either way.
- Outside its role: Outside its role. Happens in another system.
- Consequential step — it changes a candidacy, a record, a request or a rating
- Conditional — software reaches that state only under a setting, rule, plan or deployment you control, named under the row
- Software acts
- Software produces the step’s outcome and it takes effect without a person — sometimes only inside a setting, rule or plan the row names.
- Software assists
- Software proposes, prepares or carries out the mechanics; a person accepts, edits, sends or decides.
- Person decides
- Documented as a person’s step, with no software route through it found.
- Not verified
- The vendor’s own sources did not establish the step either way. Not a mark against the product.
- Outside its role
- The step happens in another system, so there is no boundary to draw here.
- Rank
≠rejectionA score is read at Screen. Leaving the pipeline is only ever read at Route.
- Answer
≠system writeTelling an employee the policy is Answer. Changing their record is Act / Write.
- Draft review
≠ratingDrafts and coaching are Draft / Coach. The rating is only ever read at Decide.
When a product has more than one route through a step
When a product has more than one route through a step, the cell shows the furthest the software is documented to go without a person: acts over assists over person. The row names the setting, rule, plan or deployment that route depends on, and the manual route remains. Marketing adjectives such as “autonomous” never set a state.
01 Recruiting & screening
How far should recruiting software go between scoring a candidate and rejecting one?
Every product here says it screens. The difference that matters is whether software can move a candidate out of the pipeline, or only puts them in an order a recruiter works through.
Your ATS is enough when
Inbound volume is manageable and your ATS already scores applicants and sends scheduling links.
Add a specialist when
Outbound sourcing is the bottleneck, first-round screens eat recruiter days, or one talent profile has to serve many requisitions.
1Source
Add a person to the pipeline
2Screen
Assess: score, rank or interview
3Engage
Contact and schedule
4Route
Move the candidate on, or out
≠A rank is not a rejection
5Hire
The offer decision
| Product | Human Decision Boundary | When it fits | Price |
|---|---|---|---|
 Workable↗Applicant tracking with an AI agent that sources, scores and messages |
At Route, software advances or disqualifies by your rule, once switched on; otherwise a recruiter moves candidatesA person decides at Hire
| Choose whenYou want sourcing, scoring, messaging and the pipeline in one ATS. Skip ifYou are keeping an ATS you already run. | From $299/mo (Standard, 1–20 employees) Company-size band, plus AI credits per agent action |
Why this reading · 6 first-party sourcesBoundaryThe Workable Agent sources and scores candidates, and can move them to a stage past a score threshold or disqualify those who miss criteria you mark. Separately, yes/no knockout questions on the application form can disqualify automatically on any plan. Any manual activity on a candidate stops further agent moves, and a disqualified candidate can be reverted. Workable’s help center says the customer makes every hiring decision, and offers can wait on the hiring team’s approval before they go out.
Reviewed Sep 18, 2026 | |||
 Gem↗Sourcing CRM that runs on your ATS or ships its own |
At Route, software rejects on a knockout answer, under a rule you set; otherwise a recruiter rejectsA person decides at Hire
| Choose whenOutbound prospects and applicants should share one record. Skip ifInbound applicants are the whole pipeline. | Quoted Employee count and configuration |
Why this reading · 6 first-party sourcesBoundaryGem’s AI app review scores and ranks applications, and Gem says it makes no automated decision to advance or reject from that rank. Rejection comes from somewhere else: knockout questions in Gem ATS, where an admin sets the answer that rejects and when the rejection email goes out. That rule is read at Route. AI Sourcing surfaces prospects, and they enter a project only when a recruiter marks them or adds them. Offers go to named approvers, who approve, reject or request changes.
Reviewed Sep 18, 2026 | |||
 hireEZ↗Agents that screen, message and schedule on top of your ATS |
A person decides at RouteA person decides at Hire; software carries it out, where deployed
| Choose whenYour ATS stays, and sourcing and first screens are the bottleneck. Skip ifYou want rules, not recruiters, to move candidates. | Quoted for teams Per seat for a solo recruiter; the replaced tool stack for enterprise |
Why this reading · 7 first-party sourcesBoundaryhireEZ’s AI phone screen runs the interview and produces a transcript, highlights and summarized signals, and its agent stack-ranks applicants against your criteria. Its agent pages say recruiters decide who moves forward and that nothing is rejected without them, and applicant review leaves Advance and Reject to the recruiter. One platform page says the agent updates statuses without saying which; that line does not outweigh the specific statements. Who adds candidates to a project in the current agent platform was not documented, so Source stays Not verified.
