STAYCHARTED BLOG · Government · Federal agencies
Federal agencies publish the codes. Let the public check their answer against your own decisions.
Federal agencies publish codes the public must apply alone. Train one AI model on your own rulings and let anyone check their code, with a confidence.
Who applies federal codes, and how often they get it wrong
Some examples of codes the public applies on its own:
- Import tariff numbers. An importer declares a 10-digit HTS number for every product and is held to "reasonable care" in choosing it. U.S. Customs and Border Protection publishes its rulings in the CROSS database, searchable by keyword.
- Industry codes. Businesses may select NAICS codes for registrations based on their primary activity, while agencies also assign codes for their own programs. The Census Bureau uses NAICS to classify establishments for economic statistics; follow the requirements of the particular registration or program.
- Export codes, product codes for regulated goods, occupation codes, complaint categories. Each has a manual, a search box and the same gap between the two.
When a code is wrong, everyone pays for it: the filer in corrections, delays or penalties, and the agency in staff time, rework and statistics built on inconsistent codes.
What the public has today
A keyword search over the code list, a manual, published past decisions to read through, a help line, and for some programs a formal ruling that takes weeks. These work for the people who already know the standard. The importer of a first product, the founder registering a business and the citizen filing a complaint are left to guess, or to pay someone to guess for them.
One model, trained on your own decisions
The agency already has the best training data that exists: its own past decisions. Rulings, reviewed filings, corrected registrations or classified complaints, each a description and the code the agency's staff decided.
StayCharted AI Model Trainer learns from that record. It checks the examples first (the same description under two codes, codes with too few examples to learn), then tests the model on decisions it never saw, with a held-out test. For a time-based check, keep newer decisions in a separate test file and compare predictions with those known answers; this tests what comes next against what was decided before. The report shows how often it's right for each code, which codes it confuses, and what share of answers are right above each confidence.
The agency publishes the model when its staff are satisfied with the report, and publishes the next version when the report on newer decisions is better.
How the public checks a code
Three ways, depending on who is asking:
| Who | How |
|---|---|
| Anyone, from the agency's website | A classification form built by the agency on the agency's own page: the visitor describes the product or business, the agency’s server sends the text to the model through the Prediction API, and shows the suggested code with its confidence. The agency's web team owns the page and its wording |
| Groups the agency works with | A trade association, a port's brokers or a program's applicants get an access code for the StayCharted AMT phone app. One code can serve up to 500 phones, with its own daily limit, expiry, and pause or revoke at any time |
| Agency staff | Filling a spreadsheet of filings or cases, or asking from Power Automate, Google Sheets or an AI assistant |
The agency chooses what a visitor sees below the cutoff: "We're not sure. Here is how to ask us," or a link to the formal ruling process. A confident answer can be shown as guidance with the agency's own disclaimer.
The public form is an integration your web team builds, not a hosted public StayCharted page. Keep API credentials on the server and add input limits and rate controls. API and mobile answers below the model’s cutoff enter the Review Queue on Essentials and Business. Access codes also require the Owner’s permission; all use counts against the owning workspace’s allowances. Agree capacity for a public service with us.
Reviewing a case improves the next model version; it does not automatically notify the visitor or send them a revised answer. Your website must provide any follow-up process.
What the agency gets
Consistent answers, from the agency's own practice. Every visitor is answered by the same model, trained on the same decisions. The same published model applies across channels. Different descriptions can still produce different suggestions.
Hard cases in one queue. Answers below the confidence cutoff go to the agency's Review Queue, marked with the channel they came from. Staff see where the public is getting stuck: which products, which codes, which wording.
A model that improves as staff decide. Each case staff decide in the Review Queue becomes a training example for the next version. Publishing the new version updates every channel at once, and the published version keeps answering while the next one trains.
Review and activity records. Review items retain their confidence and source channel. Activity records training, publication and other changes; agree the retention and audit requirements for your program before deployment.
Where it fits
Good first programs have four things: a published code list, a record of past decisions with the text that led to them, a public that applies the code itself, and staff who already answer classification questions by phone or email. For example:
- Import classification guidance from published rulings.
- Industry code guidance for registration and contracting systems, from registrations staff have reviewed.
