America.gov answers now and acts in 2027. That is the right order for your agents too.

America.gov launched as a reader. It answers questions and sends you to the right agency page. It does not file anything for you yet.

The actions come later. Coverage from CNN, Nextgov and Bloomberg describes a 2027 roadmap: applying for passports with photo upload and status tracking, enrolling in Social Security, Medicare and Medicaid, changing an address, reserving a federal campsite. The White House fact sheet says some of it will arrive "later this year." Either way, the order is clear. Read first. Act later.

Whatever you think of the rest of the launch, that sequence is right.

Why read first is right

An agent that reads can be wrong and nobody is harmed. You check its answer, you ignore it, you move on.

An agent that acts can be wrong and something happens. An email goes out. A record changes. A payment moves. Some of those cannot be undone.

The way to earn the right to act is to be right about reading for long enough that people trust you. That is how we promote people. It is how we should promote agents.

The rule we use at Sneeze It

Our AI agents work inside our CRM, our automation tools and our accounting system. Here is the rule, which is written into our operating instructions.

Read by default. Every agent can read. None can write unless a specific process has been granted write access by name.
Writes are gated in code. Our CRM wrapper refuses any write unless the process is on an approved list. If an agent tries, the call fails with a clear error.
Deletes are impossible by design. The CRM key has no delete permission, AI deletes are turned off in the CRM admin settings, and the wrapper refuses any command with "delete" in it. Three layers, so no single mistake loses client data.
Automation tools are read only. Agents can look at our Make and Zapier workflows and report on them. They cannot change or run them.

It is not glamorous. It has also never cost us a client record.

Privacy is the other half of read first

America.gov made a second promise at launch. Its privacy notice says the AI providers do not retain your prompts or responses, that answers are cached for up to two hours using a hash rather than your prompt text, and that it warns you before you share personal information.

Critics fairly asked how those promises will hold once people are filing real applications. That is the right question, and it applies to your company too. Ask it now, while your agents are still readers.

What are your people pasting into AI tools? Client lists, contracts, health details, passwords?
Which of your AI tools keep prompts, and for how long? Most business plans have a setting. Most companies have never looked.
Who owns that answer? One name.

The promotion path for an agent

Read. It answers questions and writes reports. A person acts on them.
Draft. It prepares the action, a draft email or a proposed change, and a person approves it.
Act within limits. It acts on its own inside a narrow, written boundary, with a log a person reviews every week.

Each step is earned with a track record, not granted on a hunch. At OTP we put that track record where everyone can see it: an agent holds a seat on the org chart, has a number on the scorecard, and logs its work, so the weekly meeting can see whether it earned the next step.

What to do this week

List every AI tool that can write, send or change something in your company.
For each one, ask whether it needs to. If not, make it read only.
Check the prompt retention setting on the two AI tools your team uses most.

Frequently asked questions

Can you apply for a passport on America.gov?

Not yet. At launch on September 29, 2026, America.gov answers questions and routes people to agency sites. Passport applications and enrollment in programs like Medicare are planned for a later phase, described in coverage as 2027.

Does America.gov store what you type?

Its privacy notice says AI providers do not retain prompts or responses, and that responses are cached for up to two hours using a hash rather than the prompt text.

How should a company give AI agents permission to act?

Start read only, move to drafting actions for human approval, and grant limited autonomous action only after a track record, with writes gated in code and deletes disabled entirely.

Sources

Ask your AI who owns what

OTP is readable from any MCP client. In Claude Desktop, Cursor or any MCP client, add this block:

"otp": {
  "command": "npx",
  "args": ["-y", "@orgtp/mcp-server"]
}

Restart the client. Then ask: "Use OTP to show me Sneeze It's chart. Who owns each seat, and what number is each one measured on?"

Start free at orgtp.com. Every seat is free, for people and agents. You pay only for the AI you use.

The series

All ten on one page: What America.gov teaches business about its knowledge problem.

Series: What America.gov teaches business about its knowledge problem. Part 9 of 10. Read this post as markdown: /blog/read-first-act-later-ai-agents.md.