How Eva works under the hood, and why it matters for anything that touches pay
By Vivek Suriyamoorthy, CTO, Everstage
The fastest way to lose a comp team's trust is to give them a number that changes when they ask twice.
That was the design constraint for Eva from day one. Large language models are remarkable at understanding a question, deciding how to approach it, and explaining an answer. They are not reliable calculators, and they don't know what your company means by "closed deal." In most software, an answer that's right 95% of the time is useful. In compensation, the other 5% is a dispute, a restatement, or an audit finding.
So we split the job. The AI decides what to ask. The platform guarantees the answer.
Why AI on revenue data usually goes wrong
Point a general-purpose model at raw revenue data and ask for attainment by region. To answer, it has to work out which tables matter, how they join, which deal stages count as closed, which fiscal calendar you use, how splits are credited, and which records are test data. Every one of those is a chance to be plausibly wrong. Ask the same question tomorrow, and it may make different choices.
The issue isn't the model. It's that the model is guessing the business logic that already exists in your organization, usually in someone's head or a spreadsheet.
Definitions first, AI second
Everstage already holds the most precise version of that logic: the commission plans, crediting rules, quota assignments, and data transformations that produce official payouts. Eva is built on top of it.
Each customer's business definitions are encoded once as governed, queryable views, each documented in plain language: what counts as a closed deal, how quota credit applies, which period a deal lands in. When Eva gets a question, it searches a catalog of those definitions, reads what each field means, and queries the right one. It doesn't write its own joins against raw tables. The calculation runs as a standard query against the same logic that runs commissions.
The result is deterministic. Same question, same answer, every time. And every answer traces back to the definition it came from.
Keeping the history, not just the current state
Most systems only know what a record looks like right now. Revenue questions are often about what it looked like then: what a rep's quota was when the deal closed, or what stage an opportunity was in on the last day of the quarter. Eva can answer questions about how records looked at any past date.
Permissions live in the data, not the prompt
A common way to secure an AI system is to tell it what it isn't allowed to show. We don't trust that approach, and neither should you.
Eva never queries raw tables. Each user gets views already filtered to the rows and fields their role allows, before Eva runs a single query. Eva can't reveal data it never had in front of it. This is the same role-based access that governs everything else in Everstage. We didn't build a separate security model for AI.
Proposals, not actions
Eva reads freely within a user's permissions, but anything that changes the system goes through a gate. Eva proposes the change, shows what it will do, and a person reviews it through the approval chains admins already control. Each session runs in an isolated environment, and every action is logged with who asked, what changed, and the before and after.
LLM providers retain none of a customer's data, and Eva runs under our ISO 42001-certified AI management program.
Keeping the model out of the hot path
Language models are slow and expensive compared to a database query, so we use them where judgment is needed and nowhere else. When Eva builds a dashboard in Analytics Studio, the model helps design it once. After that, the dashboard renders from plain queries, and scheduled reports run without calling a model at all. That's why they're fast, and why they're consistent from one run to the next.
What this foundation unlocks
Because Eva sits on the platform's data, definitions, and permissions instead of beside them, every new capability inherits the same guarantees.
- Eva is autonomous and can run in the background. Watchers run on a schedule or the moment something happens, such as a statement locking or attainment changing, and deliver the result where people already work.
- Eva where you already are. Through the Everstage MCP server, Eva's capabilities are available inside Claude, ChatGPT, and Microsoft Copilot, under the same permissions.
- One engine across the platform. The same definitions-first approach, permission model, and approval flow carry across planning, quoting, and incentives as we extend Eva across Everstage.
The short version
We built Eva the way we'd want a new analyst on a comp team to work: learn the company's definitions before touching the numbers, see only what you're allowed to see, show your work, and never change anything without sign-off. The AI is what makes Eva fast and easy to use. The platform underneath is what makes it trustworthy.
