The Right Salary Is a Number – Most Companies Are Just Guessing It

01 /The Concept

Overpaying burns margin. Underpaying loses people. For most companies, the gap between those two failure modes keeps shifting – and they’re navigating it blind. Compa is built to fix that.

The platform helps companies set optimal compensation for new hires and contract renewals, covering the full package: base salary, bonuses, and equity.

At its core is an AI analytics agent that continuously tracks what the market is actually paying for people with specific skills, qualifications, and experience levels.

The most interesting data source isn't job boards or public salary databases – it's the companies themselves. Compa operates on a give-to-get model: the platform integrates with clients' HR systems and automatically pulls offer and contract data, anonymizes and aggregates it, then distributes the insights back to all participants.

This creates a clear network effect: the more companies on the platform, the richer the data each individual company receives.

Importantly, the data isn't averaged into meaningless benchmarks.

First, compensation figures are tied to precise combinations of skills, seniority, experience, and additional parameters – at least five of them. The platform doesn't report "average Python developer salary." It reports something closer to "median compensation for a Senior Python developer with 10+ years of experience specializing in data analytics, plus specific additional dimensions."

Second, each company can build its own peer group – a reference set of at least 10 comparable companies whose compensation benchmarks it wants to track. The system even supports weighting individual companies in the peer group, to prevent outlier payers from skewing the averages.

A companion AI assistant for HR professionals When determining what to offer a candidate or extend to a renewing employee, the assistant draws on two inputs simultaneously: current market data (general and peer-group specific) and the company's own internal compensation rules. That combination makes recommendations immediately actionable.

Pricing starts at $35,000 per year for basic market data access. Additional capabilities – equity compensation data, skills-based tagging, AI agent features – add cost, meaning total spend can be several multiples of the base price.

For large companies hiring in-demand talent, that cost structure makes economic sense: better compensation targeting protects against both talent loss and unnecessary overspend.

Compa already counts OpenAI, Stripe, Okta, and Moderna among its customers.

The startup raised $3.9 million in 2021, then $10 million in 2023 – when it was [first reviewed here](/review/hochesh-interesnuju-model). A few days ago it closed a new $35 million round.

02 /Why It Matters

The underlying driver is simple: compensation for skilled roles doesn't stay stable long enough for annual benchmarking to remain useful. Companies can no longer run a market analysis once and coast on it for two or three years.

AI engineering compensation, for instance, has followed an unusually volatile curve from 2022 to 2024 – overall trending up, but in a jagged pattern that reflects waves of demand, supply shifts, and market corrections.

And it's not only tech roles showing this pattern. Nurses, HVAC technicians, plumbers, electricians, marketing content strategists – compensation trajectories in many fields are accelerating and becoming less predictable. Real-time market data is increasingly a competitive necessity, not a nice-to-have.

Skepticism about sharing sensitive compensation data with a third-party platform is understandable. But Compa isn't alone in this structural approach, and give-to-get models have worked in other sensitive B2B contexts.

Crossbeam ([covered here](/review/vmeste-prodadim) and [here](/review/protiv-kogo-druzhim)) raised $116.9 million and acquired Reveal (which raised $54.3 million). B2B companies connect their CRM data to Crossbeam – including current and potential customer lists. The payoff: identifying partner companies that are already working with a difficult-to-reach prospect, enabling warm introductions. The data is sensitive, but the value exchange is concrete.

The original give-to-get model in B2B goes back to Jigsaw, founded in 2004 – a contact database where members earned credits by contributing data and spent credits to retrieve it. Salesforce acquired Jigsaw in 2010 for $142 million.

David Sacks, the venture investor, argued in a 2023 article that give-to-get should become a standard architecture for AI startups operating in healthcare information, legal services, finance, scientific research, and manufacturing.

A recent example: Trata ([covered here](/review/chtoby-vygodno-investirovat-nuzhny-ne-horoshie-sovety-a-horoshie-fakty)), a Y Combinator graduate, aggregates anonymized market and company analytics from its fund clients and feeds the enriched data back to each participant – dramatically improving the quality of analysis available to every individual fund.

03 /Opportunities

Building a Compa-equivalent for your own target market is the most direct play. The business model is proven, the demand is real, and this kind of platform remains scarce outside of a few verticals.

The broader direction is applying the give-to-get architecture to AI platforms in other domains. It has structural advantages that compound over time.

First, it creates a network effect – the platform becomes more valuable to each existing client as new clients join.

Second, it creates a moat that's fundamentally harder to replicate than any technology advantage. In a market where AI capabilities can be copied quickly, being ahead on technology alone doesn't protect you for long. But having a unique, growing dataset that makes your AI materially better? That's much harder to replicate.

The OpenAI vs. Google dynamic illustrates this. OpenAI had a significant head start with the release of ChatGPT. But Google closed the gap – in part because its search engine accumulates data at a scale that no LLM startup can easily match from a standing start.

The real competitive advantage in AI isn't the model. It's the data the model learns from and operates on. Which means the most important architectural decision for an AI startup isn't which LLM to use – it's how to build data collection into the business model itself. Compa figured this out.

So: what data collection mechanism can you build into your AI platform? Or better yet – what AI platform can you design that is fundamentally dependent on a give-to-get data loop with its customers?