AI Agents
Missionforce: Policy Engine, Operations, and Field Operations for government, explained
The day after the Dreamforce keynote, Salesforce expanded Missionforce, its government platform, with three products: a Policy Engine that turns approved policy documents into deterministic rules, an Operations product for procurement and invoice work in air-gapped environments, and Field Operations with offline mobile. This is what each does, where OpenAI and NVIDIA fit, what is generally available, and the decision that stays with a person.
A government agency’s hardest automation problem is not the technology. Every rule the automation applies was written by a person, approved by a committee, and can be challenged in a hearing. An agent that improvises a policy interpretation is a liability with a chat interface, not a productivity gain.
Salesforce is aiming at that problem with the Missionforce expansion it announced on September 16. Three products, two model partnerships, and one design choice that runs through all of them: the model drafts, the rules are deterministic, and a person approves. This post covers what shipped, what is dated, and where the line between the agent and the official sits.
What Missionforce is
Missionforce is Salesforce’s government-focused platform, run by Kendall Collins as CEO of Missionforce and Government Cloud. Salesforce describes it as letting agencies develop, customise and deploy tailored AI while keeping operational control inside secure environments.
Missionforce sits on Government Cloud, which is not commercial Salesforce with a compliance sticker. I covered the constraints in the public sector post: a deliberately short model menu, features that trail the commercial release, and one decision the agent never gets to make. Missionforce inherits all three, and the September announcement is best read against them.
Policy Engine
The Missionforce Policy Engine takes approved policy documents and turns them into structured, deterministic rules, using OpenAI’s generative models to do the reading and drafting. The output is an auditable workflow that runs the rule the same way every time.
The important word is deterministic. A language model reads the policy and proposes the rules. The rules that run are not the model.
They are a fixed structure a reviewer can read, test and sign off, in the same way Agent Script lets an Agentforce builder write the parts of an agent that must not vary. When a citizen asks why a decision was made, the answer is a rule with a document behind it, not a probability.
Where the person stays is at the approval of the rule set. A model can misread a clause, and a rule derived from a misread clause will be wrong consistently, which is worse than wrong occasionally.
The review board that approves the policy has to approve the rules extracted from it. Salesforce’s announcement describes the output as auditable, and the audit trail only helps if someone reads it before the rules go live.
Operations
Missionforce Operations converts manual back-office processes, procurement, supplier management and invoice audits are the examples given, into digital workflows with agents orchestrating the tasks. Salesforce says it runs in air-gapped environments, and the Koa press release lists Missionforce Operations as generally available now in the US.
Air-gapped is the detail that separates this from the commercial supply-chain agent. An environment with no connection to a public model provider cannot call a hosted frontier model at all. That is why the NVIDIA part of the announcement matters, and I come to it below.
Invoice audit is the process I would start with, because it has the shape agents handle well: a document, a set of rules, a match or a mismatch, and a human who reviews the exceptions. The agent assembles the case. The auditor decides. The same pattern I described for accounts payable in the private sector applies, with a stricter audit trail.
Field Operations and Asset Management
The third product automates scheduling, work orders and asset maintenance for field teams, with offline mobile for staff working where there is no signal. Field Operations is the government version of Agentforce for Field Service, and the questions are the same. Which scheduling decisions can the agent make alone, which does a dispatcher confirm, and what happens to a work order created offline when the device reconnects and the asset’s status has changed.
Salesforce has not published an availability date for Field Operations in the announcement. Treat it as announced rather than shipped until the documentation says otherwise.
Where OpenAI and NVIDIA fit
Two partnerships run through the announcement, and they answer different constraints.
OpenAI’s frontier models reach Government Cloud through Amazon Bedrock, which is the same Bedrock lane Salesforce opened for commercial Agentforce customers on the same day. The announcement also describes ChatGPT access that lets government staff retrieve secure data and trigger Missionforce actions, which is Missionforce’s version of the AIforce idea: the work happens where the person already is, under the permissions they already hold.