Reviewed Sep 18, 2026 | |||
 Eightfold AI↗Skills matching, an AI interviewer and a candidate agent |
At Route, software hands applicants into interviews, where deployed; otherwise a recruiter moves candidatesA person decides at Hire
| Choose whenOne skills profile has to match talent across many roles and internal moves. Skip ifYou cannot run AI interviews or scores through a candidate-disclosure process. | Quoted Not published |
Why this reading · 6 first-party sourcesBoundaryEightfold produces match scores and runs AI interviews, and says the interviewer supports human decisions rather than making hiring decisions. Candidate Agent hands applicants directly into the interview process. No automatic rejection or knockout route was documented, so Route here is forward movement only. Offer approvals appear only on a design-partner page for early access, so Hire stays with recruiters and hiring managers.
Reviewed Sep 18, 2026 | |||
Hard lineA score is evidence about a candidate, never a decision about them. If any setting lets software disqualify, get in writing who can switch it on, and run the knockout test before launch.
How this decision is read · Reviewed Sep 18, 2026
Rejection, disqualification and advancement are read only at Route, whichever screen the vendor puts them on. Scores, ranks and AI interviews are read at Screen. A “Hired” status is a record update; the offer decision is read at Hire.
02 Employee service & HR operations
Should an HR assistant only answer employees, or also change their records?
An assistant that explains the leave policy and one that books the leave are different purchases. The second one acts with someone’s permissions inside your system of record.
A knowledge base is enough when
The questions are answerable from policy pages, and every change still goes through an HR ticket.
Add a service agent when
Routine changes such as time off and contact details should finish without an HR person keying them in.
1Retrieve
Find the governing policy or record
2Answer
Reply to the employee
3Request
Open a case or submit a request
4Act / Write
Change a record in the HR system
≠An answer is not a write
5Escalate / Approve
Who grants, denies or takes over
| Product | Human Decision Boundary | When it fits | Price |
|---|---|---|---|
 Guru↗Verified company knowledge with cited AI answers |
Act / Write happens in another systemEscalate / Approve not verified | Choose whenAnswers are the problem and HR records are not. Skip ifEmployees need requests completed, not explained. | Quoted Organization scale and knowledge complexity, not seats |
Why this reading · 5 first-party sourcesBoundaryGuru’s knowledge agents search verified cards and connected sources and deliver cited answers, including proactively. Its HRIS sync runs one way into Guru and its MCP server edits Guru content, so changing an HR record is outside its role. Assign Expert routes a flagged answer to a person inside Guru. That escalates a question; it is not a request object or an approval, so Request and Escalate / Approve stay Not verified.
Reviewed Sep 18, 2026 | |||
 Leena AIAn agent that completes employee requests inside your HRIS |
At Act / Write, software updates HRIS records via connector, where deployedAt Escalate / Approve, software can approve or reject on timeout, under a rule you set; otherwise the approver decides
| Choose whenRoutine requests should finish in your HRIS without an HR ticket. Skip ifYou are not ready to give an agent write access to HR records. | Quoted Employee headcount, plus usage for self-running agents |
Why this reading · 6 first-party sourcesBoundaryLeena answers from connected policy sources and cites the source. Through the Workday connector it submits a time-off request as a pending business-process event, and its connectors also expose record writes: Workday contact details, SuccessFactors record changes and Oracle Fusion worker actions. An approval node pauses a workflow for an assigned person to approve or reject. With Auto-Action set, the system approves or rejects on its own when the approver does not respond in time — which is what the silent-approver test is for.
Reviewed Sep 18, 2026 | |||
 ServiceNow HR Service DeliveryHR cases, requests and the employee portal, with ServiceNow Otto as its AI layer |
At Act / Write, software updates personal details in HCM, in Autonomous mode; in Supervised mode it pauses for a person at each stepA person decides at Escalate / Approve; software carries it out
Assistant layerServiceNow’s documentation names the AI layer for this product ServiceNow Otto, which ServiceNow says replaces Now Assist and Moveworks across its products. Moveworks, now owned by ServiceNow, is read as lineage for how the assistant reaches other systems — evidence for this row, not a separate recommendation. | Choose whenHR cases already live in ServiceNow, or IT and HR share one service desk. Skip ifYou do not run ServiceNow and do not want a platform program. | Quoted Tier, plus Assists consumed by AI features |
Why this reading · 10 first-party sourcesBoundaryIn ServiceNow’s HR Service Delivery documentation, AI agents search HR knowledge articles and catalog items, answer inquiries classified as noncritical, create cases and requests, and update personal details in an HCM system. Each ships inactive and runs in an execution mode the buyer sets. The marketing claim that agents handle approvals autonomously resolves, in the documentation, to agents routing an approval to a manager and surfacing approvals from Workday and SAP Concur. The manager decides. Approval-chain permissions and the audit trail for agent actions were not established in the documentation reviewed.
