Veterans are already asking AI assistants about their VA claims, and the assistants are already answering. Some of those answers are useful. Some are confidently wrong in a way that costs money. This guide is about telling the two apart.
Most of it applies whether or not you ever connect anything.
Why AI gets VA questions wrong#
The dangerous failure is not a model saying "I do not know." It is a model answering in the same steady voice whether it is right or wrong, with nothing in the answer to tell you which you got.
Combined rating math is the cleanest example. VA ratings do not add. Under 38 CFR 4.25, VA combines them in severity order using a published table, applying each next rating to the efficiency you have left rather than to the whole person. A 50 percent and a 30 percent do not make 80. They combine to 65, and VA rounds that to the nearest 10, which is 70. Ask a model to do it from memory and it may add, multiply, or recall the table imperfectly, and each looks exactly like an answer. Our VA math guide walks the procedure, and the combined rating calculator runs it against the actual table.
The second failure is time. Criteria get rewritten, rates move every December with the cost of living adjustment, and presumptive lists expand. A model trained two years ago answers a rate question with a two year old rate and never mentions the year. That is not a hallucination. It is a stale fact delivered as a current one, which is harder to catch.
The third is jurisdiction. A model can tell you what a criterion says. It cannot tell you what a rater will do with your file, because that turns on evidence it has never seen. An assistant that slides from describing the rule to predicting your outcome has stopped being useful.
What AI is actually good for on a claim#
The honest answer is not "avoid it." A specific set of things it does well:
Understanding what a rating criterion requires. The rating schedule is written for raters, not for you. Asking for a criterion restated in plain language, tier by tier, is a good use of an assistant, and it is checkable: open the regulation and confirm.
Finding the governing regulation. Knowing which diagnostic code covers your condition, and which section governs your question, saves hours. Verify the citation before you rely on it.
Organizing your own records. Sorting what you have, building a timeline out of scattered dates, naming the gap between your file and what the criterion asks for. That is clerical work, and none of it requires the model to know anything about VA. Our evidence guide covers what a file needs to contain.
Preparing questions before a claim exam. Walking in knowing what the examiner tends to ask, having thought about how to describe a bad day honestly, changes the exam you get. Our claim exam guide covers the room itself.
Translating a decision letter. Notices are dense and the reasoning is buried. A plain language pass over a document sitting in front of you is low risk, precisely because the source is right there to check. Our guide to reading a rating decision explains the structure.
Notice the pattern. Every good use is one where you can verify the answer against a document. Every bad use is one where you cannot.
What AI must never do on your claim#
Three hard lines.
It cannot replace a medical opinion. No model can examine you, and the link between your condition and your service is a medical judgment that has to come from a clinician who has. An assistant can help you understand what a medical opinion needs to address. It cannot supply one, and text that reads like one carries no weight with a rater.
It cannot be your representative. Only a VA accredited person may act for you in a claim. That is a legal status, not a capability, and no software has it. Accredited representation is available for free: our free claim help guide explains the three types and what each may lawfully charge.
It cannot be trusted on a number you have not verified. A rate, a combined rating, an effective date, a back pay figure. If a number will change what you do, check it against VA's published figures or a tool that reads the table.
If you are new to all of this, read VA Claims 101 first.
What a connector is, and what ours does#
Normally an assistant answers from memory: it recalls text it saw in training and produces something that resembles it. A connector changes the mechanics. It gives the assistant real tools it can call, so when you ask about a combined rating it runs the actual calculation on our servers and reports what came back instead of reconstructing it. The open standard behind this is the Model Context Protocol, and in practice it means our connector works with any MCP client that speaks Streamable HTTP.
That is narrow, and it is the entire point: it moves the answer from recall to lookup. The tools also instruct the assistant not to second guess what they return and not to substitute its own arithmetic, because a model that helpfully recomputes a table value has reintroduced the exact error the tool existed to prevent.
The ten free tools#
These ten need no account, no payment, and no credential. They are the regulation and math side of the platform.
| Tool | What it answers |
|---|---|
calculate_combined_rating | If VA rates me 50, 30, and 10, what is my combined rating |
compare_rating_criteria | What the rating schedule requires at each tier for this diagnostic code |
analyze_rating_gap | Where my rating sits against the criteria, and what the next tier would need |
calculate_rolling_window | How many days of incapacitating episodes fall inside a 12 month window |
find_secondary_conditions | Conditions commonly linked to one I am already service connected for |
check_presumptive_eligibility | Is this a condition VA presumes is service connected, so I do not have to prove the link |
search_legal_authority | Which regulation or court decision governs this, with the citation |
prepare_cp_exam | What the examiner will ask, and what to be ready to describe |
lookup_compensation_rate | The current monthly payment at this rating with these dependents |
compute_retroactive_pay | Roughly what back pay this effective date implies |
None of these read anything about you. They take what you type and return what the regulation or the rate table says, and the math tools carry an explicit instruction not to recompute the result and to describe an estimate as an estimate. For more depth, see our presumptive conditions guide and our secondary conditions guide.
