Your New Coworker Has an Email Address. Now Hand It the Case File.
Why AI coworkers still need domain rules, connected documents, and human sign-off to handle regulated trade.
On October 8, 2026, Google introduced a new Gemini agent for work at its Gemini at Work event. One announcement stood out to us: the coworker agent.
A coworker agent has its own Google Workspace account, email address, calendar, and Drive. It works under its own identity, sees what the team shares with it, and leaves an audit trail of its actions. It can connect to enterprise systems and tools, including through MCP.
The AI agent that works like a member of the team is becoming a product category. Google is bringing it into Workspace, and other enterprise software providers are investing in agent identity and governance.
For companies building domain-specific AI, this is good news. As the infrastructure for running agents becomes standard, attention moves to the work itself.
An agent may have an email address and access to the right systems. Does it know how to do the job?
The plumbing is getting solved
Enterprise agents need identities, access controls, persistent context, tools, audit logs, and limits on what they can do or spend.
Until recently, teams building enterprise agents assembled most of this themselves. We did too.
Now large platforms ship it as standard. Google's announcement covered agent identity, governance controls, multi-model routing, and spend caps, with industry versions for financial services and legal in preview.
That leaves specialized companies free to focus on the knowledge, documents, and decisions that make a particular job hard.
Consider an export compliance desk.
Day one on an export compliance desk
Imagine a Korean company exporting packaged food or cosmetics to the United States.
A new AI coworker joins the trade compliance team. It has an inbox, a calendar, and access to the shared files. It can read emails, retrieve documents, and coordinate tasks.
What does it need to know before anyone trusts it with a shipment?
Start with a single regulation. For food imports subject to FDA prior notice, the notice generally must be submitted and confirmed before the shipment reaches the U.S. port of arrival. Under 21 CFR 1.279, the minimum lead time depends on the mode of transport: two hours by road, four hours by rail or air, and eight hours by water.
The filing window matters too. Through FDA's Prior Notice System Interface, prior notice can be filed no more than 15 calendar days before anticipated arrival. Through ABI/ACE/ITDS, the limit is 30 calendar days.
That is one rule, for one agency, in one part of the import process.
Then come the documents. The ingredient list on the label must be consistent with the product specification. A certificate of analysis may need to support specific quality or composition claims. The commercial invoice and packing list must agree on product identity, quantities, and packaging.
A classification decision made last year might still apply. Its reasoning may be buried in an email thread, and the product or the rule may have changed since.
A human employee learns all of this through experience: which document is authoritative, which decisions need escalation, which past approvals still hold. An AI coworker needs the same knowledge, in a form it can use.
What the job requires
When we break down an export compliance desk, three capabilities stand out.
1. The rulebook: knowing why a decision is valid
An agent working with regulated products has to identify the applicable rule, connect it to the product and shipment, and show the basis for its conclusion.
"This looks fine" gives a reviewer nothing to check. "Meets 21 CFR 1.279(a)(4): prior notice confirmed 11 hours before the vessel reaches port, against an 8-hour minimum" gives them everything they need.
The same goes for internal precedent. The agent should retrieve a relevant past decision, cite its source, and check whether the conditions behind it still hold. Regulations change. Formulations change. A decision that was right six months ago may need another look.
So the rulebook holds two things: current regulatory references, and a traceable history of how the company has applied them.
2. The product record: documents as a connected whole
The documents in a shipment describe different parts of the same commercial and regulatory transaction, and much of the meaning sits in how they relate.
The product specification describes what the product is. The certificate of analysis provides evidence about a batch. The invoice identifies what is being sold. The packing list describes how the shipment is organized.
Now suppose the specification changes after the commercial documents are already prepared. The agent should flag every document built on the old version, along with any earlier regulatory decision that relied on it. Those records may live in three different systems. The agent needs to know how they connect, which one is authoritative, and what has changed since the last approval.
This is the document intelligence we are building with VaultSage: scattered files turned into structured, connected context an agent can work from.
3. The sign-off line: knowing when a person decides
Regulated work requires accountability.
An agent can collect evidence, compare documents, flag inconsistencies, prepare submissions, and recommend next steps. When an error could mean a customs hold, a delayed shipment, or a regulatory violation, approval has to be explicit.
The agent shows what it checked, which sources it used, what remains uncertain, and which issues need human review. A person makes the final compliance decision, faster and better informed, with a clear record of how it was reached.
All three capabilities depend on the company's products, documents, obligations, and past decisions. They have to be built around the actual work.
Where NURIE fits
This is the direction behind RegSage, TraderSage, and VaultSage.
RegSage focuses on regulatory affairs for food, feed, cosmetics, health supplements, and pharmaceuticals. It helps teams assess market-specific requirements, check product information against applicable rules, and prepare regulatory documentation.
TraderSage focuses on trade documentation and logistics for cross-border transactions. It helps teams collect and verify shipment documents, flag inconsistencies, and prepare information for review.
Both run on VaultSage, our document intelligence infrastructure. VaultSage reads and structures complex business documents, preserves their context, and connects information across files.
VaultSage also provides an MCP server, so compatible external agents can reach that structured context through authorized integrations.
That matters as companies adopt AI coworkers from larger platforms. A company may choose Gemini, Microsoft, or another platform as the place where employees delegate work. Choosing one should not mean rebuilding the specialized knowledge behind the work.
We do not need to own the front door. We want the agent that walks through it to know the job.
The next question
The enterprise AI conversation has mostly been about which model or agent platform to adopt. As major vendors ship agents with enterprise identity and controls, a second question gets harder to avoid: what does an agent need to know to do a specific job, and how will you check its work?
For regulated trade, our answer is the three layers above. We are building them one workflow at a time.
If your team handles regulatory affairs or trade documentation for regulated products, we'd like to hear how you work today: how you verify documents, track regulatory decisions, and handle approvals across shipments.
FAQ
- What is an AI coworker agent? A persistent AI agent that works as part of an organization, with a defined role and its own identity. Google's coworker agent model includes a dedicated Workspace account, email address, calendar, and storage, and operates under the organization's access and governance controls. Google introduced it at Gemini at Work on October 8, 2026.
- Can a general-purpose AI agent handle regulatory affairs or export compliance? A general-purpose agent can support communication, research, document retrieval, and coordination. Regulated work also requires current market-specific rules, product-level documentation, traceable decisions, and human review. Specialized systems and domain knowledge provide that context.
- What are the FDA prior notice requirements for imported food? Under 21 CFR 1.279, for food imports subject to the standard prior notice timing requirements, FDA must receive and confirm prior notice at least two hours before arrival by road, four hours by rail or air, and eight hours by water. Prior notice can generally be filed up to 15 calendar days before anticipated arrival through FDA's Prior Notice System Interface, or up to 30 calendar days through ABI/ACE/ITDS. Different requirements apply to international mail.
- How do RegSage and TraderSage work with agents from other platforms? Both are built on VaultSage, NURIE's document intelligence infrastructure. VaultSage provides an MCP server that can make structured document context available to compatible external agents through authorized integrations. Specific interoperability depends on the external platform and integration configuration.
- Does the AI make the final compliance decision? No. AI prepares, verifies, and flags information, and an authorized person gives final approval. The agent provides supporting evidence and traceable references so reviewers can verify its work.
TechCrunch, Oct 8, 2026: https://techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses
Futurum, Oct 9, 2026: https://futurumgroup.com/insights/google-collapses-enterprise-ai-into-a-single-gemini-agent-at-gemini-at-work-2026/
eCFR, 21 CFR 1.279: https://www.ecfr.gov/current/title-21/section-1.279