Run an ITR end-to-end with the agentic firm.
This guide is for every Triarc team member preparing or reviewing an Indian Income Tax Return. Read it once before your first engagement — then keep the dashboard open in one tab and refer back here when something's unclear.
Overview
The Agentic ITR module is a three-agent pipeline — preparer, manager, partner — that turns a stack of client documents (Form 16, 26AS, AIS, broker statements, ULIP / MF holdings, bank statements, challans, residential-status sheet) into a Winman-uploadable ROI workbook with a partner-attested draft.
For a typical ITR-2 with non-resident status + VDA income, the full pipeline takes ~5 minutes and costs roughly ₹300 in Anthropic tokens. The deterministic engine still owns the math (slabs, rebates, surcharges, cess); the agents handle document comprehension, classification, and partner-grade judgement.
Your first engagement, step-by-step
Clients → + New Client — enter full name, PAN, mobile, residential status, and city. Date of birth is optional but useful for TRACES/AIS auto-unlock.
Dashboard → + New engagement — pick the client, pick the assessment year, leave the ITR form blank (the preparer will auto-route ITR-1 vs ITR-2 based on the docs). Optionally assign a manager and partner.
Open the engagement, drop every relevant file into the Documents accordion in the Workspace tab. The upload route auto-extracts text and runs OCR fallback for scans. Most TRACES PDFs (Form 16/16A/26AS) and incometax.gov.in PDFs (AIS/TIS) unlock automatically using PAN + DOB; demat / bank / broker statements with custom passwords show an "Unlock" prompt for you to fill in.
Click the primary Run preparer button. The Trace drawer opens on the right and streams every tool call — extract_26as, read_document, compute_regimes, raise_finding, etc. Each event is also persisted so you can replay later. The preparer typically takes 2–3 minutes and 11–22 tool calls.
When the preparer finishes, the run auto-chains into Manager review and then Partner approval (provided assignees allow it). On a stuck engagement, the chain stops at a human gate and the action button updates to "Run manager review" or "Run partner approval" — click to continue.
Open the Findings tab — every issue is severity-tagged with section references, amount-at-risk, and a suggested action. Then open the Output tab → Winman ROI workbook → click Download. Open in Winman, paste the Schedule VDA line if applicable, fill in the refund bank IFSC, and you're ready to file.
The three agents
Preparer
- Reads every uploaded document, classifies its type
- Extracts 26AS / AIS into structured TDS + capital-gains rows
- Builds the TaxpayerFacts blob (income heads, deductions, regime)
- Runs compute_regimes (both old + new) and recommends one
- Computes 234A/B/C interest from challan dates
- Raises every gap or risk as a structured finding with evidence
Manager
- Re-runs compute_regimes from scratch to verify the math
- Spot-checks every preparer finding against the docs
- Catches numerical errors (often saves ₹15–30 per engagement)
- Drafts v2 of the TaxpayerFacts with corrections
- Decides which findings to surface to the partner
Partner
- Reviews the manager's v2 draft and findings list
- Records explicit accepted_risk notes for filing-critical items
- Locks the regime choice
- Calls approve_draft to generate the Winman ROI workbook
- Sets engagement stage to 'approved'
Per-tab guide
Where every artefact for the engagement lives. Default landing tab.
- Documents — uploaded files, classification, OCR status
- Clarifications — questions the agents flagged for you (replied here resumes the run)
- Current draft — the latest TaxpayerFacts blob the agents are working with
- Timeline — every state transition, stamped with actor and timestamp
Live SSE log of agent tool calls + historical runs list.
- Runs history — every preparer / manager / partner / evaluator run with cost, duration, token count
- Live trace — streams events as they happen during an active run
Every issue the agents raised, with evidence pointers.
- Severity filter — Critical / High / Medium / Low / Info
- Status filter — Open / Resolved
- Each finding has: title, detail, suggested_action, evidence document IDs, amount_at_risk
- Manager + partner can resolve as 'fixed', 'accepted_risk', or 'not_an_issue'
The final deliverables — workbook download + evaluation history.
- Winman ROI workbook — signed 24-hour download URL, regenerate any time
- Quality evaluations — partner-level scoring of the engagement (optional)
Cost & quality — what to expect
A representative ITR-2 engagement (Non-Resident, light VDA, 7 documents) we ran cost ₹587 across 50 tool calls in 5 minutes 1 second. The partner attested it with two explicit accepted_risk notes (refund bank IFSC missing; VDA needs manual Schedule VDA in Winman). Quality score: 88/100.
Simpler ITR-1 returns (salary + interest, 3 docs) typically land at ₹200–250 and 90 seconds. Heavier ITR-2 returns (capital gains + multiple property + foreign assets) can run ₹400–700.
Troubleshooting
Open the document row → click Unlock → enter the customer-specific password (DEMAT, bank, broker, CAMS). TRACES + IT-portal documents auto-unlock via PAN + DOB; everything else needs a manual password.
The preparer still reads the file correctly via read_document so it doesn't break extraction. Override the type by clicking the dropdown next to the document and selecting the right kind. The classifier learns from your corrections.
The deterministic engine doesn't yet have a § 115BBH bucket for VDA income, so the Winman download reflects engine output only. The partner's accepted_risks block tells you the manual Schedule VDA + interest adjustment to make in Winman before filing. Tile = post-adjustment number, Winman download = pre-adjustment.
Always collect the full bank block from the client before submitting the ITR. The partner's accepted_risk note will record this as a blocker; operations should follow up via mobile / email before e-filing.
The dashboard label is fixed; the engagement form lets you pick a different AY via the API (use ?ay=2025-26 when accessing the New Engagement flow, or create the engagement programmatically through the workspace API).
Vision-guardian guardrails
A few rails are deliberately in place to keep us out of trouble while we ramp the product up:
The compute core supports all seven ITR forms, but only ITR-1 and ITR-2 have golden-test bench coverage today. Extended forms refuse engagement creation unless the firm explicitly opts in via TAX_EXTENDED_FORMS_ENABLED=true. The current threshold to flip that flag: 50+ ITR-1/2 engagements shipped without partner override.
The agents never file an ITR, never send a notice reply, never email a client, never post to the IT portal. Every action that leaves the firm goes through a human approval step. The 'Output' tab generates files for download; you push them outbound.
The IT-portal connector is built but the auto-pull flow is gated behind an operator approval step. For now you upload AIS/TIS manually from the e-filing portal.
The WhatsApp intake classifier identifies inbound messages (client query / document upload / acknowledgement) but never auto-replies. Every reply is routed to your inbox for review.
Per ITR-2 instructions, Non-Resident taxpayers are exempt from Schedule FA. The preparer flags this as an info finding so the auditor knows the omission is intentional.
Ready to try?
Pick a clean ITR-1 from your current workload, create the engagement, upload the docs, hit Fast-track. Compare the output to the manual review you'd otherwise do. If anything feels off, tag George and the team — every flag from the first 50 engagements goes straight into the next training cycle.