People Desk: How the AI agents work
The short version
- People Desk agents are AI that drafts and prepares. Named people at your company decide and approve.
- Nothing reaches an employee, candidate or manager without approval by a person authorized for that kind of item, and the approval is tied to the exact text approved. The agents have no tool that can send anything.
- Agents never make, score, rank or recommend hiring, pay, performance, promotion, discipline, termination, layoff selection, severance or individual benefits decisions.
- Legal questions are flagged for your counsel, not answered.
- No training on your data, by us or by our model provider.
- AI can be wrong. Approvers must read every draft.
1. The agents
| Agent | What it does | What it never does | Approver |
|---|---|---|---|
| Integration | Compares the acquired company's policies and benefits with yours, drafts Day 1 communications, prepares role-mapping inputs, keeps a 100-day tracker and weekly status | Decide harmonization, select roles or people, set pay or severance | Integration lead |
| Policy and Handbook | Answers employee policy questions from your handbook with section citations, logs gaps, drafts handbook updates | Answer medical leave, complaint, pay dispute or accommodation questions on the substance (it escalates them); state what a law requires | HR owner |
| Onboarding | Builds dated 90-day checklists with owners, tracks overdue items, drafts check-in messages | Collect identity, tax, bank, medical or immigration documents | HR owner or manager |
| Recruiting Coordinator | Drafts job descriptions, checks that postings include a pay range per your rules, proposes interview slots, drafts invitations and status updates for decisions a person already recorded | Read resumes, screen, score, rank or compare candidates, or explain a decision with reasons of its own | Hiring manager |
Every queued item carries the label: "Drafted by a People Desk AI agent. Reviewed and approved by a named person before delivery." Approvers may keep this line in the delivered message; PortLev recommends keeping it on employee-wide announcements and on any message about pay, benefits or role changes.
2. Models and data
Agents use Anthropic's Claude models through the Claude API (or Amazon Bedrock or Google Cloud Vertex AI if your Order Form says so). Each run sends the model the agent's instructions, the operator's task and the tool results for your company only. Anthropic's Commercial Terms of Service state that Anthropic "may not train models on Customer Content from Services" (section B, https://www.anthropic.com/legal/commercial-terms). If your Order Form selects Amazon Bedrock, AWS states that model providers "don't have access to Amazon Bedrock logs or to customer prompts and completions" (https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html). If it selects Vertex AI, Google states that it "won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction" (https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance).
3. Human oversight and controls
- A per-agent and global kill switch, per-client and per-run budgets and an append-only, hash-chained audit log of every run, tool call and decision.
- Approval is refused for the person who started the run, by default, so two people are involved.
- Agents treat documents and tool results as data rather than instructions and are tested against planted prompt-injection text.
- PortLev's evaluation suite checks each agent's controls before release. Live model evaluations are being added; until a live baseline is recorded, treat agent quality as unmeasured.
4. Known limitations
- Drafts can contain errors, including wrong dates, names or policy references. The approver is responsible for checking them.
- The Policy agent knows only your handbook. If the handbook is silent, it says so and routes the question to your HR owner.
- Policy comparisons show differences in written terms; they do not tell you what the law requires in any state. Items marked "flag for legal review" need your counsel.
- Pay-range checks confirm a range is present and matches the marks you configured. They do not determine compliance with any pay transparency law.
5. Your notices and assessments
Your company makes the employment decisions, so laws on AI in employment decisions apply to your company as the employer. People Desk is designed so its outputs are preparatory and do not select or rank people. Even so, decisions you make with help from agent outputs (for example harmonizing benefits after an acquisition or mapping roles) may be covered by:
- Colorado SB 26-189 (signed May 14, 2026; in force January 1, 2027), which requires notices before using covered automated decision-making technology for consequential employment decisions and explanations after adverse ones (summary: https://www.seyfarth.com/news-insights/colorado-enacts-artificial-intelligence-replacement-law.html);
- the Illinois Human Rights Act as amended by HB 3773 (in force January 1, 2026), which requires notice when AI is used for covered employment decisions;
- the California Civil Rights Council regulations on automated-decision systems (2 Cal. Code Regs. section 11008.1 and following, in force October 1, 2025), which treat a vendor acting for an employer as the employer's agent and require 4 years of records;
- the California Privacy Protection Agency regulations on automated decisionmaking technology, risk assessments and pre-use notices (in force January 1, 2027).
On request we will give you the information about the agents you need for these notices and assessments, including the agent descriptions above, the data each agent uses and our audit records for your tenant.
6. Report a problem
Email support@portlev.com or tell your PortLev operator. Security issues: security@portlev.com.