Skip to content
New: ask the Rexfin Analyst Agent about your model. Every figure comes back cited.

Platform

The planning platform that shows its work.

Rexfin plans budgets, forecasts, scenarios, headcount, sales, and consolidation on one governed model, with every figure still tied to the filed document it came from. Built GCC-first, bilingual Arabic and English.

AES-256 at rest TLS 1.2+ in transit Per-organization isolation Pursuing SOC 2 Type II

Every figure clears a policy gate, or it stops right there.

Animated loop: data flows through governed platform layers, policy gates allow or block each pass.

Works with what you file

Agents

Three modules, one governed model.

None of them work off a copy. Each reads and writes against the same reconciled model, so nothing they say can drift from what finance already trusts.

Analyst agent

Answers questions over your model and its cited actuals in plain language: retrieving figures, running the calculation, and explaining the variance.

Modeler agent

Builds and edits models and scenarios from a plain-language instruction, working inside the governed definitions rather than inventing new ones.

MCP server

Gives other AI tools and internal systems programmatic access to the governed model, under the same permission and citation rules as the interface.

The stack

Four layers. One governed model.

Data climbs the stack; trust is built at every step. The agents on top only ever speak for the model underneath, never around it.

Layer 01 · Ingestion

Connected sources

Accounting systems, banking feeds, warehouses, and uploaded statements flow in continuously. Every account, period, and dimension normalizes into one chart before it ever reaches a model.

  • Excel, CSV, or PDF upload works standalone, no live connection required to start.
  • Documents are tagged against an IFRS taxonomy on intake.
  • A figure only becomes citable once independent extraction engines agree on it.
Drill · evidence
Line item FY24 FY25
Revenue 48,210 61,930
Cost of sales (29,340) (36,470)
Gross profit 18,870 25,460
Evidence · Cost of sales p.25

Ties out 36,470

Layer 02 · The model

The reliability layer

A structured financial model where every metric, segment, and adjustment is defined once and reused everywhere, so revenue means the same thing on Tuesday as it did on Monday.

  • Every figure is a single addressable atom, not a cell copied across drifting sheets.
  • Snapshots per close keep historical answers stable over time.
  • Nothing exports until the reconciliation checks agree.
Export · gated

Export review

FY2025 board pack.xlsx

  • 2 extractors agree
  • Reconciles to printed total
  • Identity checks pass
3/3 checks Export Export approved

Layer 03 · Compute

Deterministic engine

Rollups, ratios, and dependency-graph recalculation run outside the language model, in an exact engine. Same inputs, same output, every time, on every scenario.

  • Change an assumption in plain language; every dependent line recomputes.
  • Three statements are modeled together, with identity checks enforced by the engine.
  • Scenarios stay tied to the reconciled base case, never a detached copy.
What-if · recomputed
Price increase +8%
+4.1%
Revenue
+9.6%
EBITDA
+2.8%
Cash

Layer 04 · Interface

Grounded AI interface

The Analyst and Modeler agents plan which figures to retrieve and what to compute, then answer from the model, assumptions and lineage attached, never asserted without a source.

  • Every answer cites the source it computed from.
  • The MCP server extends the same rules to outside tools and agents.
Ask · cited answer

What drove the margin jump in Q3?

Gross margin rose 3.2pts on lower cost of sales p.24 and favorable mix in exports.

Computed · 2 sources

How trust is built

The chain that never loses the source.

01 Filed

The source document is captured exactly as filed: a PDF statement, a bank feed, an ERP export, a ZATCA e-invoice.

02 Extracted

More than one independent extraction engine reads the figure; it only advances once they agree with each other.

03 Reconciled

Every derived total is checked against the subtotal the source document actually printed, not assumed.

04 Cited

The figure keeps a citation drawer back to the page, the extraction, and the check that confirmed it.

05 Exported

Nothing leaves as a board pack or a statement until every check agrees; unverified numbers get refused.

Why it holds up

Three guarantees the architecture enforces, not the model.

01

One definition, always

Metrics, segments, and adjustments are defined once in the model and reused everywhere, for every person, every cycle.

02

Math outside the model

Totals and ratios run in a deterministic engine, not the language model, so they reproduce exactly, every time.

03

Every answer, traceable

The AI attaches assumptions and lineage to what it answers: nothing asserted that it cannot also show you.

Capabilities

Everything in service of one thing: the number is right.

Reconciliation
Every derived figure traces back to source entries and ties out to your statements automatically.
Definitions library
Metrics, segments, and adjustments are defined once and reused across every planning surface.
Scenario engine
Change an assumption in plain language; dependent lines recompute across the model.
Lineage & audit
Open any number to see how it was built: defensible in diligence, audit, and the boardroom.
Versioned models
Snapshots per close so historical answers stay stable even as the model evolves.
Roles & access
Per-organization isolation with granular access to sensitive figures.

Full planning coverage

The reliability layer underneath every planning surface, not just the evidence trail:

  • Budgeting & forecasting
  • Scenario planning
  • Headcount & workforce
  • Sales, capacity & quota
  • Supply chain & S&OP
  • Consolidation
  • Three-statement modeling
  • Dashboards & reporting

Get started

Put a governed model under your next answer.

Book a demo and we'll connect a sample of your stack, build the model live, and let you interrogate every figure it produces.