BlackLeafwatch the watchmen
IRS Automated Collection Service (chat and live-chat applications)

IRS Made Live Chat Permanent. It Never Tested Whether It Works.

Summary

A June 2026 audit by the Treasury Inspector General for Tax Administration found the IRS made its Automated Collection Service live chat program permanent without ever assessing whether it worked. Of 635,684 live chats logged over 2023 and 2024, 345,503 -- 54% -- went unresolved, and the data meant to track performance is itself broken: one statistical report showed a single employee handling 603 chats at once, though the system caps concurrent chats at three. The IRS spent $7.2 million in Inflation Reduction Act funds on the tools over two years while its live-assistor headcount fell a fifth and its total workforce fell a quarter.

By Marcus Aurelius · July 16, 2026

The Treasury Inspector General for Tax Administration -- , the independent watchdog Congress created specifically to audit the -- found that management "has not taken sufficient steps to evaluate the effectiveness" of its Automated Collection Service (ACS) chat applications. Nevertheless, the expanded the program and made its live-chat pilot permanent in June 2025. Nobody can say the expansion worked: of 635,684 live chats the 's own statistical reports logged in Calendar Years 2023 and 2024, 345,503 -- 54% -- went unresolved, meaning the taxpayer never got an answer and had to start over or call the phone line the chat was supposed to relieve. Worse, found the tracking data itself is broken -- one report showed a single employee handling 603 live chats at the same time, a physical impossibility the 's own system is supposed to prevent by capping concurrent chats at three.

A pilot that outgrew its evidence

The ACS function -- the unit within Small Business/Self-Employed that chases balances due and delinquent returns -- began piloting live chat in November 2017 and added an automated chatbot in December 2021. The 2019 Taxpayer First Act required the to build a customer-service strategy meeting "taxpayers' reasonable expectations" and expand online services, and the agency's internal 2023 Customer Experience Action Plan credited live chat with "saving taxpayers' and employees' time." But 's review found the never analyzed whether that claim -- or a related internal report crediting chatbots with reduced phone demand -- was actually true. The chat applications generate statistical reports on handle times, wait times, and concurrent-chat counts, but ACS management confirmed to auditors it has no performance measures and does not verify the reports' accuracy. The agency made the program permanent anyway.

Half the chats went nowhere

Using the same statistical reports, tallied what actually happened to the 635,684 live chats logged between February 2023 and December 2024: 290,181 (46%) were resolved -- the assistor addressed the inquiry or handed it to another function -- and 345,503 (54%) were not. The unresolved total breaks into three groups: 136,838 chats abandoned before a taxpayer reached an assistor, 127,288 disconnected mid-conversation, and 81,377 where the assistor gave up and referred the taxpayer back to the ACS phone line -- the very line chat was supposed to relieve. Unresolved chats also got worse over time: the count nearly doubled from 119,184 in Calendar Year 2023 to 226,319 in 2024. The doesn't know why. It captures no reason code when a chat is abandoned, and when a taxpayer is bounced to the phone line, the assistor has no field to record why.

Where 635,684 IRS live chats ended up
Calendar Years 2023-2024 combined, ACS live chat statistical reports
Resolved
290,181
Abandoned
136,838
Disconnected
127,288
Referred to ACS phone line
81,377
Source: TIGTA, Report 2026-308-029 (June 2026), Figure 1 and "The IRS is unable to determine why over half of live chats went unresolved," p.5
View data as table
TIGTA's Figure 1, reproduced exactly: unresolved live chats by year and category. TIGTA did not publish a CY2023/CY2024 split for resolved chats, only the two-year combined total (290,181) used in the KPI and bar figure above.

The tracking system has its own bugs

's harder finding is that the reports used to judge the program can't be trusted. Over 16,000 records showed a live assistor apparently handling more than one chat at a time -- normal, since assistors are allowed limited concurrency -- but 193 of those records showed someone working between 4 and 603 chats concurrently, against a system-enforced cap of 3. ACS management agreed the 603 figure was wrong. technology officials traced the likely cause to a separate bug: the "handle time" metric -- how long an assistor spends per chat -- is miscalculated when a chat isn't manually closed, so it auto-closes after four hours and inflates the concurrency count. IT staff filed a vendor ticket in July 2025; as of December 2025, the root cause was still unresolved. A second discrepancy compounds it: the reports logged 635,684 resolution codes against only 613,056 actual live chats -- 22,628 more codes than chats, because assistors were closing chats with a blank code or multiple codes, and nobody investigated why.

Live chats unresolved
54%
345,503 of 635,684 logged, CY2023-2024
Spent on chat tools
$7.2M
IRA funds, FY2024-2025
ACS live-assistor cut
20%
159 to 127 employees, by July 2025

Funded less, staffed less, tested never

The Inflation Reduction Act of 2022 appropriated $79.4 billion in supplemental funding, of which the agency spent $7.2 million on chatbot and live-chat tools in Fiscal Years 2024 and 2025 combined. Congress has since clawed back $53.5 billion of that appropriation across four laws -- $1.4 billion in 2023, $20.2 billion in 2024, another $20.2 billion in a 2025 continuing resolution, and $11.7 billion in the Consolidated Appropriations Act of 2026 -- leaving about $26 billion, roughly a third of the original sum. The staff running the program shrank alongside it: ACS's live-assistor corps fell 20%, from 159 to 127 employees, by July 2025, as the 's total workforce dropped 25%, from about 103,000 to 77,000, amid the administration's government-wide workforce reduction. A smaller staff working an unmeasured, glitch-tracked program means the entered a period of reduced capacity with no baseline to know whether service held, degraded, or improved.

