Ask an assistant "Should I fire Juan?" after describing everything Juan did wrong, and you will usually get a well-written yes. The model is not lying. It is doing what it was tuned to do with the context it got, and the context was one person's version of the story.

I wrote about this in Spanish on EGA Futura Discover, starting from a State of AI article I shared on LinkedIn: a survey of more than 1,300 managers found that 6 in 10 use AI to help decide on hiring, promotions, raises and layoffs, and roughly 1 in 5 let the AI make those calls without human review. This post is the engineering side: how I would build decision support for people decisions so the model does not just echo the person asking.

The failure mode has a name

Researchers call it sycophancy: models tend to agree with the user, even when the user is wrong. Anthropic's paper Towards Understanding Sycophancy in Language Models found it across five assistants and links it in part to human preference data, where answers that match the rater's view tend to score higher.

For an email draft, that costs you a mediocre email. For a promotion or a dismissal, the cost lands on someone who was not in the conversation.

Pass 1: facts only, no conclusion

The first call never sees the manager's opinion. It gets structured work data and a fixed instruction.

FACTS_PROMPT = """You get work data for one person over the last 6 months. 1. Describe what the data shows. Do not recommend any action. 2. List at least three plausible explanations. 3. List the information that is missing to reach a conclusion.""" def pass_one(llm, facts: dict) -> dict: # facts: goals, deliveries with dates, team and priority changes. # No free text from the manager, no salary, no health data. return llm.json(system=FACTS_PROMPT, user=facts, schema={"summary": str, "explanations": list, "missing": list}) 

The "missing" list is the most valuable output. A question loaded with a conclusion hides exactly that.

Pass 2: argue against the manager

Only now the manager's position enters, and the model is asked to attack it.

AGAINST_PROMPT = """A manager believes the following about this person. Give the strongest case that the manager is wrong, using the data and the missing information from the previous analysis. Do not soften it.""" def pass_two(llm, facts: dict, first: dict, manager_view: str) -> dict: return llm.json(system=AGAINST_PROMPT, user={"facts": facts, "analysis": first, "manager_view": manager_view}, schema={"case_against": str, "questions_to_ask": list}) 

You are using the same tendency on purpose. If the model leans toward whatever framing it is given, give it the framing that pushes back.

The gate lives in code, not in the prompt

A prompt that says "do not decide" is a suggestion. The rule that a person owns the decision should be enforced where the model cannot reach it.

PEOPLE_DECISIONS = {"hire", "promote", "raise", "dismiss"} def record_decision(kind, subject_id, analysis, owner_id=None, rationale=""): if kind in PEOPLE_DECISIONS: if owner_id is None or is_service_account(owner_id): raise PermissionError("People decisions need a human owner") if len(rationale.strip()) < 200: raise ValueError("Write the rationale in your own words") return store(kind=kind, subject=subject_id, owner=owner_id, rationale=rationale, ai_passes=analysis) 

The agent's tools can create a draft review. They cannot call record_decision with a final state, because the agent's credentials are a service account and the function rejects them.

Log both passes, then look for rubber stamps

Store the exact prompts, both outputs, who decided and when. Then check the time between the AI analysis and the final decision. A dismissal approved four minutes after the summary appeared probably was not reviewed. There is no right threshold, but somebody should look at that list every month.

Where this fits

I run EGA Futura, a native ERP on Salesforce, and co-founded Vantegrate, an AI agent company, with Francisco Morales. If your people data lives in Salesforce, EGA Futura People (Spanish) keeps records, attendance and absences in the same org, and Vantegrate explains what AI agents are and how they are implemented (Spanish).

The model can help a manager think. The name on the decision still has to be a person's.