ReconPe — frequently asked questions

ReconPe is an AI-agent reconciliation platform with stateful memory across runs. It pairs ACRE — a deterministic 6-stage probabilistic matching engine — with two AI agent modes (Investigate and Ask) and a cross-run exception memory. Finance teams upload settlement, payment, and bank files; the AI does the first pass; humans review, approve, and close. ReconPe is launching in May 2026.

ACRE — Adaptive Cascade Reconciliation Engine — is ReconPe's deterministic matching pipeline. It runs six stages: data profiling, multi-level blocking (exact key, fuzzy, LSH nearest-neighbour), Bayesian Fellegi-Sunter scoring 0–100, Hungarian-algorithm N:M assignment, domain validation (conservation law, temporal ordering), and typed exception classification. ACRE is deterministic and reproducible — no LLM calls on the matching path.

AI handles four jobs around the deterministic matching engine: (1) auto-detects file format and proposes column mappings; (2) suggests matching rules from your data; (3) runs a second pass on unmatched records to surface POTENTIAL_MATCH exceptions; (4) auto-generates resolution suggestions for every exception, scores reconciliation-level financial risk, and writes a plain-language narrative. Humans review decisions; AI does the heavy preparation.

Anthropic Claude, OpenAI, DeepSeek, and Google Gemini are all supported and configurable per tenant. Buyers choose based on data residency, cost, or compliance policy. Vendors with hard-coded single-provider architectures cannot offer this flexibility.

Stateful reconciliation means the accumulated history of past decisions — exception dispositions, counterparty patterns, and weight adjustments — materially changes how the current run is computed. Most reconciliation tools are stateless: they re-match each cycle from scratch. ReconPe carries open exceptions, learned patterns, and rejected-candidate feedback across runs.

Every exception gets a deterministic fingerprint (counterparty + amount + direction). When today's run produces an exception that opposite-signs a prior-run exception, ReconPe surfaces a candidate: 'this looks like the settlement of exception #1832 from Feb 14.' One click closes both sides together. Manual link UI is also provided for cases the fingerprint misses.

After three confirmed settlements for a recurring counterparty, ReconPe tags the pattern (for example: 'late settler, average lag 5 days') and shows it as an amber banner on future exceptions for that counterparty. This lets analysts treat predictable late settlement as expected behaviour rather than re-investigating it every cycle.

Agent Investigate Mode is a fixed-pipeline root-cause analysis that inspects a reconciliation and narrates why exceptions were raised — streaming its reasoning trace via Server-Sent Events. Ask Agent Mode is a real ReAct planner loop with typed tool access to the reconciliation state (exception pool, pattern library, risk analytics, memory tools). Both run downstream of ACRE; neither makes the matching decision.

Auto-detection is live for seven Indian formats: Amazon India (deep audit including Settlement Report V2 GST), Flipkart (deep audit including the four-fee structure), Meesho (deep audit including price-base semantics), Razorpay (deep audit including UTR matching and MDR variance), Cashfree (detection only), PhonePe (detection only), and PayU (detection only). Deep audit on Cashfree, PhonePe, and PayU is on the roadmap.

Upload your Amazon India MTR (Merchant Tax Report) and bank statement. Iris auto-detects the Settlement Report V2 format. Argus proposes matching rules. ACRE runs probabilistic matching with a SUM_AGGREGATE pass for combined payouts. Commission variance is audited against your category rate card per order; COD remittance ageing is tracked from delivery; A-to-Z claim dispositions are surfaced. Returns and RTO entries are matched against original orders.

Upload your Flipkart settlement report and bank statement. ReconPe parses Flipkart's four-fee structure (commission, fixed, collection, shipping) as separate audit axes. Each fee category is reconciled against the configured rate card with per-order variance flagging. Returns and reversals are matched against original payouts.

Upload your Meesho Supplier Payments report and bank statement. ReconPe detects the price-base structure (supplier price vs customer price) and asks you to confirm once per supplier. Return-cycle reversals are netted against forward settlements. COD remittances are aged from delivery date.

Upload your Razorpay settlement export and bank statement. ACRE matches every settlement payout to its bank credit by UTR canonicalisation, applying fuzzy fallback when banks truncate or reformat the UTR. MDR (merchant discount rate) variance is computed against your Razorpay rate card per transaction. Refund-cycle reversals are matched to original payouts.

No. ReconPe's schema detection reads marketplace and bank exports as-is. Format changes that Amazon, Flipkart, and Meesho have pushed over the last three years are all supported. No column mapping required for the first run on standard formats.

FinanceOps brings the ACRE engine to the controller close. Subset-sum matching for sub-ledger to GL tie-out, Bayesian confidence per group, stateful memory across periods, and an audit trail Big Four reviewers will accept. Designed for the controller closing the same books every month, not the operator chasing a settlement file.

Upload AR aging and the GL extract for the receivables control account. ReconPe proposes a SUM_AGGREGATE rule on amount partitioned by counterparty (cardinality is many-to-one — many invoices roll up into one journal entry). ACRE's aggregation blocker searches for subsets of AR rows summing to each GL JE within tolerance. Twelve invoices matched to a ₹25 lakh JE in 240 ms is typical.

Upload the bank statement and your cash GL extract. Iris detects the bank format. Argus proposes UTR canonicalisation rules to handle cross-system truncation. ACRE matches with date-shift tolerance for transactions that posted to GL on a different day than the bank settled them. Cross-period memory carries in-flight wires forward to the next month.

Yes. Approval chain rules are configurable for high-value or regulated dispositions, with separated reviewer and approver roles. Audit trail (rule, residual, confidence, approver, timestamp) exports as a workpaper. Segregation-of-duties on rule edits and approvals is enforced; the audit log is immutable.

Yes. ReconPe has a permanent free tier with 3 reconciliations per month, 1 marketplace, core ACRE matching, and a downloadable review report — no credit card required. Paid plans start at ₹3,999/month (Starter), ₹6,999/month (Growth), and ₹16,999/month (Pro). Annual billing saves the equivalent of two months.

GST reconciliation (matching GSTR-2A / GSTR-2B against purchase invoices and settlement deductions) and TCS reconciliation (Section 52 of the GST Act, deducted by Indian e-commerce marketplaces) are on the Pro plan. Output is finance and CA-ready with discrepancy context.

Alternatives to Cointab include ReconPe (AI-native with a free tier, probabilistic ACRE matching, and risk scoring), Paxcom (rules-based, ₹10,000+/month), and Excel (free but fully manual). ReconPe's differentiators over Cointab are AI exception review, Bayesian confidence per match, adaptive field weight learning per organisation, and stateful cross-run memory.

Yes. Admins invite team members by email; invitees accept via a one-click link and land in their organisation. Roles include Analyst, Finance, Compliance, Admin, Org Admin, Member, and System Admin. Capabilities per role are enforced server-side. Per-plan seat caps are enforced at the invite endpoint.

Yes. Each exception can be routed to a specific team member with a full audit trail of who was assigned, by whom, and when. Each analyst gets an 'Assigned to me' view — a focused queue of exceptions they own.

No. ACRE is deterministic. AI reasons over ACRE's output but never makes the match decision itself. Re-running on the same inputs produces the same output. This keeps the audit trail intact and removes hallucination risk from the matching path.

Yes. Every match decision decomposes into per-field Fellegi-Sunter evidence. Re-running on the same inputs produces the same output. Every material action (assignment, disposition, approval) is timestamped and attributed in an immutable log.

Data is processed in India. AI provider choice (Anthropic, OpenAI, DeepSeek, Gemini) is configurable per tenant — buyers select based on data residency requirements and compliance policy.