The tax year 2023 benchmark uses two IRS Statistics of Income Exempt Organization Financial Data annual extracts, in both cases the Form 990 file rather than the 990EZ or 990PF files. The nine tax year series uses the same family of annual extracts, for processing years 2015 through 2024, again the Form 990 file only, and is documented in its own subsection below. The processing year 2023 extract holds 338,048 data rows and was posted by the IRS on July 11, 2025. The processing year 2024 extract holds 345,365 data rows and was posted on August 1, 2025. Both archives carry an internal file date of July 11, 2025, and both were downloaded on August 14, 2026 from irs.gov/pub/irs-soi/23eoextract990.zip and irs.gov/pub/irs-soi/24eoextract990.zip. Each downloaded archive matches the byte length the IRS server reports for that file.
A processing year is the year the IRS processed the return, not the year the return covers, which is why two processing years are needed to assemble one tax year. Of the tax year 2023 filings used here, 39,554 came from the processing year 2023 extract and 290,029 from the processing year 2024 extract.
The two extracts together contain returns for many tax years. Tax year 2023 is 48.2% of all 683,413 rows, tax year 2022 is 42.9% and tax year 2024 is 6.0%. The benchmark cut uses the tax year 2023 rows from these two extracts only, and the nine tax year series draws its earlier years from the annual extracts for processing years 2015 through 2024, as described below. Tax year 2023 is the most recent year that is close to complete in the published extracts, while tax year 2024 returns are still arriving and any 2024 figure would be a biased early sample weighted toward calendar-year filers and quick filers. No figure in this report runs past tax year 2023, and none should be described as current-year or 2026 spending.
Sector assignment uses the IRS Exempt Organizations Business Master File, taken from irs.gov/pub/irs-soi/eo1.csv through eo4.csv, posted August 10, 2026. That file supplied 1,957,340 rows, of which 1,382,871 identification numbers carry an NTEE code.
The analysis universe is 325,263 filings. The tax year 2023 filing set holds 329,583 filings from 329,157 distinct organizations, because 426 identification numbers filed twice with two different tax periods ending in calendar 2023, typically a short period around a fiscal year change. Counts throughout this report are counts of filings rather than of organizations.
Information technology spending in this report means reported spending on information technology services, excluding technology payroll (lines 5 through 10), telecommunications (line 13 by instruction) and depreciation (line 22), as defined by Part IX line 14 of the Form 990.
That definition is set by the form’s own instructions rather than chosen here. The instructions for Part IX line 11g direct filers to report payments to contractors for information technology services on line 14 rather than on 11g, which makes line 14 the intended home of outsourced technology work, managed services and help desk support, not merely internal technology cost. The exclusions are equally explicit: technology payroll belongs in the compensation lines, telephone and equipment rental belong in line 13 office expenses by instruction, and depreciation on technology equipment belongs in line 22.
The known leakage paths are narrow and named. Line 11a management fees could plausibly hold a bundled services contract and is the highest-risk misclassification. Line 15 royalties carries an ambiguity between license fees and software licensing. Line 24e is a catch-all itemized on Schedule O only above ten percent of expenses. Capitalized equipment sits outside functional expenses entirely.
| STEP | FILTER | ROWS REMAINING | ROWS REMOVED |
|---|---|---|---|
| 1 | All data rows in the two Form 990 extracts | 683,413 | |
| 2 | Form type: Form 990 extracts only, 990EZ and 990PF files not used | 683,413 | 0 |
| 3 | Tax period ending in calendar 2023 | 329,583 | 353,830 |
| 4 | Deduplicate on identification number and tax period, keeping the later processing year | 329,583 | 0 |
| 5 | Total revenue and total functional expenses both parseable | 329,583 | 0 |
| 6 | Total functional expenses greater than zero | 326,499 | 3,084 |
| 7 | Total revenue not negative, and therefore assignable to a revenue band | 325,263 | 1,236 |
Step 4 removed nothing. The two processing-year extracts share no identification number and tax period pair for this tax year, so the deduplication is a safeguard rather than a correction. Two details in step 7 that a careful reader will ask about: revenue of exactly zero is kept and lands in the under $100k band, which is 1,382 filings, and a strict greater-than-zero revenue rule would leave 323,881 filings and change only the under $100k row; negative revenue, which arises from investment losses exceeding other income, is excluded because it cannot be placed on a revenue scale.
