# Counting America: how census technology failed in 2010, then came in under budget in 2020 The 2010 census handheld computers failed and reverted to paper, adding billions. A decade later the 2020 count finished near $13.7B, under its $15.6B estimate. Author: J.A. Watte Published: July 16, 2026 Source: https://jwatte.com/blog/decennial-census-technology-failures/ --- Every ten years the United States has to do something no other routine of government requires: find and count every person living in the country, all at once, from a standing start. There is no permanent counting machine humming between decades. The Census Bureau, which sits inside the Department of Commerce, builds the operation almost from scratch each cycle, tests it, staffs it with hundreds of thousands of temporary workers, runs it, and then largely stands it down. The stakes are constitutional. The count from Article I, Section 2 apportions seats in the House of Representatives, and it steers the distribution of hundreds of billions of dollars in federal funding. Because the census is a once-a-decade megaproject built on new technology each time, the Government Accountability Office has repeatedly placed it on its High-Risk List. This is the story of two cycles and two very different endings. In 2010, the flagship modernization, a fleet of handheld computers, failed under testing and was pulled back to paper for the biggest field operation, adding billions and making that count the most expensive in the nation's history at the time. In 2020, the Bureau again bet on new IT, again drew a High-Risk designation, saw its estimate climb past $15 billion, and then finished under that estimate. Both arcs are true, and neither cancels the other. ## The 2010 arc: a handheld program that fell back to paper In 2006 the Census Bureau awarded the Field Data Collection Automation (FDCA) contract to Harris Corporation. The plan was to modernize the most labor-intensive parts of the count by putting handheld computers in the field, both for address canvassing (walking the country to verify where housing units are) and for nonresponse follow-up, or NRFU, the door-knocking that reaches households that do not mail back a form. The award is commonly cited at about $600 million over five years, with the underlying announced value near $595 million, per the Census Bureau's own release. The Bureau planned to deploy more than 500,000 devices. That is not what happened. As the program advanced, the cost of the revised effort climbed well past its original value. Harris's estimate for the revised program reached roughly $1.3 billion, and reported figures across documents span a range of about $900 million to $1.3 billion. It is important to keep the distinction clear: that higher number is an estimate of the revised program, not money already spent and not the original award. The reasons matter, and the auditors were specific about them. The Commerce Department Inspector General and GAO attributed the overruns largely to government management rather than to simple contractor failure. The Bureau delivered more than 400 requirement changes late, many arriving in January 2008, well into development. As Government Executive reported on the auditors' findings, mismanagement rather than the technology itself drove the handheld trouble. When requirements churn that late and that heavily, cost and schedule follow. The devices also performed poorly in testing. During the 2006 to 2009 Census Dress Rehearsal, the handhelds were reported to be too slow, prone to freezing, and unreliable at transmitting data. GAO's January 2008 testimony, titled to flag "significant problems" of a critical automation program contributing to risks facing the 2010 Census, documented reliability risk directly. The precise symptom words, slow and freezing and connectivity problems, come from contemporaneous reporting rather than verbatim GAO text, so they are best read as press-characterized descriptions of a documented reliability failure. ### The April 2008 decision On April 3, 2008, Commerce Secretary Carlos Gutierrez announced the pivot. For nonresponse follow-up, the largest and most demanding field operation, the Bureau would revert to a paper-based process. Harris and the handhelds would focus on address canvassing instead. This was not a wholesale scrap. More than 100,000 handhelds still ran the 2009 address canvassing operation. The reversion applied specifically to NRFU, which is why the accurate framing is "dropped for nonresponse follow-up, kept for address canvassing." Reverting a nationwide door-knocking operation to paper on a compressed schedule is expensive. The same 2008 release put the revised life-cycle estimate for the 2010 Census at $13.7 billion to $14.5 billion, up from the $11.8 billion figure in the FY2009 budget request. One nuance the record insists on: that $11.8 billion prior estimate included about $1.8 billion for the American Community Survey, a separate ongoing survey, so not all of it was decennial field cost. Measured against the prior request, the reversion added an estimated $2.2 billion to $3.0 billion over five years. The outcome was a record. GAO's post-count review, GAO-11-193, found the 2010 Census to be the costliest in the nation's history at that time, at roughly $13 billion, with the average cost per housing unit rising from about $70 in 2000 to about $97 in 2010, in