Reading Other Nonprofits' 990s With AI

What a 990 Actually Tells You, and What It Doesn't

Before pointing a model at anything, it helps to know what is genuinely in the document. The Form 990 is an information return rather than a tax computation, which means most of it is disclosure rather than arithmetic. The parts that matter most for comparative work are concentrated in a handful of places, and knowing which ones saves enormous time.

Part I gives you the summary: total revenue, total expenses, net assets at the beginning and end of the year, and headcount for employees and volunteers. Part VII lists officers, directors, trustees, key employees, and the highest-compensated employees, with reportable compensation and estimated other compensation for each, plus average hours worked per week. Schedule J expands on that for individuals above the reporting threshold, splitting compensation into base pay, bonus and incentive, other reportable compensation, deferred compensation, and non-taxable benefits.

Part VIII breaks revenue into contributions, program service revenue, investment income, and other categories, and separates government grants from other contributions. That separation is one of the most useful pieces of comparative information in the entire document, because it tells you how much of a peer's model depends on public funding. Part IX is the statement of functional expenses, allocating each expense line across program services, management and general, and fundraising columns. Part VII also has a second section that names the five highest-compensated independent contractors receiving more than one hundred thousand dollars, which is how you learn who your peers hire for technology, fundraising counsel, and consulting.

The limits deserve equal attention. The 990 is a lagging document. A filing you download today may describe a fiscal year that ended eighteen months ago, and organizations routinely file on extension. It is also self-reported and shaped by judgment. The functional expense allocation in Part IX depends heavily on how an organization chooses to allocate shared costs, and two organizations with identical operations can report meaningfully different program expense ratios based on allocation policy alone. Schedule B, which lists major contributors, is generally redacted in public copies, so you cannot see who funds a peer from the 990 itself. And the narrative sections describing program accomplishments are marketing copy written for an audience of funders and watchdogs, which is worth remembering when a model summarizes them back to you as fact.

Five Questions Worth Asking of a Peer's Filing

Benchmarking fails most often not because the data is bad but because the question is vague. Asking a model to tell you how your organization compares to a peer produces a bland summary that changes nothing. Asking a specific question with a decision attached to it produces something usable. Five questions repay the effort consistently.

The first is compensation. If you are about to post a director-level role, or your board is reviewing the executive's pay, the compensation tables across ten comparable organizations give you a defensible range rather than a guess. Boards setting executive compensation have a specific interest here, since the process for establishing reasonable compensation generally involves reviewing data on comparable positions at similar organizations, and 990 filings are the most accessible form that data takes. The exercise is more useful when you capture average hours per week alongside the dollar figure, because a part-time executive at a peer organization is not a comparison point for a full-time role.

The second is revenue concentration and mix. Pull the revenue breakdown for every organization in your peer set and the picture that emerges is strategic rather than financial. If most of your peers derive a large share of revenue from government grants and you do not, you are either missing an opportunity or deliberately avoiding a dependency, and either way the board should know which. This is also the most reliable early warning available about sector-wide fragility, since a peer set heavily weighted toward one funding stream is a peer set that will struggle together when that stream contracts.

The third is expense structure. The functional expense table shows how peers distribute spending across program, management, and fundraising, and more usefully how they distribute it across specific lines such as salaries, occupancy, professional fees, and technology. Comparing your technology spend as a share of total expenses against ten similar organizations is a far better argument for a budget increase than an assertion that you are underinvested.

The fourth is vendors and contractors. The independent contractor section names the highest-paid firms a peer works with and describes the service provided. This is how organizations discover which fundraising consultants their peers actually use, what a comparable organization pays for accounting or IT support, and which firms have deep sector experience. It is genuinely practical market intelligence that would cost real money to buy elsewhere.

The fifth applies mainly when you are looking at foundations and grantmaking organizations. The grants schedule lists recipients and amounts, which turns a funder's 990 into a map of what they actually fund rather than what their website says they fund. For a development team evaluating whether a foundation is a realistic prospect, the distribution of past grant sizes answers the question faster than any conversation. This complements the work covered in monitoring peer nonprofit performance more broadly.

Using AI Without Inventing Numbers

This is the section that determines whether the whole exercise is trustworthy. Language models are excellent at reading a document and pulling out requested values, and they are also capable of producing a confident, well-formatted, entirely fabricated figure when the value is not where they expected it. The failure is not random noise that you would notice. It looks exactly like a correct answer.

The single most effective safeguard is to require citation of location. Instead of asking for total program service expenses, ask for total program service expenses along with the part and line number where the figure appears and the exact text surrounding it. A model that has to point at the source is far less likely to invent a value, and more importantly, you can verify a claim in seconds rather than rereading the filing. Any figure that comes back without a location should be treated as absent rather than as a number worth checking.

The second safeguard is to insist on an explicit not-found response. Models default to helpfulness, and helpfulness in this context means guessing. Telling the model directly that returning a null value is the correct and preferred answer when a figure is absent changes behavior substantially. Small organizations filing the short form will not have most of the fields you are asking for, and a workflow that cannot say so will silently fill the gaps.

The third safeguard is arithmetic validation outside the model. Once you have extracted a table, check that revenue components sum to total revenue, that the three functional expense columns sum to total expenses, and that the change in net assets equals revenue minus expenses. These checks cost nothing in a spreadsheet and catch a large share of extraction errors, including the ones a human reviewer would miss. Any row that fails a check goes back for manual review rather than into the analysis.

The fourth is sampling. Pick three organizations at random from the finished table, open their filings, and verify every extracted figure by hand. If all three are clean, your error rate is probably acceptable. If any of them has an error, the workflow needs fixing before the results go anywhere near a board packet. This discipline is the same one that makes any AI-assisted data work defensible, and it applies equally to the data cleanup work most organizations face in their own systems.

One more practical note on document handling. Long filings can exceed what a model handles well in a single pass, and accuracy degrades when a model has to locate a small figure inside a very large document. Splitting the filing and asking for specific sections, or working from the structured XML where the relevant fields can be isolated directly, produces noticeably better results than dropping the entire document in and hoping.

A Starting Workflow You Can Run This Month

The version of this work that actually gets done is small. A first pass with ten organizations and four data points is more valuable than an ambitious sector study that never finishes, and it teaches you where the friction is before you commit to anything larger.

Start by writing down the decision the analysis is meant to inform. If the answer is that a compensation committee meets in October, your data points are compensation, hours, and budget size, and everything else is a distraction. If the answer is that the board wants to understand revenue risk, your data points are the revenue breakdown across three years. A benchmarking project without a decision attached produces a document nobody reads.

Then build the peer list and write down the inclusion rule before looking at any numbers. Ten to fifteen organizations matched on size, mission, and region is enough. Pull each filing for a single common tax year, preferring the structured version where it exists. Extract the specific fields you need with location citations, load them into a spreadsheet, run the arithmetic checks, and hand-verify three filings at random.

Finally, present the result as a range rather than a verdict. Boards and executives respond well to a statement that peer executive compensation for organizations of comparable size and region falls between two figures, with a note on how many organizations were reviewed and which were excluded. They respond badly to a single number presented without provenance, and they should, because a number without provenance is indistinguishable from a guess. The value of this whole exercise is that your guesses become documented ranges, and documented ranges are what allow a board to make a defensible decision.