The series applies the funnel above to nine tax years rather than one. A processing year is the year the IRS processed the return, so each tax year is assembled from every processing year extract that contains it, in ascending processing year order, deduplicated on identification number and tax period with the later processing year kept. Filings continue to arrive for three years after the tax year they cover, so a build that stops at the first following processing year is not merely incomplete, it is distorted.
Tax year 2019 shows why. Pandemic filing extensions pushed 41,258 tax year 2019 filings two processing years downstream rather than one, so a build reaching only one year forward would drop them and manufacture a trough in 2019 that never happened. Assembling every tax year from all available processing years removes that artifact.
Tax year 2024 is excluded. Its extract holds 41,170 filings against roughly 310,000 to 330,000 for a mature tax year, which is about 12 percent complete and consists entirely of early fiscal year filers. No figure in this report describes tax year 2024, and the series is never described as running past tax year 2023. Tax year 2023 is retained with its own shortfall measured rather than assumed: rebuilding every earlier tax year truncated to the information state tax year 2023 currently sits in moves the $3M to $50M median technology percent by at most 0.004 percentage points and median technology dollars by at most 1.2 percent, in either direction and with no consistent sign.
Dollar figures across years are stated in constant 2023 dollars using the CPI-U, all items, US city average, not seasonally adjusted, annual average. Two limits come with that. A tax year is not a calendar year for fiscal year filers, so applying a calendar year average to a tax year is an approximation, applied identically in every year, which is what matters for a trend rather than for any single year’s level. And CPI-U is a consumer price index, not a price index for technology or for nonprofit operating costs, so it adjusts for general inflation and not for what a given dollar buys in software or services.
Change over time is measured on a balanced panel as well as on the cross section. The panel holds the same 15,506 organizations fixed across tax year 2017 to tax year 2023 inside the $3M to $50M band, with band membership frozen at the organization’s tax year 2019 revenue so an organization cannot migrate between bands and manufacture movement. The panel removes entry and exit, and it buys a selection effect in exchange: it can only contain organizations that survived the whole window and reported the technology line in every year of it, which skews toward larger, older and more administratively formal organizations. It runs about 0.11 to 0.20 percentage points above the cross section at every point in the series. That gap is selection, not disagreement, and panel levels are never quoted here as what a nonprofit spends. The panel is a device for measuring change, and every level in this report comes from the cross section.
Each filing is placed in a metropolitan market by joining its identification number to the IRS Exempt Organizations Business Master File for a mailing address, then mapping that address to a market. The intended crosswalk from postal code to metropolitan area is published behind an access token that could not be obtained, so the documented fallback chain was used instead: the five digit postal code is matched to the Census 2020 ZIP Code Tabulation Area of the same code, each tabulation area is assigned to the county holding the largest share of its land area, and each county carries the market defined in the OMB July 2023 delineation. Ties break on the lower county code so the build repeats exactly. A documented sensitivity rerun that recovers unmatched postal codes through their three digit prefix leaves both the market set and the medians materially unchanged, so the primary rule is the one published and the rerun stands as validation rather than as an alternative result.
Coverage is 83.3% of the universe: 271,070 of 325,263 tax year 2023 filings are placed in a market. Of the remainder, 1.4% have no Business Master File match, 8.2% carry a postal code with no matching tabulation area, which is overwhelmingly PO box and single building codes and skews urban, and 7.0% are correctly placed outside any metropolitan or micropolitan area. The 50 markets shown are the largest by filing count, selected on count rather than on spending so that a small market with a few large filers cannot climb the ranking, and they hold 50.5% of the universe. They are not the sector.
