14,000 people applied. Thirty-three got in. The jobs are at Homeland Security.

Nathan Labenz admits a blind spot. After a year of covering private-sector AI, he still did not have a clear picture of how the U.S. civil service was approaching the same tools. His guest is Michael Boyce, director of the Department of Homeland Security’s AI Corps, a career civil servant who will stay through the presidential transition and who will not spend the hour on electoral politics. The Hatch Act and the nature of the job, Boyce says, keep the focus on the core functioning of the government.

Ask what DHS does and most people reach for the TV show or a news clip. Boyce has drawn a DHS paycheck since he was 25, so the org chart feels natural to him and strange to almost everyone else. Created after 9/11, it is, he argues, the largest civilian department in the federal government — you can quibble over VA hospitals — with 260,000 employees plus contractors, and 22 agencies and offices inside it. FEMA, the Coast Guard, the Secret Service, the immigration agencies, TSA, and CISA are all in the building. Naming which federal acronyms sit under DHS, he jokes, is the most boring cocktail-party game in the world.

His high-level split is prevention and response on one side, service delivery on the other. Borders, security, and “bad things happening” are half the story. The other half is who can become a citizen, which packages may enter the country, and what disaster survivors are entitled to from FEMA. DHS is the largest federal law-enforcement agency *and* a huge civilian benefits shop. Continuity is almost total: in the CIO’s world of about 700 federal staff and 3,000 including contractors, three people are political appointees. Across a transition, 99.9 percent of the department keeps showing up.

AI is not new here. Using standard definitions, Boyce says DHS has been using it for about a decade, and a Trump-era executive order already forced big agencies to publish inventories. More than 40 public reports now sit at dhs.gov/data/ai_inventory. Travelers meet the work at the airport: CBP kiosks that scan a face, TSA checkpoints that match an ID to a face. Those systems are narrow. The phase change Labenz wants to talk about is generative models — general-purpose, open-ended, and wrong in new ways — which is why a dedicated team suddenly mattered.

Boyce had been on loan to OMB’s internal AI policy shop in late 2022, when colleagues still treated the beat as a “fun little fad.” ChatGPT landed, and interest inside government rose with everyone else’s. DHS announced the AI Corps in February 2024, after the executive order that called for a national AI talent surge. The department, he says, wanted to put its money where its mouth was: not one or two extra AI postings, but the largest civilian AI team in the federal government, 50 positions aimed at delivery and mission implementation rather than policy theater.

The model is an internal consultancy, partly the brainchild of CIO Eric Hysen, another U.S. Digital Service alumnus. About half the Corps will roam in teams of two to five, sitting with component partners to ship priority projects. Roughly a fifth works on scaling adoption — today that means policy, governance, and oversight; later, shared infrastructure. The rest is an incubator that centrally owns products, takes risks operational offices cannot afford, and exists so the Corps can walk the walk before it lectures anyone. Boyce was hired in part because he already knew the department’s scars. “I’m not here from Silicon Valley to save anybody.” Demand, he insists, is not the bottleneck: grassroots ideas, top-down visions, and his team’s matchmaking all show up. If you think federal agencies are uninterested in AI, he has a bridge to sell you in Brooklyn.

His own path is why the “bureaucracy hacker” title on his LinkedIn is not a joke. A New York public high school with a serious CS program; a 2007 neural net, built with a friend, that found Waldo and made the city science-fair semifinals; two or three years of college CS abandoned for seminars; a first job as a refugee officer who flew around the world interviewing applicants and running security checks. Process and technology pulled him back. He wrote code, joined USDS as a White House-adjacent tech squad, then ran technology for USCIS humanitarian and asylum systems and shipped machine-learning projects around 2017–2018. He now calls himself technical enough to be dangerous — Streamlit and Next.js on the side, not the next frontier model — and his first speech to new hires is to listen. Soft skills, in a 22-agency maze, are not a bonus. The team went from Boyce in April to three people two weeks later to 33 at recording, adding about five a month, while the rest of DHS still expected them to know every rule. His antidote to meeting culture is a templated one-to-two-page memo: stakeholders sign the paper before anyone books the room.

