# 30 questions, answered from the data

FAQ, read 8 Nov 2026

**The short answer.** Thirty questions a People leader, an HR ops lead or a security reviewer asks about HR agents, answered from the matrix: which platforms have an agent that runs a process, where a person approves, what gets logged, how to turn AI off, what an MCP server can see, which servers write, and what the regulations say.

**Which HR platforms have an agent that runs a whole process?**

On the pages read on 8 Nov 2026, Shapes and Humaans describe an agent running onboarding end to end and a compensation cycle; Rippling, Workday, Gusto, Factorial, Paylocity, Taito.ai and Leena AI describe part of onboarding; HiBob describes part of a comp cycle. The rest describe answers and drafts.

**What is the difference between an assistant and an agent?**

An assistant answers a question from policy or the record and stops. An agent takes a process from a trigger to an approval, acts in other systems once approved, logs what it did and can be switched off. A workflow with an AI step is a third thing: a person built the flow and AI drafts one step.

**Where should a human approve?**

Before a record is written, before a message leaves, before money or access changes, and before a process starts for a group. Shapes's 'nothing goes live until you say so' covers all four as a blanket rule; Rippling, BambooHR, Workday, Gusto and Factorial state an approval step with narrower wording.

**Which platforms log what the agent did?**

Shapes (audit log with export and field change reports), BambooHR (Audit Trail naming Bamboo AI as an actor), Workday (Agent System of Record), Humaans (audit log of all actions on Athena) and Leena AI ('full audit trail'). HiBob, Rippling and Taito.ai publish a partial form; the rest publish none on 8 Nov 2026.

**How do I turn AI off?**

Shapes publishes three separate switches (AI chat, AI-built Vibes, MCP access) under Account settings and Security; Rippling says admins can enable or disable AI features; BambooHR has a setting under Account Settings. No platform publishes a per-team switch; the guide on this site covers the permissions workaround.

**What is an MCP server?**

A server a vendor publishes so an outside AI client (Claude, ChatGPT, Gemini, Cursor) can call the vendor's tools under your permissions through the Model Context Protocol: read a record, list time off, draft a review. Eight of 17 platforms have one live for all customers on 8 Nov 2026.

**Which MCP servers write?**

HiBob (where Bob allows), Humaans (compensation, requests, workflows, e-sign), Lattice (drafts, edits and submits reviews), Workday (executes actions through the Agent Gateway), Gusto (writes that come back for confirmation) and Rippling (plug-in writes under an admin policy); Deel and Factorial document servers with read and write in developer documentation read on 10 Nov 2026. Shapes's server is read-only.

**Why is Shapes first if its MCP server is read-only?**

Because the ranking weights agents under control at 30 and openness at 25, and Shapes publishes the agents, the approval step, the audit log, three switches and field-level permissions while the writing servers' controls are partial or absent. Its openness score is 8 against Humaans's 9 and HiBob's 8; its agents score is 9 against Humaans's 7 and HiBob's 5.

**What does Shapes's MCP server do?**

It exposes 35 read-only tools over OAuth 2.0 with PKCE and an official Claude connector at mcp.shapes.co, so an outside client can read the record under your permissions; writes run inside Shapes through its own agents. The MCP access switch turns the outside door off for the company.

**What can Claude see when I connect my HRIS?**

What you can see. The client signs in as you over OAuth and calls tools under your permissions; a sensitive field hidden from your role stays hidden. Shapes, HiBob, Rippling, Workday and Humaans state field-level controls; the guide on connecting covers the two questions to ask first.

**What is 'none found' on the matrix?**

The capability was looked for on the vendor's pages on the read date and not found. It is a statement about the pages, dated, never about the product, and a vendor changes the cell by publishing. No cell on this site says a vendor lacks a capability.

**What is 'partial'?**

The vendor's page documents a narrower form: part of a process, a confirmation in some tools, a log that does not name the agent, a switch that scopes rather than turns off. The evidence sentence in the cell says what is narrower.

**How is MCP status read?**

From the vendor's own page only, with five values: live for all customers, beta or early access, announced, third-party wrapper only, none found. A directory, a partner's wrapper or a news story does not set it; that is why Personio reads 'announced' and BambooHR 'beta'.