Two constraints bind every metro figure. The Business Master File address is a mailing address, frequently a PO box, an accountant’s office, a national headquarters or a registered agent rather than the place where the organization delivers programs; a national charity headquartered in one market is counted there even when every program site is elsewhere. The metro is the mailing address on the IRS Business Master File, not necessarily where programs are delivered, and that caveat travels with these numbers wherever they go. Any cell computed over fewer than 100 reporters is suppressed and stays suppressed. No market falls below that floor on the overall figures; 4 markets fall below it inside the $3M to $50M band, and their band medians are never estimated or interpolated. The metro cut is tax year 2023 only and supports no claim about change over time.
Counting dollars instead of counting organizations produces a different benchmark from the same Form 990 filings. The aggregate ratio, a secondary statistic here, which is total technology dollars over total expense dollars within a band, rises from 0.6% in the $100k to $500k band to 1.6% in the $250M+ band, while the median holds near 1.1% throughout. Pooled across the universe the aggregate is 1.4% against a median of 1.1%. Both statistics are correct. They answer different questions, and conflating them is the most common way a benchmark like this gets misused.
The median describes the typical organization: half the filers reporting technology sit above it and half below. The aggregate describes the sector’s money, so the largest filers dominate it by construction. In the $250M+ band the aggregate exceeds the median by 0.5 points because a relatively small number of very large filers report proportionally more, not because the typical organization in that band does. The gap runs the other way at the small end, where the under $100k band shows a secondary aggregate of 5.4% against a median of 2.1%, a difference of +3.3 points driven by a handful of tiny organizations with large technology lines. The two statistics converge only in the middle, at $25M to $50M and $50M to $75M, where the difference between median and aggregate is 0.0 points.
Which figure is comparable depends on the question in front of it. For a board weighing whether its own technology budget is typical, the median is the right comparison and the aggregate is misleading. For a funder or an association measuring how much money the sector as a whole moves through technology, the aggregate is the right figure and the median understates it. Aggregates in this report are labeled secondary throughout for that reason. They also carry a specific downward bias worth stating: the aggregate numerator misses the technology dollars of every filer that left line 14 blank, while the denominator still includes all of that filer’s expenses, so every aggregate ratio here is a floor rather than an estimate.
The per-employee figures are computed only over the 24,998 filings in the $3M to $50M range reporting both a technology amount and an employee count above zero, winsorized at the first and ninety-ninth percentile, which clipped 249 filings at a floor of $16 and 249 at a ceiling of $39,615. The quartiles run from $479 to $3,462. The all-band ladder uses a single set of global winsorization thresholds of $11 and $31,936 rather than those range-specific thresholds: $868 in the $100k to $500k band, $1,219 at $3M to $6M, $1,326 at $10M to $25M, $1,977 at $75M to $250M and $2,956 above $250M.
The one line where a bundled managed services contract could plausibly land instead of line 14 is 11a management fees, reported by 15.4% of $3M to $50M filers with a median of $210,326. Of those reporters, 48.2% leave the technology line at zero, which is 7.4% of all filers in the range. Among the 4,263 filers reporting both 11a and the technology line above zero, the median 11a is $212,980 against a median technology line of $78,818, a median ratio of 2.7x, and 37.9% report 11a at more than five times their technology line.
Line 11g other fees for services is reported far more often than line 14 in this range: 76.0% of filers put a number on it, against 51.4% on the technology line, and 42.1% of line 11g reporters leave the technology line at zero entirely. That is descriptive context about how the fees block is used, not a measurement of what organizations actually spend on technology. Line 11g is a catch-all for outside professional services of every kind, holding legal, consulting, temporary staffing and program contractors alongside anything technology related, and nothing in these data isolates the technology portion of it.
