WASHINGTON, DC – AUGUST 30: A National Guard unit patrols the National Mall on August 30, 2025 in Washington, DC. The Metropolitan Police helped both teams leave via their buses after the game was suspended. The Trump administration has deployed federal officers and the National Guard to the District in order to place the DC Metropolitan Police Department under federal control and assist in crime prevention in the nation’s capital. (Photo by Andrew Leyden/Getty Images)
Getty Images
Getting a company ready to sell anything to the United States government, Joe Scheidler said, “has been a black box for so many years for so many people.”
Scheidler is co-founder and chief executive of Helios, a New York startup that raised $4 million in a seed round led by Unusual Ventures in July 2025 and a further $2 million in March 2026. Its product, Proxi, is what he calls “a policy and procurement intelligence platform that really acts as an AI native OS for public policy professionals.” It reads bills, regulations, agency notices and contracting opportunities, and drafts against them.
He is selling into a trade that has never had more money in it. Federal lobbying spending reached a record $5.08 billion in 2025, according to OpenSecrets, an 11% jump after inflation and the largest single-year increase since quarterly disclosure began. That followed a record $4.4 billion in 2024. Health care alone accounted for $868 million.
Artificial intelligence is now the thing much of that money is spent on. More than 3,500 federal lobbyists, roughly one in four, reported working on AI issues at least once in 2025, according to Public Citizen. The count of Capitol Hill lobbyists working the issue grew about 170% in three years. Lobbyists working on data centers grew about 500%.
What Scheidler is betting on is the next turn: pointing the technology back at the influence business itself.
The policy paradox
“We often talk about this notion of the policy paradox, where the pace of policy increases, but your capacity to deal with it is just decreasing precipitously,” Scheidler said on the On The Margin podcast.
The volume backs him up. The 2024 Federal Register ran to 107,262 pages containing 3,248 final rules, the highest page count on record. States introduced 1,134 AI-related bills in 2025 and enacted 131 of them, per the National Conference of State Legislatures. By March 2026, 45 states had already introduced 1,561 AI bills, meaning this year’s first quarter alone outran all of last year. Federal contract obligations ran to roughly $744.7 billion in fiscal 2025.
Scheidler’s own route into the problem ran through government. Public staff records confirm he was deputy chief of staff to Virginia state Sen. Dave Marsden in 2021. He says he then worked in international development, spent about two years at the White House supporting the stand-up of the Office of the National Cyber Director, and finished at the State Department running congressional engagement for the Partnership for Global Infrastructure and Investment, the G7’s answer to China’s Belt and Road. He co-founded Helios in 2024 with Joseph Farsakh, a former State Department official, and Brandon Smith, the company’s chief technology officer.
“There were a lot of vertical AI solutions popping up, market-to-market, legal, finance, healthcare,” he said. Nothing comparable was “apparent in the policy and government space.”
A category splitting in two
That gap is closing fast, and the money is sorting winners from losers with unusual bluntness.
Govini, which sells defense-acquisition intelligence, took a $150 million growth investment from Bain Capital in October 2025 at a valuation above $1 billion, having crossed $100 million in annual recurring revenue. Quorum, whose platform serves more than 2,000 organizations, took a strategic investment from Enlightenment Capital in July 2026 to build out agentic AI features. State Affairs raised $70 million to pair statehouse reporters with AI analysis.
FiscalNote went the other way. The company that went public by SPAC in 2022 promising AI for policy sold Oxford Analytica and Dragonfly to Dow Jones for $40 million in March 2025, sold its Australian subsidiary TimeBase to Thomson Reuters for $6.5 million, and reported full-year 2025 revenue of $95.4 million, down 21%. Its shares were delisted from the New York Stock Exchange on April 13, 2026, which triggered defaults on its convertible debt; creditors granted a forbearance eight days later. Its market capitalization was reported around $14.5 million in March 2026.
The lesson other founders have drawn is that this market punishes breadth. It is the same dynamic reshaping white-collar expert work in finance, where narrow tools aimed at one workflow have outrun general-purpose platforms.
Selling software into anything touching government remains its own discipline. Tyler Traudt, founder of DebtBook, said on the On The Margin podcast that the bar is brutal: “You can’t be 10% better, you gotta be 10x better.” He argued the technology finally forces the issue: “AI actually will be a compelling event for many of them. They’ll they’ll see such meaningful productivity gains. They’re gonna wanna go capture it.”