The headline internal project is a department-wide LLM chatbot, which Boyce cheerfully admits half the audience could stand up in fifteen minutes of Python. The work is the plumbing: logging, hooking the system into DHS cybersecurity monitoring, and running it on a hyperscaler cloud that has already passed federal controls. The point is not a clever chat UI. It is letting thousands of staff put in information that is not classified and not public — including personal information — so the model sits inside real workflows. Magic, he says, happens at the layer where you are analyzing actual data.

The application he loves most is not even his team’s from the start. At USCIS, asylum officers run three-to-four-hour interviews. Training used to mean pulling other officers off the interview rotation to pretend to be seekers. The replacement is an LLM, steered mostly by careful prompting, that has to recreate a messy experience: a translator in the middle, miscommunication, legal language the applicant may not have the schooling to parse. The first drafts answered with “according to the Immigration and Nationality Act section 104 subpart B.” Tuned, the bot keeps a consistent story, switches scenarios on the fly, and does not wander off-script the way a human actor does. Evaluation is qualitative — does this feel as good as a person, and does it make a new officer calmer before the first real interview? Colleagues who had heard of ChatGPT still asked how the team made it *this* good for their specific job.

Those pilots were not random. An AI task force put headquarters money on the table and asked the department for three or four generative projects to accelerate. Homeland Security Investigations — child sexual abuse material, human smuggling, the *Zero Dark Thirty* desk-officer-reading-documents job — won a slot with a Llama 3 70B RAG system on cloud GPUs. The engineer mixed standard k-nearest-neighbor semantic search with lexical search once he saw that phone numbers and names wanted exact matching, then used the 70B model as a judge of whether retrieved chunks were actually relevant. Boyce’s line, after quoting an Air Force Research Lab CIO on RAG giving people a “personal relationship” with their own data: there is no one RAG algorithm to rule them all. You get into the data, try hunches, and you know it when the results come back clean.

A quieter pattern, he thinks, is quality check. One unnamed team hand-jams unstructured reports into another team’s tracker all day, then bureaucracy adds a second and third review layer because people make 5 p.m. mistakes. An LLM as a second pair of eyes can lift accuracy from 94 percent to 99 percent without taking the decision away from humans — a much easier sell to executives whose careers were built on putting people on high-stakes work. Narrow systems still pay the bills too. A border vision stack flagged a truck in a second and a half; secondary inspection found 165 pounds of narcotics. FEMA’s disaster imagery went from more than a million human-reviewed aerial photos to on the order of 70,000. USCIS still takes millions of paper filings a year; an image classifier plus OCR labels birth certificates and benefit forms. Running those millions through an LLM, token by token, still loses to a cheap in-system classifier. General models will not eat the narrow ones on cost alone, not yet.

Labenz’s guess about the chatbot — Azure, contractual promises not to retain or train — is the private-sector anxiety in one sentence: the boss will not let us hit an external API. Boyce’s record is the opposite of a dark data center. In 30 to 50 software projects at the department, he has never shipped one off fairly standard cloud infrastructure. DHS already runs AWS, Azure, and the other big clouds. Of four public generative-AI pilots, some call frontier commercial APIs; one that his team touched ran an open-source Llama on a cloud GPU. He does not expect the department to rack its own GPUs, Science and Technology Directorate aside, and he does not think it will train a GPT-2 of its own.

The gate is FedRAMP, the Federal Risk and Authorization Management Program run by GSA. NIST’s risk-management framework spells out controls; a third party certifies them; logs and verification packages go to reviewers before a service lands on the FedRAMP marketplace anyone can search. The point is not that a fun API should never see a shipping label or an immigration file. It is that vendors spend real money proving the controls first. A person at a frontier lab told Boyce the process was a pain — and also one of the few he had seen that actually added security. Fortune 5000 companies building private GPU clusters “for cost,” Boyce and Labenz agree, are usually making a mistake the federal government has already declined to make.