The five tools that need an account#
Five more tools read your own VeteranHQ data. All three of these must be true:
- A paid membership. An account that is signed in but not entitled sees exactly the same ten tools an anonymous request sees.
- Your explicit consent. You read a disclosure and acknowledge it before anything is enabled.
- A sign-in from your assistant. When it first reaches VeteranHQ it sends you to veteranhq.app to sign in and approve. No token to copy.
| Tool | What it returns |
|---|---|
get_veteran_summary | Your profile, ratings, and active claims |
get_va_claim_status | Live VA.gov claim data, if you synced it with our browser extension |
get_claim_detail | One claim in full: evidence items, milestones, and status |
list_veteran_documents | Titles, types, and AI summaries of your uploads, 50 per call, with the true total |
get_document_content | The stored text of one document, 16,000 characters per page, only with the document-text permission |
All five are read only. Not one tool on the connector can write, change, upload, submit, or file anything.
What leaves, and what stays#
This is the same disclosure you acknowledge in Settings and see on the setup page.
What leaves, when an authenticated entitled client asks:
- Name, email, and service history: branch, dates, MOS, deployments
- Combined rating, and each condition's rating, diagnostic code, effective date
- Claim status, claim details, milestones, evidence lists
- Denied and deferred conditions, and any appeals
- VA payment history, including dates and amounts
- Dependents
- Document titles, types, and the AI-written summaries of what is in them, including files in your Medical Vault
- Stored text from documents on your account, including Medical Vault documents, only if you allow document text when your assistant signs in
What stays:
- The uploaded files themselves: stored text can leave when you allow it, the file never does, and stored text may not contain everything in the original
- Nothing can be changed. The connector can only read, never write.
The disclosure is blunt about the part we do not control: "Once your data reaches your AI client, it is in that conversation and under your agreement with that AI provider. We cannot delete it there, we cannot see how it is used, and we cannot promise it is safe."
Turning the connector off blocks new connector requests. A request already in progress may finish, and it does not reach into a conversation you already had elsewhere.
How to connect#
Free path: point your AI client at the connector endpoint with no credential, and you get the ten tools above. The endpoint and current setup steps are on our AI Connector setup page.
Paid path, if you want the five account tools:
- Open Settings, then AI Connector.
- Turn the connector on and read the consent disclosure.
- Add the connector address in your AI client.
- Sign in when it asks. Your client sends you to veteranhq.app, you approve there, and it holds a short-lived credential you never see.
- After you approve, open the tools menu in your chat and make sure VeteranHQ is switched on. Claude only uses connectors that are enabled for that chat.
- Claude asks before it uses each tool. To answer once for all of them, open Customize, then Connectors, choose VeteranHQ, and under Tool permissions set the read-only tools to Always allow. Every VeteranHQ tool is read only, so that one setting covers all of them.
Approving is not the same as enabling. Step 4 grants the connector on your account; steps 5 and 6 are settings inside your assistant, and skipping them is the most common reason a connection that looked successful ends with the assistant saying it has no VeteranHQ tools.
If it did not work
The sign-in tab showed your dashboard instead of the permission screen. Go back to your assistant and choose Connect again. That happens when your veteranhq.app session had already expired; signing in fixes the session, and the next Connect lands on the permission screen.
Your assistant says it has no VeteranHQ tools. Open the tools menu in that chat, switch VeteranHQ on, and ask again. Connectors are enabled per conversation, so a chat you opened before you connected will not have it.
Claude keeps asking permission before each tool. Set the read-only tools to Always allow under Customize, then Connectors, then VeteranHQ, then Tool permissions. Answering "Allow always" on the prompt itself works too, but that choice only lasts for that chat.
Two things to know, both in the disclosure's own wording. "Your approval lasts 90 days, then every assistant you connected stops until you approve again": the clock is on the approval, not on a key you hold. And "Turning it off blocks new connector requests. Requests already in progress may finish, and it cannot take back information already sent": one confirmation in Settings, no waiting period.
Honest limits#
The connector is read only by design, not as a stage we are passing through. It cannot file a claim, submit a form, upload a document, or change a field in your record or ours. Tools that could were deliberately left off. Reading the stored text of a document is the one thing that was added, and it is fenced behind its own permission: it is off unless you switch it on when your assistant signs in, it returns one document at a time, and it never returns the uploaded file.
It also cannot decide anything. Only VA decides ratings, effective dates, and payment amounts. Every forward looking number is an estimate resting on stated assumptions, and it says so. Treat it as a planning figure.
VeteranHQ is independent software. It is not part of, endorsed by, or affiliated with the Department of Veterans Affairs.
And none of this is required to file a good claim. Every regulation these tools cite is free at the eCFR, every rate is published on VA.gov, and an accredited VSO representative will help you at no cost.
Where to go next#
If you are early, start with VA Claims 101. If a rating number brought you here, read the VA math guide. If you have an exam coming, read the claim exam guide. If a decision has landed, how to read a rating decision explains what VA decided and why.
VeteranHQ explains these programs and helps you organize your evidence. We never prepare or file claims for you, and we are not your representative before VA. Nothing here is legal or medical advice, and only VA can decide a rating, an effective date, or a payment amount.