- Product codes for regulated goods entered on import or premarket filings.
- Public complaint and report categories, so a report lands with the right office the first time.
- Occupation or activity codes on applications and surveys.
How accurate is it? Two public benchmarks on federal data
We've tested StayCharted on two sets of decisions federal agencies publish:
| StayChartedAI Model Trainer · model comparison | ||
|---|---|---|
| Benchmark | Test | Result |
| CBP classification rulings | 1,167 rulings from Dec 2025 to Aug 2026, from the one-line subject, after training on 5,621 earlier rulings | Exact 10-digit HTS code: 22% to 25%. Right chapter: up to 74%. StayCharted AI Classifier Model at confidence ≥0.90: 6.7% coverage, 96.2% exact-code accuracy |
| NHTSA vehicle complaints | 5,000 complaints from 2025 onward, filed by subsystem from the narrative, after training on earlier complaints | 85.2% filed where NHTSA filed them · StayCharted Dedicated AI Model across 70 subsystems. At 0.90 confidence or more: 70.9% of complaints, 94.4% right |
The CBP result shows why the confidence matters. An exact 10-digit code from one line of text is hard for any model. The StayCharted AI Classifier Model reached 96.2% exact-code accuracy on the 6.7% of rulings at confidence 0.90 or higher. The remaining 93.3% would need review at that cutoff; human-review outcomes were not measured. More detailed records may help, but test that on the agency’s own cases.
What it isn't
- Not a ruling or a determination. The model suggests a code from past decisions. It doesn't apply the law, read the notes of the code list or replace the agency's formal processes. The agency says so on its page, in its own words.
- Not a replacement for staff. It sends the hard cases to them, and their decisions train it.
- Not a judge of edge cases. Missing facts can cause wrong answers, including confident ones. Staff must check descriptions that lack the facts needed to choose a code.
Security and procurement
- A database schema of the agency's own, holding only its organization's data, with each workspace walled off.
- Encrypted at rest and in transit, hosted on AWS in the United States.
- Used only for the agency's models. Its data never trains anyone else's.
- Its own cloud, if preferred. On the Business plan, models can be trained in the agency's own Amazon Bedrock or Microsoft Foundry account.
- Bought on AWS Marketplace, on the agency's AWS invoice, or directly.
StayCharted is not FedRAMP or GovRAMP (formerly StateRAMP) authorized. Marketplace availability and training in your cloud do not provide those authorizations. Data still passes through StayCharted for the cloud-training workflow. Before a pilot, the agency should approve the data, privacy notice, retention and security arrangements; public submissions can be stored as predictions or review items.
A pilot in 60 days
- Weeks 1 to 2: choose one program and export a few thousand recent decisions: the description and the code. Remove names, addresses and anything confidential.
- Weeks 2 to 3: train, read the report per code with program staff, and set the cutoff.
- Weeks 3 to 6: staff test incoming cases. Review spreadsheet results in the returned file; use the API or phone app for below-cutoff answers to enter the Review Queue.
- Weeks 6 to 8: open it to one outside group with an access code, or as guidance on one page of the agency's site. Compare the report on the newest decisions with the first one.
Common questions
Can a government agency let the public check a classification code with AI?
Yes. The agency trains a model on its own past decisions and shows its suggestions, with a confidence, on its own website or through a phone app. Unsure answers point to the agency's normal process.
Is the model's answer an official determination?
No. It's guidance from past decisions. The agency's formal processes, such as binding rulings, stay as they are.
What data does the model learn from?
Decisions the agency has already made: a description and the code staff decided. It learns only from the agency's data, and that data trains no one else's model.
How does the agency know whether it's accurate enough?
Before publishing, the model is tested on decisions it never saw. The report shows accuracy per code and what share of answers are right above each confidence, so the agency sets the cutoff from evidence.
What happens to questions the model is unsure of?
API and mobile answers below the model’s cutoff enter the agency’s Review Queue on Essentials and Business. Staff decisions can become training examples for a later version. The visitor is not automatically notified; the agency must provide its own follow-up process.
Is StayCharted FedRAMP authorized?
No. Talk to us about what your program needs before a pilot, including which data the model would hold.
Can we run it for one program first?
Yes. One program, one code list and one group of outside users is the usual start.