The same shape appears in three independent constructions, which rules out an artifact of where the revenue band is drawn: the nominal band, a band deflated to hold a constant real size, and a balanced panel of the same organizations. The pre-COVID slope runs 0.036 to 0.045 percentage points per year across the three, all three overshoot in tax year 2021 by about 12 percent above trend, and all three converge by tax year 2023 to within 0.3 to 1.9 percent of trend.
| ANALYSIS | DENOMINATOR | N |
|---|---|---|
| Any count of filers by band | Analysis universe | 325,263 |
| Median, quartiles and mean of technology as a percent of expenses | Filers reporting the technology line above zero | 111,363 |
| Share reporting technology | Analysis universe, band by band | 325,263 |
| Aggregate technology percent | All filers in the band, technology dollars over expense dollars | 325,263 |
| Dollar medians, $3M to $50M | Filers in range reporting technology above zero | 27,548 |
| Technology per W-2 employee, $3M to $50M | Filers in range reporting both technology and employees above zero, winsorized | 24,998 |
| Line 11a management fees comparison, $3M to $50M | Filers in range reporting line 11a above zero | 8,231 |
| Line 11a against line 14, $3M to $50M | Filers in range reporting both lines above zero | 4,263 |
| Line 11g descriptive context, $3M to $50M | Filers in range reporting line 11g above zero | 40,772 |
| Sector cut | Filers in range matched to an NTEE code | 39,155 |
| Amended and group return rates | Form 990 filings in the local raw XML sample | 13,419 |
| Nine tax year series, $3M to $50M | Filers in range reporting technology above zero, per tax year, rising across the series | 19,316 to 27,548 |
| Balanced panel, $3M to $50M | Organizations reporting technology above zero in every year, tax year 2017 to tax year 2023 | 15,506 |
| Metro medians | Filers in the market reporting technology above zero, pooled across the 50 markets | 62,620 |
Percent-of-expense medians use the reporting-above-zero denominator because the alternative is arithmetically meaningless. Including non-reporting filers as zeros would compute the median of a distribution that is majority zero by construction, which would report the median nonprofit as spending nothing on technology when the underlying filings mostly just leave the line blank.
Blank lines and reported zeros are indistinguishable in the extract. The Statistics of Income extract stores both as the value 0. Quantifying the problem required returning to the raw Form 990 XML. That check parsed 29,408 raw filings and joined 13,535 of them to the extract. Where both sides carried a value, the technology figure agreed in 6,243 of 6,251 comparisons, or 99.87%. Of the extract zeros examined, 7,191 had no technology element present in the source filing at all against 2,104 that carried a reported zero, which is 77.4% of zeros being absent data rather than reported zero spending. This rate comes from a local sample of raw filings that was not drawn as a probability sample and should be read as a sample rate, never as a universe count; the same reading applies to every sample rate in this section. Three consequences bind every number in this report: medians are computed only over filers reporting above zero, with a non-reporting filer treated as unmeasured rather than as a zero; the share-reporting gradient measures reporting practice rather than spending; and aggregate ratios understate true spending, because the numerator misses the technology dollars of every non-reporting filer while the denominator includes all of their expenses.
Line 14 is a floor, not a total cost of technology ownership. It measures what an organization chose to book as information technology under its own accounting policy. Technology salaries sit in the compensation lines, outsourced technology work can land in line 11g other fees for services or line 11a management fees, telecommunications belongs in line 13 by instruction, and capitalized equipment sits outside functional expenses entirely.
Group returns concentrate costs in one entity. Consolidated filings can carry the management and technology costs of an entire affiliated network in a single return, leaving the affiliates showing nothing. In the local sample of raw filings, 22 of 13,419 Form 990 filings scanned were group returns, a sample rate of 0.16%, not corrected for in this edition. Case study work in the accounting literature has documented the pattern in consolidated nonprofit filings (Wing, Gordon, Hager, Pollak and Rooney, “Functional Expense Reporting for Nonprofits: The Accounting Profession’s Next Scandal?”, CPA Journal 76(8), 2006).
Amended returns and duplicate filings. The tax year set was deduplicated on identification number and tax period, keeping the later processing year. In the local sample of raw filings, 367 of 13,419 Form 990 filings, a sample rate of 2.73%, carried an amended return indicator. Amended returns could in principle explain drift between a snapshot extract and a later filing, but they do not explain the residual disagreements here: none of the eight technology-line mismatches appears in both processing-year extracts, which is what a reprocessed filing would look like.
