Internal consistency is the constraint people discover too late
Offers are set from the band and from the candidate's expectations. What people in that role are already paid often becomes visible only after the offer has been discussed.
The offer that causes the longest-lasting problem is the one that pays more for the same work than an existing team member already receives. That problem shows up many months later and is costly to resolve.
Market data works the other way around, and it is easy to give it too much weight. A range compiled from different markets, seniority levels, and definitions is a weak standard presented as a benchmark.
The agent prepares the pack and does not propose a figure. Compensation decisions create equal-pay obligations, affect internal fairness, and affect the budget. The people who are accountable for all three should make those decisions.
How the agent is built in Agent Studio
The focus here is restriction. When the agent assembles and presents a figure, it must state the internal comparator position before any external data, and it must make no recommendation. This sequence is deliberate.
Prompt skills hold your banding structure and your stated compensation philosophy. That is why the presentation uses your terms rather than generic ones.
A PostgreSQL or HTTP Request node supplies internal comparator data as anonymized aggregates. Those aggregates show position within a band for equivalent roles and never name individuals.
When you use market data, it comes from a Google Sheets node that holds the sources you chose, with dates and methodologies. A benchmark without a date cannot be treated as evidence.
Agent Chat assembles the pack. Notion then receives it: band position, the internal comparator distribution, sources and dates, and any points where the data does not support the claim.
Access is narrow. Compensation data is some of the most sensitive information in the business, so the agent must do only this task.
The output contains no proposed figure. The system prompt states this explicitly.
What this agent is built from
- System prompt: assemble and present only, propose no figures, and lead with internal comparators.
- Prompt skills: your banding structure and stated compensation philosophy.
- PostgreSQL (Data): supplies internal comparators in anonymized aggregate and never names individuals.
- Google Sheets (Integration): holds the chosen market sources, with dates and methodologies recorded.
- Agent Chat (Util): assembles band position, comparators, and market context.
- Tool scope: narrow to this task. Compensation data gets no adjacent access.
- When an ActionFlow is enough: If the role is a standard one, pulling the band and the approved range is a workflow, and that is the whole preparation.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.