constant 2010 dollars. GAO noted the cost had been roughly doubling each decade and called for fundamental reform. ## The 2020 arc: high risk, high estimate, lower final cost For the next cycle, the Bureau again reached for technology, and GAO again raised its hand. In February 2017, GAO added the 2020 Decennial Census to its High-Risk List, citing untested innovations, IT system development challenges, and significant cybersecurity risk. The 2020 design leaned on new systems, including the Census Enterprise Data Collection and Processing (CEDCaP) program, and introduced the first-ever internet self-response, letting households answer online rather than only on paper. Then the estimate climbed. In October 2017 the Bureau put the 2020 life-cycle cost at $15.6 billion, including about $1.2 billion in contingency. That was a jump of more than $3 billion, described as a 27 percent increase over the 2015 estimate of roughly $12.3 billion. GAO reviewed the quality and reliability of that estimate in GAO-18-635. It is worth stating plainly that $15.6 billion was the peak official estimate, not the final spend, and the two should not be blurred together. The 2020 count then had to absorb shocks that had nothing to do with IT design. It ran through litigation over a proposed citizenship question, and it ran through the COVID-19 pandemic, which disrupted field operations and schedules mid-count. And yet the money landed differently than the 2010 story would predict. In March 2023, GAO reported in GAO-23-105819 that by the time 2020 Census activity ends in 2024, the count will have cost roughly $13.7 billion, below the $15.6 billion estimate. GAO attributed part of the savings to field productivity gains, including the use of laptops and modern devices in place of paper, which made enumerators more efficient. (For accuracy against a figure that sometimes circulates: the correct pair is the $15.6 billion estimate and the roughly $13.7 billion final, not a "$14.2 billion" number that is not GAO's headline.) ## Timeline - **2006:** Census Bureau awards the FDCA handheld contract to Harris Corporation, about $600 million, with more than 500,000 devices planned. - **2006 to 2009:** Census Dress Rehearsal. Handhelds are reported slow, freezing, and unreliable at transmitting data. - **January 2008:** More than 400 late requirement changes delivered; GAO testifies to "significant problems" in the automation program. - **April 3, 2008:** Commerce Secretary Gutierrez reverts nonresponse follow-up to paper; handhelds kept for address canvassing. Revised life-cycle estimate rises to $13.7 to $14.5 billion from $11.8 billion. - **2009:** More than 100,000 handhelds run the address canvassing operation. - **2010:** Census completed at roughly $13 billion, the costliest ever at the time (GAO-11-193). - **February 2017:** GAO adds the 2020 Census to its High-Risk List over IT and cybersecurity. - **October 2017:** Bureau's 2020 life-cycle estimate reaches $15.6 billion, including about $1.2 billion contingency. - **2020:** First-ever internet self-response; count absorbs citizenship-question litigation and COVID-19. - **March 2023 (GAO-23-105819):** 2020 count projected to finish about $13.7 billion through 2024, under the $15.6 billion estimate. ## The money, dated and sourced - **2006 FDCA award:** about $600 million (announced value near $595 million), U.S. Census Bureau. - **Revised FDCA program:** estimated roughly $1.3 billion, with a reported range of about $900 million to $1.3 billion; an estimate of the revised program, not spent dollars. - **2010 reversion impact:** life-cycle estimate to $13.7 to $14.5 billion from $11.8 billion; an added $2.2 billion to $3.0 billion over five years. - **2010 total:** roughly $13 billion, the costliest census in U.S. history at that time; cost per housing unit about $70 (2000) to about $97 (2010) in constant 2010 dollars (GAO-11-193). - **2020 October 2017 estimate:** $15.6 billion, including about $1.2 billion contingency; a 27 percent increase over the 2015 estimate of about $12.3 billion. - **2020 final:** roughly $13.7 billion through 2024, below the $15.6 billion estimate (GAO-23-105819). ## The honest critique and the honest defense **The failure critique.** A once-a-decade, constitutionally mandated count has become a recurring GAO High-Risk program precisely because of technology. In 2010 the flagship modernization failed under testing after the Bureau churned hundreds of late requirement changes, forcing the largest field operation back to paper and adding an estimated $2.2 billion to $3.0 billion. That pushed the count to roughly $13 billion, the most expensive in history at the time, with cost per housing unit rising from about $70 to about $97. In 2020 the Bureau again bet on new IT, GAO again kept the census on the High-Risk List over IT development and cybersecurity, and the estimate ballooned to $15.6 billion. The pattern is a count that repeatedly over-promises on technology and then over-runs or reverts late, at real cost to taxpayers. The auditors' own diagnosis, that much of the 2010 damage came from government-side requirement churn rather than contractor incompetence, is not exculpatory. It points at a program management weakness that recurs across cycles. **The mission defense.