Intake agent
- Extracts taxpayer data from source forms (W-2, 1099, K-1) and pins each value to its bounding box on the page — so the number carries its provenance from the first hop.
- Datapoint provenance
AI agents for tax
The governed Tax AI that proves every number.
A 4-agent pipeline — intake → deduction-finder → return-reviewer → compliance-check — prepares, reviews, and compliance-checks returns. Unlike any horizontal copilot, every deduction and figure traces to its source form in a tamper-evident ledger you can hand to an auditor.
Aligned with IRS Circular 230 · SOC 2 · EU AI Act · high-risk: bias monitoring
Object Explorer
ontology ▸ tax · published · object · property · action permissions
The regulatory pressure
A preparer is responsible for every number — so the AI must be too.
IRS Circular 230 makes the practitioner accountable for diligence and accuracy; an AI that quietly hallucinates a deduction or paraphrases a figure is a liability, not a productivity tool. Cortex grounds every figure to the source form, scores accuracy against a gold eval suite, and seals each return so the number on the line can always be traced back to the box on the page.
IRS Circular 230EU AI Act · high-riskBias monitoringSOC 2ISO 42001 · AI records
Regulatory pressure
IRS Circular 230EU AI Act · high-riskBias monitoringSOC 2ISO 42001 · AI records
Typed objects
Restrictedlocked
frameworks ▸ encoded as policy · gated · sealed
Use cases
A 4-agent pipeline — each step grounded and gated.
Every agent is a registered, identity-bound actor with allowed models and actions; each hands off through the Action Fabric and seals its output into the ledger.
Deduction-finder agent
- Identifies eligible deductions and credits grounded to the extracted facts and IRS rules — citing the source figure for each, never inventing one.
- Grounded RAG
Return-reviewer agent
- Reviews the assembled return against the gold eval suite (faithfulness, citation coverage, safety) and flags discrepancies for a human in ≤ 30 minutes.
- Observability & evals
Compliance-check agent
- Runs the Circular 230 diligence checks and bias monitoring as Policy-as-Code, blocking sign-off if a figure can't be traced to its source.
- Policy-as-Code
Sealed, auditable return
- Every figure links to its source form, page, and bbox in a hash-chained ledger — click a number, see exactly where it came from, verify it wasn't altered.
- Trust Ledger
The published SLA — the promise we sell on
97% accuracy. 95% citation coverage. ≤ 30-minute review.
These aren't aspirations — they're the measured targets of the @cortex/pack-tax gold eval suite. Every figure cites its source form, the reviewer turns a return in under thirty minutes, and the whole thing ships in your own environment with the evidence pack attached.
- 97% accuracy measured against a gold evaluation suite — not a vibe check
- 95% citation coverage: every figure traces to its source form, page, and bbox
- ≤ 30-minute review turnaround with discrepancies pre-flagged
- EU AI Act high-risk posture: bias monitoring and human review built in
Object Explorer
ontology ▸ tax · published · object · property · action permissions
Prove it — don't just claim it
The verdicts you'll see in the demo.
Challenge the pipeline and it fails closed — an uncited number never makes it onto a signed return.
runtime · verdicts
Cited deduction200 cited
Uncited figure451 blocked
Altered numberhashOk:false
Sign-offHOLD · review
fail-closed ▸ challenge it and it denies, on the record
- 01
Model the domain
Map the work to typed ontology objects — Taxpayer, Return, Form, Deduction, Finding — with restricted fields like ssn / ein locked at the property level.
- 02
Encode the rules
Translate the obligations above into Policy-as-Code and Action Fabric approvals — the gates fail closed, returning a precise code, not a guess.
- 03
Prove the outcome
Every run lands in a hash-chained Trust Ledger with signed receipts and 10-hop provenance — a record you can hand the regulator.
97%Tax pack accuracy (gold eval suite)
95%Citation coverage on every figure
30-minReview turnaround target
4 agentsIntake → deduction → review → compliance
Works with the rest of the runtime
The controls behind tax.
The same shared runtime powers every vertical — here are the capabilities and the Solution Pack this industry composes from.
Tax Solution Pack
Install @cortex/pack-tax — signed, SLA-backed, and verifiable before import.
Trust Ledger
Datapoint provenance pins every figure to its source form, page, and bbox.
Observability
Faithfulness, citation coverage, and safety scored against the gold eval suite.
Policy-as-Code
Circular 230 diligence and bias monitoring encoded as enforced policy.
Ontology
Model Return, Form, and Deduction; lock SSN/EIN at the property level.
Trust Center
Map the pipeline to Circular 230, SOC 2, and EU AI Act high-risk controls.
Security & compliance
Built for the diligence Circular 230 puts on the practitioner.
The grounded pipeline, gold evals, sealed ledger, and bias monitoring map directly to the diligence, accuracy, and AI-management-record controls a tax practice reports against. Aligned with — never claiming certification you don't hold.
IRS Circular 230 · diligenceEU AI Act · high-riskBias monitoringSOC 2 · auditISO 42001 · AI recordsNIST AI RMF · Measure
FAQ
AI agents for tax — questions, answered.
How does Cortex Tax AI prove a number is correct?
Each figure is pinned to its source form, page, and bounding box at intake, carried through the pipeline with a value hash, and sealed into a hash-chained ledger. Click any number to see exactly where it came from, and verify it wasn't altered after extraction.
What is the Tax pack's published SLA?
97% accuracy measured against a gold evaluation suite, 95% citation coverage so every figure traces to its source form, and a review turnaround of 30 minutes or less.
How is the Tax AI a 4-agent pipeline?
Intake extracts and grounds the data, deduction-finder identifies eligible deductions, return-reviewer checks the assembled return against the eval suite, and compliance-check runs Circular 230 diligence and bias monitoring before sign-off.
Does this satisfy EU AI Act requirements for high-risk AI?
Cortex treats tax preparation as high-risk: it adds bias monitoring, mandatory human review, grounded outputs, and a sealed AI-management record. It is built for alignment with the EU AI Act — Cortex does not assert certification you don't hold.
More industries
One governed runtime, every regulated vertical.
The controls are the same — the obligations they satisfy differ. Explore another vertical, or see the full set.
Prove every number — then hand it to the auditor.
See the 4-agent pipeline, the published SLA, and the source-form provenance behind every figure on the return.