Reviewed Sep 18, 2026 | |||
WatchAn agent that writes to your HRIS acts with someone’s permissions. Name whose, and which fields, before connecting it — then run the change-my-record and silent-approver tests.
How this decision is read · Reviewed Sep 18, 2026
Answering is read at Answer and never counts as a write. Opening a case or submitting a request is Request. Only a change to a record in the system of record, or the fulfillment itself, is Act / Write. Escalate / Approve reads who holds approval authority.
03 Performance & people insight
Can AI help managers write reviews and see risk without deciding the rating?
Drafting, summarizing and coaching are now standard. What separates these products is whether anything they produce reaches a rating, a pay change or a promotion before a person does.
Your review tool is enough when
Managers write reviews on a set cycle and nobody is asking for attrition or pay-equity analysis.
Add insight when
Leaders need workforce answers across HR data, or managers need coaching between cycles.
1Collect
Feedback, check-ins, surveys or HR data
2Summarize
Condense what was collected
3Draft / Coach
Draft review text or coach the manager
4Recommend
Suggest an action, a flag or a score
5Decide
Rating, pay, promotion, exit
≠A draft is not a rating
| Product | Human Decision Boundary | When it fits | Price |
|---|---|---|---|
 LatticeReviews, goals and feedback, with cited AI review drafts |
A person decides at Decide
| Choose whenManagers write reviews in one place and want drafts that cite the evidence. Skip ifWorkforce analytics is the question, not reviews. | $13/seat/mo, billed annually Seats per module and add-on, on an annual contract |
Why this reading · 10 first-party sourcesBoundaryLattice AI summarizes feedback and review packets and drafts open-text review answers with citations. Lattice says the agent cannot select ratings or submit a review; the reviewer stays accountable for the final submission. Each draft’s confidence score describes how much evidence it had, not the employee. The AI recommends growth areas that a person adds and publishes. Only admins, facilitators or a calibration role can input calibrated scores, and final approval of compensation changes sits with admins and compensation admins.
Reviewed Sep 18, 2026 | |||
 15FiveReviews, check-ins and surveys, with an AI coach for managers |
A person decides at Decide
| Choose whenManager coaching is the gap, and you want AI kept out of ratings in writing. Skip ifYou need people analytics beyond your own review cycles. | $11/user/mo (Perform), billed annually Users, plus per-manager and per-employee AI add-ons |
Why this reading · 8 first-party sourcesBoundary15Five says its AI-assisted reviews never score ratings and never block submission: AI drafts answers, and managers write summaries and finalize results. Kona turns meeting transcripts into notes and coaching prompts delivered to the manager, which is why Draft / Coach reads Software acts while a review draft still needs the reviewer. Recommend covers compensation-fairness and pay-risk insights, engagement-driver impact and Kona’s manager competency scores. None of them is a performance rating.
Reviewed Sep 18, 2026 | |||
 Culture AmpSurveys and reviews, with an AI coach drawn from both |
A person decides at Decide
| Choose whenEngagement surveys drive your people agenda and reviews should sit beside them. Skip ifReviews alone are the need and surveys are not. | Quoted Employees, products and service tier, billed annually |
Why this reading · 8 first-party sourcesBoundaryCulture Amp runs review cycles and collects feedback, summarizes comments once there are enough of them, and its AI Coach drafts and improves feedback that a person finishes and submits. Retention insights flag turnover risk and AI Coach proposes action plans; Culture Amp documents that AI Coach does not suggest ratings. Final ratings are set by the manager, HR and managers can review and adjust them in calibration before they are finalized, and compensation reviews happen in Pave.
Reviewed Sep 18, 2026 | |||
 VisierWorkforce analytics with an AI agent, Vee |
Decide happens in another system
| Choose whenLeaders ask workforce questions that reviews cannot answer. Skip ifYou need to run reviews or collect feedback. | Quoted Not published |
Why this reading · 5 first-party sourcesBoundaryVisier pulls data through connectors, and Vee answers workforce questions from it within the asker’s permissions. Vee Boards surface trends, resignation risk and suggested actions. Visier describes its insights as helping executives and HR make decisions, and nothing in the product records a rating, pay or promotion decision, so Decide sits outside its role. No review-drafting or coaching feature was established either way.
Reviewed Sep 18, 2026 | |||
Hard lineNothing here should put a rating, pay change or promotion in front of a manager that no person entered. If a demo shows one, that is the finding.
How this decision is read · Reviewed Sep 18, 2026
Drafts, summaries and coaching are read at Summarize and Draft / Coach. Recommend says exactly what is suggested: an action, a flag, or a score about something other than a person’s performance. Ratings, pay and promotions are read only at Decide.