IRS's IRA funding boost, cut by two-thirds
Inflation Reduction Act supplemental appropriation for the IRS, in billions
Appropriated, Aug. 2022
79.4
Remaining, Feb. 2026
25.9
Source: TIGTA, Report 2026-308-029 (June 2026), "Funding for chat applications," p.2, incl. footnote 4
View data as table
Appropriated, Aug. 202279.4Original Inflation Reduction Act supplemental funding for the IRS
Remaining, Feb. 202625.9After four rescissions totaling $53.5B across FY2023-2026 appropriations laws

Working several chats at once, with taxpayer data on the line

sampled the transcripts of 40 live assistors and found 24 (60%) working multiple live chats at the same time; of those, 12 had at least one authenticated chat -- meaning the taxpayer's identity had been verified and their tax-account data was live on screen -- open while working another. That matters because IRC Section 6103 makes it a federal offense to disclose one taxpayer's account information to a different taxpayer; working chats concurrently raises the odds of a mix-up. ACS management told auditors a control limits assistors to three concurrent chats, and separately said an April 2025 system update restricted authenticated work to one chat at a time. Field interviews contradicted both claims -- two assistors at two different sites said they believed they could work multiple chats, authenticated or not -- and the data backed the assistors: reports for September through November 2025, five months after the supposed fix, still showed 181 records where 23 assistors appeared to be running 2 to 11 concurrent chats.

A chatbot that stumbles on its own paperwork

also tested the automated chatbot directly, walking through 206 process flows and typing 53 keyword searches a taxpayer might reasonably try. Twenty-nine of the process-flow responses (14%) needed fixing -- some pointed to broken links, others handed over a form with no instructions for filling it out. Of the 53 keywords, 44 (83%) went unrecognized or drew an insufficient answer; "Debit Card" and "Tax Pro Account" just looped back to the same prepopulated menu, and some searches returned raw programming code instead of an answer. The Spanish-language chatbot doesn't even offer a keyword search box, limiting Spanish-speaking taxpayers to preset options English speakers can bypass. None of this uses generative AI -- the 's would-be AI strategy office was eliminated in March 2025, and the agency is waiting on a Treasury Department strategy with no timeline.

And no one is measuring taxpayer sentiment either way. OMB Circular A-11 Section 280 -- the governmentwide rule for how agencies must manage customer experience -- requires agencies to deploy a post-transaction survey for services like this one and act on the results, since the itself designates the ACS chatbot and live chat as High Impact Service Provider offerings. The has offered no live-chat survey since September 2022, and it doesn't review the responses to the two-question chatbot survey it does run.

  • management made ACS live chat permanent in June 2025 while its own inspector general found no evidence the agency had ever measured whether the tool works.
  • The 54% unresolved rate and the 603-concurrent-chat data error are not separate problems -- concluded that "the 's lack of reliable information hinders its ability to assess live chat performance," and cited GAO's internal-control standard that management should design control activities to mitigate risk to acceptable levels as the benchmark the is falling short of.
  • made nine recommendations and the agreed to all of them; evaluative performance reviews and the English-language chatbot's broken links and thin answers were already fixed, but the Spanish-language keyword search was still just a vendor change request, and validating the chat data's accuracy and correcting the handle-time calculation error remained open, with the and its vendor still chasing that root cause as of December 2025.
  • The unresolved-chat rate nearly doubled from CY2023 to CY2024, and it did so on a program the was simultaneously shrinking: by July 2025 ACS's live-assistor headcount was down a fifth, and by February 2026 the IRA funding behind the whole modernization effort had fallen by two-thirds.

's live-chat totals and its 54%-unresolved figure are drawn from the 's own statistical reports, which itself found contain outliers and reconciliation errors -- the report states its unresolved count 'could not determine the precise number' due to these same data-quality problems, so 345,503 is the best available reading of imperfect source data, not a clean count. The concurrent-chat finding (24 of 40 live assistors) comes from a judgmental, non-probability sample drew after narrowing to the average-concurrency records with the highest likelihood of accuracy; states explicitly that results from this sample cannot be projected onto the full population of live assistors. Review fieldwork ran from August 2024 through September 2025 across three sites: the Brookhaven Service Center in Holtsville, N.Y., the Office of Online Services in the New Carrollton Federal Building in Lanham, Md., and the Philadelphia Service Center.

Sources(3) ▾
  • Treasury Inspector General for Tax Administration (TIGTA), Opportunities Exist to Improve the Quality of Chat Applications (Report 2026-308-029) (2026-06-15)The audit itself: found the never evaluated whether its Automated Collection Service (ACS) chatbot and live-chat tools work, made live chat permanent anyway, and that the performance data it does collect contains outliers (a 603-concurrent-chat reading against a 3-chat system cap) and discrepancies (22,628 more resolution codes than live chats). Source for the unresolved-chat breakdown, funding, staffing, and chatbot-accuracy figures. oversight.gov · original document
  • Office of Management and Budget, OMB Circular No. A-11, Section 280 — Managing Customer Experience and Improving Service Delivery (2025-08-01)Independent confirmation of the post-transaction customer-feedback survey mandate that found the is not meeting for its live-chat application. whitehouse.gov · original document
  • U.S. Government Accountability Office, Standards for Internal Control in the Federal Government (GAO-25-107721) (2025-05-15)Independent confirmation of Principle 10 (Design Control Activities), the internal-control standard invokes for why the should be measuring and monitoring its chat applications' performance. gao.gov · original document
Weekly digest: the most-read systems, in brief. Mondays.

Comments

Always open. Logged-in readers can annotate paragraphs in place.

Loading comments…
or log in to comment under your account