Board of directors compensation: how the number actually gets set

Board of directors compensation is not really a pay question. It is a data question. Every retainer, meeting fee, and equity grant a committee recommends comes out of a benchmarking study, and the study is only as good as the reading operation behind it. Here is where that data comes from — and why private-company boards fly blind.

Ask a compensation committee how it set this year’s director pay and the honest answer is a process, not a number. Someone — usually a consultant, sometimes an internal team — assembled a benchmarking study: what comparable companies pay the people who sit on their boards, sliced by size and industry, landed in a deck as a recommended range. The committee debated the range. Almost nobody questioned the data. We build the data operations behind studies like these for a boutique executive-compensation consultancy with US and UK offices, and the view from inside is that board pay gets set by whoever did the reading. So this is a piece about the reading: where it comes from, where it breaks, and what changes when it runs itself.

Where board of directors compensation data actually comes from

Three sources, and none of them is a number you can trust out of the box. The first is the proxy statement. Every US public company files a DEF 14A before its annual meeting, and inside it — past the pay narrative and the executive tables — director compensation has its own table: retainers, committee and chair fees, equity grants. Public, complete, free. Also buried in filings that run 150 to 200 pages, every company disclosing in its own format. You can buy the pre-packaged version from the data subscriptions, but everyone gets the same numbers, and your study’s rules are not in the package.

The second source is the pay survey. Associations and research publishers collect compensation directly from employers — companies share because the raw file never leaves the publisher — clean every cycle’s responses by hand, and publish percentile tables. One publisher’s pipeline we have sat with moves roughly 1,600 responses through a spreadsheet before a single table ships. The survey’s strength is exactly the proxy’s weakness: it covers private companies, which file nothing at all. Its weakness is the calendar. A hand-cleaned survey publishes once a year, and the figure in the February board deck was collected the previous spring.

The third source is the peer group, and it is pure judgment. Which dozen companies count as comparable — size, industry, ownership, where the board recruits from — decides the benchmark before a single number is read. At the consultancy, peer group development is constantly iterative: analysts keep their selection queries in a Word document and rerun them by hand each time the group moves. That is not a criticism. It is what benchmarking actually is — a judgment layer sitting on top of a reading operation.

Why private-company boards fly blind

Put the three sources together and the blind spot is structural. Public-company board compensation is disclosed: the director pay table sits in the DEF 14A whether anyone reads it or not. A private company files nothing. So a private board setting its own pay has two options — the survey, annual and aging, or whatever a consultant happens to hold from past studies. The Fortune 500 board works from complete, current, comparable data. The $300M family company works from last year’s percentile table and someone’s memory. Same governance duty, a fraction of the information.

The survey half fails quietly, through cohorts. A benchmark only means something inside a band — revenue, industry, ownership — and when the tables are built by hand, the bands are whatever the annual report had room for. We have seen a $250M company read as 80th percentile overall while sitting 45th in its own size band. Director compensation set against the wrong band is wrong in a way nobody at the meeting can see, because the deck shows one number and not the cut that produced it.

What a benchmarking study actually costs to produce

Now price the reading honestly. Pulling pay fields out of proxies at study scale takes roughly 150 to 160 fine-grained extraction rules per study, each one a judgment a senior consultant used to carry in her head: disclosed figures trump plan tables, non-disclosed never defaults to zero, target values never mix with actuals — because a benchmark built on targets corrupts the moment realized pay leaks in. Applied by hand, that is one analyst, one filing, one company at a time, and a director-pay study spanning dozens of companies means weeks of reading before the first table exists. Stretch the coverage universe toward the Russell 3000 and the reading does not scale at all.

And the output has to survive a room full of scrutiny. The number goes in front of a compensation committee with the firm’s credibility riding on it, so every extracted value needs a citation back to the exact spot in the filing, every run needs to reproduce the same answer, and anything the rules cannot clear needs a person. Same prompt, different numbers is not a rounding error in this business. It is an unwinnable client meeting.

What changes when the data pipeline runs itself

At the consultancy, those same rules now run as a system. Filings across close to the entire Russell 3000, plus the FTSE 350, are watched nightly and read against the encoded rule set instead of being opened one at a time. Every value lands with a citation to the page it came from. Low-confidence fields queue for an analyst instead of guessing. Approved rows lock and carry into the next study rather than being rebuilt from scratch. None of it was trusted on faith — the encoded rules were first run against five prior hand-checked studies and matched line by line, and the review surfaced real violations a senior consultant had missed by hand.

The survey side moves the same way. Encode the cleaning methodology — outlier thresholds, smoothing, cohort bands, the privacy rule that raw responses never ship — and the 1,600 responses clean themselves as they arrive instead of in one annual crunch. Percentiles recalculate every cycle, the benchmark stops going stale between editions, and the survey becomes something a subscriber can query by title, size, and industry rather than a PDF to interpret alone. The mechanics of that pipeline are on the pay survey data page (zaigo.ai/workflows/pay-survey-data); the proxy-reading operation, director tables included, is on the proxy extraction page (zaigo.ai/workflows/proxy-extraction).

Public companies file their board pay. Private companies guess theirs. The difference is a data pipeline, not a governance philosophy.

None of this makes the judgment cheaper. The peer group is still a choice, the recommendation is still a partner’s call, and the committee still owns the number. What disappears is the reading bottleneck that made fresh benchmarking a once-a-year event. That is our shape as an AI operating partner: we encode the rules your best people carry, the system runs in your accounts against your data, and whatever the rules cannot clear lands with a person, reason attached. When the pipeline runs itself, how much board members make stops being an annual research project and becomes a question you can answer the week the committee asks it.

If your board pay numbers still arrive once a year in a PDF, the next step is a 30-minute working call: zaigo.ai/book-a-call. Bring last year’s study — we will tell you on the call what reading the filings and surveys behind it against encoded rules would take, and what the next edition costs when the pipeline does the reading.