** The census is not optional or discretionary. The Constitution requires counting every U.S. resident once each decade to apportion the House and to allocate hundreds of billions of dollars. Doing that at national scale, once every ten years, from a standing start, is genuinely hard. There is no annual production line to keep systems warm, so the Bureau must build, test, and staff a nationwide operation on a compressed schedule, and it bears the cost of learning the same lessons a decade apart. The innovation story is also not all failure. The 2010 handhelds were not scrapped wholesale; they still ran address canvassing. And in 2020 the first-ever internet self-response worked at national scale, while productivity gains from modern field devices helped bring the final cost, about $13.7 billion, in below the $15.6 billion estimate. The 2020 count did that while absorbing citizenship-question litigation and a pandemic, and still delivered the numbers the Constitution requires. Both readings survive the evidence. The 2010 arc is a cautionary tale about late requirements and unready technology. The 2020 arc shows the same institution putting a large IT bet on the field, riding out shocks, and finishing under estimate. The consistent thread, across both, is GAO watching closely and reporting the outcome in numbers anyone can check. ## Fact-check notes and sources - 2020 Decennial Census added to the GAO High-Risk List in February 2017 over IT and cybersecurity risk and untested innovations; Census Bureau sits within the Department of Commerce; the count is constitutionally mandated (Article I, Section 2). [GAO-18-141t](https://www.gao.gov/products/gao-18-141t) - 2006 FDCA handheld contract awarded to Harris Corporation, about $600 million (announced value near $595 million), covering address canvassing and nonresponse follow-up, with more than 500,000 devices planned. [U.S. Census Bureau, CB06-51](https://www.census.gov/newsroom/releases/archives/2010_census/cb06-51.html); award and scope also reported by [Government Executive](https://www.govexec.com/oversight/2008/04/census-to-scrap-handheld-computers-for-2010-count/41957/) - FDCA program cost grew from about $600 million toward an estimated roughly $1.3 billion for the revised program (reported range about $900 million to $1.3 billion); GAO documented "significant problems" and reliability risk. [GAO-08-550t](https://www.gao.gov/products/gao-08-550t) - Handhelds reported slow, freezing, and unreliable in transmitting data during the 2006 to 2009 Dress Rehearsal; the specific symptom wording is press-characterized rather than a verbatim GAO quote. [Government Executive reporting](https://www.govexec.com/oversight/2008/01/census-program-to-use-handheld-computers-said-to-be-in-serious-trouble/26029/) - Overruns attributed largely to government management, including more than 400 late requirement changes, rather than to contractor failure. [Government Executive, "Mismanagement, not technology"](https://www.govexec.com/oversight/2008/04/mismanagement-not-technology-caused-census-handheld-trouble-auditors-say/26672/) - April 3, 2008: Commerce Secretary Carlos Gutierrez reverted nonresponse follow-up to paper while handhelds were used for address canvassing; the revised life-cycle estimate rose to $13.7 to $14.5 billion from $11.8 billion (the prior figure including about $1.8 billion for the American Community Survey). [U.S. Census Bureau / Commerce release, April 3, 2008](https://www.census.gov/newsroom/releases/archives/2010_census/2008-04-03_2010_census.html) - 2010 Census was the costliest in U.S. history at that time, roughly $13 billion; average cost per housing unit rose from about $70 (2000) to about $97 (2010) in constant 2010 dollars. [GAO-11-193](https://www.gao.gov/products/gao-11-193) - October 2017 life-cycle cost estimate for the 2020 Census was $15.6 billion, including about $1.2 billion contingency, a 27 percent increase over the 2015 estimate of about $12.3 billion. [U.S. Census Bureau life-cycle cost estimate](https://www2.census.gov/programs-surveys/decennial/2020/program-management/planning-docs/life-cycle-cost-estimate_v2.pdf); estimate quality assessed in [GAO-18-635](https://www.gao.gov/products/gao-18-635) - 2020 Census final life-cycle cost about $13.7 billion through 2024, below the $15.6 billion estimate, aided by device and productivity gains. [GAO-23-105819 (March 2023)](https://www.gao.gov/products/gao-23-105819) ## Related reading - [The GAO High-Risk List and improper payments](/blog/gao-high-risk-list-improper-payments/): the watchdog behind this series and the list the census keeps landing on. - [The FBI's Virtual Case File and Sentinel](/blog/fbi-virtual-case-file-sentinel/): another federal IT modernization that failed before it recovered. - [The VA's Oracle Cerner EHR modernization](/blog/va-oracle-cerner-ehr-modernization/): a large government IT bet under sustained oversight scrutiny. - [The public-money programs index](/blog/public-money-programs-index/): the full set of "where the public money goes" write-ups. *This post is informational and journalistic, not legal or financial advice. It describes public programs and documented events; mentions of third parties are nominative fair use and no affiliation is implied.* --- Canonical HTML: https://jwatte.com/blog/decennial-census-technology-failures/ RSS: https://jwatte.com/feed.xml JSON Feed: https://jwatte.com/feed.json Hero image: https://jwatte.com/images/decennial-census-technology-failures.webp