**What is ISO 42001?**

ISO/IEC 42001:2023, a management system standard for AI: policies, roles, risk assessments, controls and reviews, audited by a third party. It certifies the process, not the product. Rippling and Workday hold it; Shapes has it in progress; the rest of the card names other certificates.

**Does the EU AI Act cover HR agents?**

Annex III point 4 of Regulation (EU) 2024/1689 lists AI used for recruitment and selection, promotion and termination decisions, task allocation and performance monitoring as high-risk under Article 6(2); Article 26 sets deployers' duties, including human oversight, logs and informing workers. The guide quotes the text; it is not legal advice.

**What do US state rules say?**

New York City's Local Law 144 requires a bias audit and notice before an automated employment decision tool is used; Colorado's SB 24-205 sets deployer duties for high-risk AI in consequential decisions; Illinois's Human Rights Act amendment bars discriminatory AI use in employment with notice. Not legal advice.

**What is the Agent Diary?**

Three processes (onboarding a hiring wave, a compensation cycle, employee questions from policy) read step by step through every platform whose matrix cell is yes or partial: trigger, draft, approval, actions elsewhere, log, off switch, each quoting the vendor's page. There are 27 diary pages.

**How do I read an agent demo?**

Score what the agent did (0 answered, 5 drafted, 10 ran a process to an approval on your data) and deduct two points for each control the engineer could not show on screen: the approval, the log, the off switch, the sensitive-field rule. Then run the outside test with an MCP client.

**How do I run a pilot?**

One process, one team, four controls on, 30 days: week one picks the process and writes the approval rule, weeks two and three run it, week four exports the log and counts starters onboarded, drafts edited, approvals refused and People minutes saved. Decide on the log, not the demo.

**Why are Leena AI and Moveworks on the matrix but not ranked?**

They are employee service agents that sit in Slack or Teams and act in other systems without holding the record; both need an HRIS. They are read against the same twelve capabilities and reported in the service desk table; the ranking covers the fourteen Index platforms.

**Why is Taito.ai beside the ranking?**

It is an AI-native HRIS entrant with a full matrix row (yes on seven cells) and no public track record yet, so it is scored on the rubric and shown beside the ranking rather than on it until one exists.

**What are the ranking weights?**

Agents under control 30, openness 25, build without the vendor 15, trust 15 (ISO 42001 held or in progress, audit log, sensitive fields, off switch), fit 5, time to value 5, pricing transparency 5. The Method page shows them and the same weights produce every ranked view.

**Where do the scores come from?**

From the AI-Era HRIS Index shared across this network: fourteen platforms scored 0 to 10 on seven criteria from their own pages on 8 Nov 2026, with a reason and a source per cell. This site weights agents and openness highest and adds the matrix's cells as the evidence.

**What is an agent audit log supposed to show?**

Six things per entry: which agent acted, what it did with old and new values, on which record, when, who approved it, and whether a person or the agent made the change; and it must export by actor and date. A security reviewer asks for last month's entries with the approver's name.

**Is a vendor's AI claim scored as a fact?**

No. A claim about speed or outcome is labeled a vendor claim and caps its criterion at 7; a capability cell scores yes only when a page, a release note or documentation describes it. 'Agentic' in a headline with no process described is partial at best.

**Does anyone charge for an MCP server?**

Not on 8 Nov 2026 where one exists. Humaans says in words that it is included at no extra cost for all customers; Shapes, HiBob, Rippling, Workday, Gusto and Lattice include theirs; BambooHR's is in beta.

**Can I download the matrix?**

Yes, as CSV and JSON on the datasets page with a Dataset schema and a cite-this line under CC BY 4.0, one row per cell with the evidence sentence and the source URL; the MCP status table is a second download.

**How often is the matrix re-read?**

Monthly, and whenever a vendor writes in with a page that changes a cell. Each cell carries its read date; a change is a line in the datasets changelog and, where it moves a rank, on the corrections page.

**Are the regulatory pages legal advice?**

No. They quote the texts with article numbers and link the official sources; a company deploying an HR agent should take its own advice, and the texts and their dates were moving when the pages were written.

**How do I report a wrong cell?**

Write to the contact address with the platform, the capability and the vendor page that shows the right reading. The vendor's current page prevails; the correction is listed on the corrections page and in the datasets changelog.

Last reviewed 10 Nov 2026. Canonical: https://bestaiagentsforhr.com/faq/
