Key Takeaways

  • Federal employee retirement processing dropped from six months to six seconds after Gebbia's National Design Studio moved the workflow online.
  • Replacing a lost Social Security card currently requires 87 separate clicks across disconnected government portals.
  • Americans spend roughly 10 billion hours every year trying to complete administrative tasks across federal agencies.
  • America.gov runs on an LLM restricted to official government publications, deliberately avoiding general web scraping to eliminate hallucinations.
  • The project shifts federal portals from passive reading directories into action layers that trigger transactions via agency APIs.

The 87-Click Reality of Federal Services

Gebbia pointed to a single metric that explains why public infrastructure feels broken: replacing a lost physical Social Security card currently requires 87 clicks across confusing legacy pages. It is an administrative maze that explains why citizens lose massive amounts of productive time to basic life admin. As Gebbia noted, “There's a stat published from the government that Americans spend 10 billion with a B hours every year trying to get things done in the government.”

Most software startups obsess over shaving one or two clicks off a checkout flow. Meanwhile, public services have accumulated decades of technical debt, layer upon layer of agency portals, and redundant paper forms. Gebbia's first major proof of concept inside the federal government tackled internal retirement processing. The legacy system took half a year to process paper records. Gebbia and the National Design Studio digitized the pipeline: “We took the whole process online. It was a huge win for the admin. Went from 6 months to about 6 seconds to retire.”

Why Search Without Action Fails

The standard approach to public digital services has always been building directories. An agency creates a website, uploads hundreds of PDF manuals, and expects citizens to decipher bureaucratic language. Gebbia realized that finding an answer is only half the problem. People do not visit public service portals to read policy; they visit to get a specific task done.

“What if there was just one website where people could go to search for the information that they need and also action it?” Gebbia asked.

America.gov is built to merge search with execution. Instead of forwarding users to an obscure regional portal to download a printable form, the interface connects directly to underlying agency APIs. You ask the system how to resolve a specific issue, and the software executes the transaction right in the interface.

Constrained AI as an Operational Standard

Deploying an AI interface for federal administration creates an obvious risk: hallucination. If a consumer chatbot invents a tax rule or fabricates an eligibility cutoff, users suffer real legal or financial harm. Gebbia bypassed this failure mode by stripping out the standard internet training corpus entirely.

“Because we're in the government, our LLM is only official government sources which means that we only return information that's published from the government. We don't scrape the internet,” Gebbia explained.

By locking the model strictly to published federal documentation, the system removes the creative guessing common in generalized large language models. The software acts as a direct translator between statutory rules and user intent, converting conversational prompts into validated API calls.

What to Do With This

Audit your core product workflows for artificial handoffs this week. Open your app, count every click required to complete your primary value transaction, and flag every step where you force customers to read documentation instead of completing the action for them. If completing a core action takes more than five clicks, replace the intermediate forms with direct API calls that execute the task in the background.