How AI is used
How AI is used, and where it is not
AI does the messy reading and labeling. It never decides a tax figure on its own, and it never touches your position.
Most of Accuren is deterministic: replay the chain, apply the rules, convert at the right rate. AI is used in the three places where a rule cannot reach — and each of them ends with a human confirming, not a model deciding.
#Where it is used
| Job | Why a rule cannot do it | What happens to the output |
|---|---|---|
| Labeling unknown transactions | New protocols appear faster than rules can be written | Flagged ai_labeled, listed for confirmation |
| Reading issuer filings | Prose, not data, and different at every issuer | Quoted with the clause, never summarised away |
| Explaining an event | Plain language, on demand | Presentation only — the figures come from the replay |
| Filtering alerts | Whether something matters depends on your position | Can only downgrade to digest, never suppress risk |
#Where it is not
- Cost basis arithmetic. Deterministic replay, exact decimals, no model in the path.
- Exchange rates. Published sources, looked up by date.
- Disposal or not. Rules first; anything unmatched is flagged, not assumed.
- Anything that moves funds. There is no such path — see Read-only by design.
An unconfirmed label never disappears into a total
Every export lists ai_labeled events individually in its warnings. If a figure depends on a guess, you find out before you file, not after.
#Confirming or correcting a label
curl -X PATCH https://api.accuren.xyz/v1/events/evt_9f21 \
-H "Authorization: Bearer $ACCUREN_API_KEY" \
-d '{ "disposal": false, "reason": "self_transfer", "confirmed": true }'Corrections are kept with the event and carried into every future export, so the same transaction is never re-guessed.
Why it is not sold as an AI product
The asset is the record: multiplier history, basis, and the right rate on the right day. AI is why the messy parts feel effortless — it is not the thing you are paying for.