NVIDIA answers the air-gapped case. The partnership brings accelerated computing and models so an agency can train, tune and deploy mission-specific models on its own infrastructure, with post-trained NVIDIA models reaching select customers in October 2026. An agency that cannot send a prompt outside its network can still run a capable model inside it.
Nothing else makes Operations work in a disconnected environment, and this is the part of the announcement with the most engineering behind it.
The reference deployment
The TSA is the customer Salesforce put in front of the announcement. Its agent, Ace, was built with Public Sector Solutions on Government Cloud and handles routine traveller questions over email, phone and social channels. Salesforce’s press release says Ace has handled around 100,000 conversations a month since launching in summer 2026 and resolves 96% of routine questions, such as packing rules and medical device procedures, without escalation. The projected saving is over 11,660 staff hours a year and $11 million in labour over three years.
Those are the TSA’s and Salesforce’s figures, and they describe a specific kind of work: high-volume questions with a published answer. Nothing in Ace decides who gets through security. Copy that pattern. Start where the answer is already written down, and keep the decisions with consequences where they were.
What is available, and where
Missionforce Operations is generally available now in the US. Post-trained NVIDIA models reach select customers in October. The Policy Engine and Field Operations are announced without a published date. Salesforce’s standard line applies to all of it: pricing and packaging are subject to change, and availability varies by region and is governed by customer agreements.
Outside the US, Government Cloud availability and the FedRAMP framing do not translate directly. A UK, EU or Indian public body should read this as a signal of where Salesforce is investing, not as a product it can buy this quarter, and should ask the account team which parts, if any, are certified for its jurisdiction.
What I would do with it
For an agency or a contractor building for one, three steps in order.
Start with the Ace-shaped problem. Find the queue of questions with published answers, put an agent on it, and measure resolution against the published rule. The number will be high and it will be defensible.
Pilot the Policy Engine on one policy with the review board in the room. Extract the rules, compare them line by line to the document, and count the corrections. That count is the evidence for how much review every future policy needs.
Use Operations for invoice audit before procurement. The exception-review pattern is proven, the audit trail is native, and the people who own the process already know what a mismatch looks like. Save supplier management, where the agent’s actions touch outside parties, for after the first two have run for a quarter.
Understanding the basics
What is Missionforce?
Salesforce’s platform for government agencies, built on Government Cloud, that lets agencies develop and deploy AI agents within secure environments. On September 16, 2026 Salesforce expanded it with a Policy Engine, an Operations product and Field Operations and Asset Management, alongside OpenAI and NVIDIA partnerships.
What does the Missionforce Policy Engine do?
It uses OpenAI’s generative models to read approved policy documents and turn them into structured, deterministic rules that run as auditable workflows. The model drafts the rules. The rules themselves are fixed, reviewable and consistent, and a person approves them before they apply.
Is Missionforce available now?
Missionforce Operations is generally available in the US. Post-trained NVIDIA models reach select customers in October 2026. The Policy Engine and Field Operations were announced without a published date, and Salesforce says availability varies by region.
Does Missionforce use OpenAI?
Yes. OpenAI’s frontier models reach Government Cloud through Amazon Bedrock, and the announcement describes ChatGPT access for retrieving secure data and triggering Missionforce actions. NVIDIA models cover the air-gapped case where no external model call is possible.
The rule is the product
Every commercial agent platform sells reasoning. Missionforce, read carefully, sells the opposite: a way to take the reasoning out of the decision and leave it only in the drafting. Government work needs that shape, because a decision has to be the same for the next citizen as it was for the last one and has to survive a hearing.
If the Policy Engine delivers rules a reviewer can read and sign, it will matter more than any model behind it. If it does not, no model will make up the difference.
If you build for a public body and want the Ace-shaped queue identified and the policy review process designed before an agent touches a rule, talk to us, or start with the Agentforce readiness assessment.