Before you sign
Nine things to get in writing, whichever decision you are making
Procurement questions, not legal advice. Where a hiring location has an AI or automated-decision law, name it with counsel and attach it to the third group.
Who or what can act3
- Which settings let software disqualify, advance, close or approve?
The list, the default of each, and the role that can change it.
- What can the agent write, and as whom?
Field-level HRIS and ATS writes, and whether it uses a service account or the employee’s own permissions.
- Where does a proposal stop?
Scores, drafts and recommendations marked as such, and never pre-filled into a decision field.
Where candidate and employee data goes3
- Is our data used to train shared models?
A contractual opt-out that covers subprocessors and model providers.
- How long are interviews, transcripts and comments kept?
Retention you set, and deletion you can prove.
- Who can read AI summaries of feedback?
A summary can expose what a rater wrote in confidence; its permissions must match the source.
What you can prove afterwards3
- Why was a candidate scored, not just what the score was?
Plain-language reasons a recruiter could repeat to the candidate.
- Is every automated action and override logged?
An exportable log with the actor, the rule and the time.
- What bias or impact data does the vendor supply?
The vendor’s part in any audit a hiring location requires, in writing.
Name the laws only after legal review.
Test before you buy
Six trials, each aimed at a step on the boundary
Run them in a sandbox with the vendor on the call. Each ends in a question with a yes or no answer.
Trial 01Recruiting
ScreenRoute
Knockout test
Apply as a candidate who misses one required criterion.
Pass or failDid a person see this candidate before they left the pipeline?
Trial 02Recruiting
Screen
Reason test
Ask for the reasons behind the top and bottom score in one requisition.
Pass or failCould a recruiter say those reasons to the candidate?
Trial 03Employee service
AnswerAct / Write
Change-my-record test
As an employee, ask the assistant to change your home address.
Pass or failDid it explain, file a request, or write the change — and did the vendor say which beforehand?
Trial 04Employee service
RetrieveAnswer
Stale-policy test
Replace a policy page, then ask yesterday’s question.
Pass or failDoes the answer cite the new version, or refuse?
Trial 05Employee service
Escalate / Approve
Silent-approver test
Submit a request that needs a manager, then have the manager do nothing.
Pass or failDoes it wait, escalate to a person, or approve on a timer?
Trial 06Performance
Draft / CoachDecide
Draft-not-rating test
Generate a review draft for a low performer.
Pass or failDoes any AI output contain a rating, score or pay suggestion no manager entered?
Switch it off and take your data
In the sandbox, turn every AI feature off, then export candidates, cases and reviews with their AI outputs attached. If either fails, the contract is the only exit.
Deliberately not compared
Purchases outside these three decisions
- Payroll and benefits administrationMoney and entitlements are a different control problem.
- Background checksConsumer reports with their own consent and dispute rules.
- Interview recording and notesConsent and retention decide it more than AI quality.
- Choosing an HRISThe system of record itself; this page assumes you have one.
- Employee monitoringProductivity surveillance is not compared here, on purpose.
Still open
Is AI screening legal where we hire?
It depends on the location and on how the tool is used. This page is not legal advice; the checks under Before you sign list what to get in writing, and counsel should name the laws that apply.
What does Conditional mean on a step?
The software does that step only after you switch something on: an action, a rule, an add-on, a tier or an execution mode. The condition is named under the row, and the manual route still exists.
Why do some steps say Not verified?
Because the vendor’s own documentation did not establish that step on the review date. It is a fact about the evidence, not a verdict on the product.
Why is Moveworks not compared on its own?
ServiceNow owns Moveworks, and says ServiceNow Otto replaces Now Assist and Moveworks across its products; its documentation names the AI layer for HR Service Delivery ServiceNow Otto. The page reads Moveworks as lineage for how that assistant reaches other systems, inside the ServiceNow row.
How this page decides
This is a buying guide, not a directory listing. Products are here because an editor put them here for the job, and they stay whether or not ToolWorthy happens to hold a record for them.
Selection and evidence
- Every state is read from the vendor’s own documentation, help center, product or pricing pages on the review date. Third-party sources never set a state.
- The furthest documented route through a step sets its state; the condition that route depends on is named on the row, and the manual route remains.
- Not verified means the vendor’s own sources did not establish the step either way. Outside its role means they establish that the step happens in another system.
- Prices are the vendor’s published figures. Where a vendor publishes none for the plans that matter, the row says Quoted.
Last reviewed Sep 18, 2026. Pricing and features change; the linked tool and category pages are the maintained record.
Keep going
Category directories are the maintained inventory for a capability, and the ranked guides are where products are scored against each other. This page only reads where software stops and a person decides.