Employee use is more open than the stereotype. The latest executive order discourages agencies from blocking generative AI; some still cut off ChatGPT at the VPN. DHS did not. Take an hour of training, get a supervisor’s sign-off, finish cybersecurity and privacy modules, sign rules of behavior, and most staff can log into ChatGPT, Claude, or Bing. Two red lines: no personal information in the commercial tools, and no using a model to make a final decision on public benefits. That is why the Corps is building the more locked-down internal chatbot. Thousands of people have already sat through generative-AI brown bags; GSA’s government-wide training drew 10,000 to 15,000 sign-ups last year. When colleagues saw the asylum role-play, the reaction was not fear. It was “how did you get it so good?” — which Boyce treats as the training program that actually works: put the tool in the job and more ideas grow.

Boyce’s first metric is hiring: 50 AI-focused seats, the largest civilian AI team in government. Plenty of people told him the talent would not come. More than 14,000 applied. His staff tease him for disliking metrics; he says he dislikes metrics invented after the fact. Members of the Corps are drafting a roughly 30-page testing-and-evaluation guide — if you are in this scenario, here are the tests; if you are in that one, try these. For the asylum role-play the numbers that mattered were qualitative: better or equal to a human, more comfortable before the first interview. For the internal chatbot he will want usage *and* the shutdown test: if we stop funding this, does a critical process break?

He bristles at “70 percent more efficient” and at the five-minutes-a-day times headcount times workdays spreadsheet that “saves $5 million.” No one ever gives the $5 million back. He would rather know whether cognitive load is moving in the right direction. Asked what five-year failure looks like, he names three: the public does not trust the work, so Congress and political appointees will not fund it; the department uses the tools for ill and fails to protect privacy and civil rights; or it spends good money after bad because design, product, and engineering practice were sloppy. This is not *The Matrix*. For the business cases in front of him, model quality is rarely the blocker. Integration into existing flows is.

Labenz asks the delicate question: a staff-free TSA checkpoint, given how easy it is to glaze over at a bag-scan screen. Boyce had just visited TSA’s innovation center outside Las Vegas airport. Policy today is not to replace people on benefits and other critical decisions. Practically, a human saying “put your shoes on the rack” is load-bearing; ten extra seconds per passenger and the line runs out of the building. The next step is better algorithms in the hands of the officers who already argue about which machine’s interface they like. Quality use cases, he repeats, come first because they are safer, they build trust, and they are how you find out whether the system works before you try to take a person out of it.

The Corps is still hiring toward 50, fully remote, with 33 already on board and more in selection. Roles are cross-functional on purpose: backend application engineers who can live with latency and fault tolerance, ML engineers, data scientists, security engineers, policy analysts, product managers, designers, and a catch-all “technologist” seat. Base pay is around $150,000 and generally tops out near $230,000, which Boyce places in the top 3 percent of American wage earners, plus a pension that kicks in after three years and a 401(k) match. People making $2 million a year will not come. Vesting schedules have been the real turndown, not the salary. He has rarely lost a candidate on pay.

The surprise in the résumé pile is a loop: Google or Meta, a few years in government, a departure, then an application back because they missed the work. The technical problems and the data sets are real, but they are not the whole pitch. “You kind of can’t actually beat the level of difficulty,” Boyce says, because the problems are also societal, personal, and political, inside what is by many definitions the largest organization in the world. People come back because they are less bored. For anyone watching, the public doors are dhs.gov/ai/join, USAJobs alerts on AI and machine learning, and ai.gov. For small firms, he adds that the government buys services as well as products, small-business paths are often easier than prime-contractor ones, and he is willing to be found on LinkedIn.