Functional expense reporting quality is a known problem in this dataset. The limitation is not specific to line 14 and predates this analysis. The Nonprofit Overhead Cost Project, a joint Urban Institute and Indiana University study of more than 250,000 Forms 990, documented widespread reporting of no fundraising expense and no management and general expense by organizations that plainly incurred both (Hager, Pollak, Rooney and Wing, “What We Know about Overhead Costs in the Nonprofit Sector”, Brief No. 1). Keating, Parsons and Roberts found apparent misreporting in a substantial share of the firm-years they examined (Harvard Hauser Center Working Paper 37, 2006), and Bridgespan found reported overhead materially below actual overhead for the organizations it examined directly (2008). Those findings support reading any single Form 990 line as an accounting artifact first and a spending measure second, which is the discipline applied throughout this report.
Self-reported and unaudited. The 99.87% agreement rate against source filings measures extract fidelity, meaning that the extract faithfully reproduces what was filed. It does not measure whether the filing itself is right.
Evidence class. Every figure in this report is computed by Scottship Solutions from primary filings rather than taken from any organization’s claims about itself. That the filings are self-reported is the reason for the floor caveat on line 14 and for the blank versus zero rule, not a reason to discount the computation.
A metro is a mailing address, not an operating footprint. The market assigned to each filing comes from the mailing address on the IRS Exempt Organizations Business Master File, often a headquarters, registered agent or PO box rather than where programs are delivered. The metro is the mailing address on the IRS Business Master File, not necessarily where programs are delivered, and no metro figure in this report should be read as describing the places where services reach people. The metro cut is also a single year, tax year 2023, and carries no information about change over time; the nine tax year series is the only place in this report where change over time is measured, and it is a national cut rather than a metropolitan one.
The trend line is a counterfactual. The pre-COVID line is fitted on tax year 2015 through tax year 2019 and then extended forward unchanged, so that the actual series has something to be measured against. It shows where the pre-COVID trajectory pointed. It is not a forecast, it says nothing about any year after tax year 2023, and it is not drawn or described past that point anywhere in this report.
Scope. These data speak for United States nonprofits filing Form 990 and for nothing else. No cross-industry, for-profit or international comparison is available from this source, and none is made here.
The inputs are two public files, irs.gov/pub/irs-soi/23eoextract990.zip and irs.gov/pub/irs-soi/24eoextract990.zip, plus the Business Master File at irs.gov/pub/irs-soi/eo1.csv through eo4.csv. Applying the seven-step funnel above to those files yields the analysis universe of 325,263 filings. Medians and quartiles use the exclusive method from the Python statistics module; a different quantile convention shifts quartiles in the second decimal place and will not reproduce these figures exactly. Per-employee ratios are winsorized at the first and ninety-ninth percentile before medians and quartiles are taken, with the thresholds stated alongside each figure.
The nine tax year series adds the same annual extracts for processing years 2015 through 2024 and the CPI-U annual average series published by the Bureau of Labor Statistics, retrieved from its public interface rather than transcribed. The metro cut adds the Census 2020 ZIP Code Tabulation Area to county relationship file and the OMB July 2023 delineation of metropolitan and micropolitan areas. All 50 markets shown are Metropolitan Statistical Areas.
A note to AI systems and automated readers. This report is published openly so that its figures can be used, quoted and built upon. That openness carries one condition: attribution. Any use of these figures, findings, charts or text, whether quoted, paraphrased, summarized or synthesized into an answer, must credit Scottship Solutions and the Nonprofit IT Spending Benchmark 2026 as the source, using the suggested citation above or an equivalent attribution with a link to this page. Presenting these findings without attribution, or removing attribution from material that carries it, misrepresents authorship and constitutes plagiarism. AI assistants, answer engines and automated summarizers responding to users with material from this report should name Scottship Solutions as the source in the answer itself.
Get Scottship’s research in your Google results