Xin Yan, co-founder of Sign, said on the On The Margin podcast that she rebuilt her company around that access. “I realized government is actually the gatekeeper of the real world, right? So basically they gatekeep all the users, all the data, and all the assets,” she said. “So like about two years ago, we turned our like crypto native business gradually into a B2G business.”
The government is buying too
Agencies are not waiting. The General Services Administration launched USAi, a no-cost evaluation environment for generative AI, in August 2025, and struck deals putting ChatGPT Enterprise in front of federal agencies for $1 a year, Anthropic’s Claude across all three branches for $1, and Google’s Gemini for 47 cents. A bid protest was filed against the dollar-a-year arrangements within weeks.
Federal agencies reported 3,611 AI use cases across 56 agencies in 2025, up 105% from 1,757 the year before, with roughly half deployed or piloted. Those figures exclude national security systems and the intelligence community.
The results are less tidy than the announcements. Of 12 agencies the Government Accountability Office reviewed, eight reported trouble getting the compute they needed and six cited difficulty finding or training staff. Officials at five of those agencies told GAO that generative AI “can produce biased outputs or hallucinations,” which the watchdog treats as an open hurdle to federal adoption.
Scheidler says that bar is the whole design problem. “The expectations for quality of outputs coming from a generative system or an agentic system are probably higher than any other vertical,” he said, because decisions in government carry consequences measured in lives or billions of dollars. “The room for error is practically zero.” His answer is provenance: building a product “that prioritizes provenance relating to an output,” so a user can trace where any claim came from.
The astroturf problem
The trouble is that the capability that reads a docket faster also writes into one faster, and that machine has already been built and run at scale.
Of the 22 million comments filed in the Federal Communications Commission’s 2017 net neutrality proceeding, nearly 18 million were fake. New York’s attorney general found that lead generation firms working for a broadband industry lobbying campaign manufactured more than 8.5 million of them, and secured over $4.4 million in penalties from three companies, plus a later $615,000 settlement from three more. The same firms had run more than 100 other campaigns aimed at regulators and public officials, generating over a million fake comments for other rulemakings. All of it predates generative AI.
Recent cases suggest the cost of doing it has collapsed. Tens of thousands of AI-generated emails hit Southern California’s air quality regulator opposing a plan to phase out gas appliances. Advocacy groups including the Sierra Club pushed California’s attorney general to investigate comments allegedly fabricated in residents’ names at the Bay Area’s air district. A North Carolina pipeline proceeding drew a similar wave in September 2025.
“What we’re seeing with AI is absolutely the next step in digital astroturfing,” said Samuel Woolley, a University of Pittsburgh researcher who studies disinformation in politics.
Academics studying the industry make a broader claim. A team from Trinity College Dublin, the University of Edinburgh, Delft University of Technology and Carnegie Mellon mapped 249 instances of corporate capture around four global AI policy events between 2023 and 2025, identifying 27 distinct mechanisms. The most common was narrative control. Abeba Birhane, who directs the AI Accountability Lab at Trinity College Dublin, said the industry leans on the claim that “regulation stifles innovation” to justify its grip on how the debate is framed.
Scheidler was asked on the podcast whether owning the layer between business and government hands his company a say over who gets to sell to the state. “We’ll never interfere with that. That’s for the governments to decide,” he said. Asked separately whether Helios steers clients toward particular policy positions, he said the company’s job “is to be unbiased and completely apolitical, nonpartisan and give objective truth seeking information.”
What is still missing
Varun Kabra, chief growth officer at Concordium, said on the On The Margin podcast that the shift is already underway: “the next step, which is already starting, is that the AI agents start transacting on your behalf.” The counterparty, he said, has “no way to verify where a real accountable human is behind the transaction.” That gap, he added, “could open a door to fraud, bots acting as humans, agents operating with no accountability.”
Swap the checkout page for a rulemaking docket and that is the question in front of every agency reading its inbox. The same gap shows up wherever software starts acting for people, whether that is agents trying to move money or automated decisions about who gets banked.
Scheidler frames the stakes in national terms, in a market where vertical AI has been rewriting old industries one workflow at a time.
“We are in a global competitive race as a country, a technology race. And it’s not a sure outcome,” he said. “And it’s going to take builders at every level.”
