The Army Shipped AI Agents to 9.2 Million People. You're Still in Committee.

The United States Army just put AI agents into production for 9.2 million people.

Your company is on its fourth pilot.

On August 5, Salesforce announced that its Missionforce National Security platform cleared Impact Level 5 authorization from the Defense Department. IL5 is the bar for handling controlled unclassified information and unclassified national security systems data. Army Human Resources Command is the first component live on it.

I spent six years in the Army National Guard. If you had told me then that the Army would beat the private sector to production on anything involving new software, I would have laughed.

It did.

The Numbers Are Not a Pilot

HRC supports 9.2 million soldiers, veterans, and military families. It runs more than 1,500 personnel cases a day and closes roughly 600,000 a year with about 3,000 analysts and HR professionals.

The agents answer routine inquiries around the clock, summarize case histories, surface policy and career information, and route complicated matters to people. Salesforce projects more than 55 million agent conversations a month at full scale and about $6 million in annual savings for HRC.

This is not a proof of concept in a sandbox. It is running against the personnel records of nine million Americans.

It did not start there either. The Air Force put Missionforce to work with the 441st Vehicle Support Chain Operations Squadron in early July. One unit, then the next.

They Cut Scope to Ship

Here is the detail everybody is skipping.

To earn the IL5 authorization, Salesforce turned off Anthropic's Claude capabilities inside the platform. A commercial feature that works fine on the outside got removed to clear the security bar on the inside.

That is a scope decision. Somebody read the certification requirements, looked at the feature list, and cut.

Most companies will not do this. They want the full feature set, the vendor's entire roadmap, and a clean security sign-off, so they sit in review for eleven months instead of shipping something narrower that clears the bar today.

The Army shipped the narrow thing.

Early reviewers also found the commercial platform handles routine, keyword-driven requests well and struggles with personalized recommendations. HRC deployed it anyway, scoped to the work it does well, with people keeping every decision that touches a soldier's career.

That is what disciplined implementation looks like. Not the best possible system. The right system for the task in front of you.

Guardrails Are Not the Brake

Retired Major General Allan Day, now at Salesforce, said it about as well as it can be said: "Guardrails, to me, built-in means I can accelerate."

Read that twice if you run technology at a regulated company.

Governance is not what slows AI down. The absence of governance is what slows AI down, because every deployment becomes a fresh argument with legal, security, and compliance that nobody has the authority to end.

Build the controls into the platform once and everything after it moves faster. That is Simple and Secure, and it is why the Defense Department can move on this while a mid-market bank cannot.

Your Excuses Are Weaker Than Theirs

Run the list.

"Our data is too sensitive." More sensitive than the personnel records of 9.2 million service members and their families?

"We are too regulated." More regulated than an IL5 authorization boundary inside the Department of Defense?

"Our processes are too complex." More complex than Army human resources?

"We cannot risk a wrong answer." Neither can HRC. So they scoped the agents to routine work and left judgment with people. That option was always available to you.

Every objection I hear from private sector leaders applies harder to the organization that just went live.

What I'd Do This Week

Pick the highest-volume, lowest-judgment workflow you own. Not the most interesting one. The one that eats hours and requires no discretion. That is where HRC started.

Write down what the agent is not allowed to decide. On paper, approved by whoever owns the outcome. That one document kills more delay than any tool evaluation.

Cut the feature that is blocking approval. If a single capability is holding up your security sign-off, ship without it. Add it back after the thing is live and earning.

Staff the human boundary. Routing to a person is only a control if the person exists and has time. Otherwise it is a queue with better branding.

The Part That Matters

The Army did not win this because it had better technology. It bought the same platform anybody with a purchase order can buy.

It won because somebody defined the mission, cut the scope to clear the bar, and put the thing in front of real users.

Prioritize and execute. That works in uniform and it works in your company.

You are not blocked by AI. You are blocked by the meeting.