Why most finance AI breaks on real data (and what agent-based reasoning fixes)

TL;DR: Most AI-native finance tools that look polished on demo data fail when they encounter real production ledgers. The reason is architectural, not a model-quality issue: finance data is high-volume (thousands of transactions per month even for lean scale-ups), exception-heavy (reversals, partial payments, intercompany loops, foreign-currency adjustments), and demands deterministic, auditable logic that a single-prompt LLM cannot deliver reliably. Agent-based reasoning, a network of specialised agents that plan, execute, check, and hand off the way a finance team does, is the architecture that survives real books. Cortena learned this during its August 2025 Xero beta with five early adopters: the integration held and the interface stayed responsive, but the AI engine became unstable once full-volume real-world finance data started flowing. The team paused testing, rebuilt the core around agent-based reasoning, and relaunched in mid-September 2025 narrowed to one proving-ground use case, cash flow, before expanding to receivables, payables, and forecasting.

The demo-to-production gap in finance AI

Most AI-native finance tools look sharp in a pitch deck. A clean chart of accounts, a handful of invoices, a natural-language question like "why did expenses move?", and a tidy answer. Then the real data lands, and the wheels come off. Transaction volumes jump 50x. Categorisation rules contradict each other. Subsidiaries follow different ledger logic. Exceptions outnumber standard cases. The same system that nailed the demo starts hallucinating, stalling, or producing answers that look right and aren't.

This isn't a model-quality problem. It's an architecture problem.

Why real finance data breaks generic LLMs

Three characteristics of live finance data expose the limits of single-prompt AI:

• Volume at the transaction layer. Even a lean scale-up pushes thousands of transactions a month across banks, invoices, and journals. Stuffing that into a context window is neither feasible nor useful.
• Exception density. Finance runs on edge cases: reversals, partial payments, intercompany loops, foreign-currency adjustments. A model tuned for "typical" breaks on atypical, which is where value lives.
• Deterministic expectations. A sales dashboard can tolerate a fuzzy answer. A cash-flow statement cannot. Finance teams need repeatable, auditable logic, not creative interpretation.

Generic LLMs are optimised for fluency, not for finance-grade rigour. That mismatch is why most finance AI stalls the moment it meets real books.

What agent-based reasoning fixes

Agent-based reasoning replaces the single all-purpose prompt with a network of specialised agents that plan, execute, check, and hand off. One agent pulls and cleans transactions. Another applies accounting logic. Another reconciles against bank data. Another surfaces the narrative: "cash dipped because of two late client payments and a VAT outflow."

The shift matters for three reasons:

1. Separation of concerns. Each agent has a narrow job with narrow inputs. Failure is contained and traceable, not a black box.
2. Determinism where it counts. Accounting logic runs as rules, not creative writing. The model reasons; the logic stays stable.
3. Auditability. Every step leaves a trace. Finance teams can inspect what the agent did, not just what it said.

Agents think like analysts, plan like CFOs, and surface the signal before the question is even asked. That's the architecture finance deserves.

Why Cortena is starting narrow, with cash flow

When Cortena ran its first beta in August 2025, five early adopters connected live Xero data. The integration synced cleanly. The interface held up. But the AI engine cracked under the full volume of real-world finance data.

It was a fast, valuable learning: the foundation needed sharper logic, not more scope. So instead of pushing breadth, the team is going deep on one essential use case, cash flow, and using it as the proving ground for a finance-native agent architecture.

Cash flow is the right starting point because it touches everything finance teams care about: timing, risk, runway, and the gap between what the books say and what the bank shows. If agents can handle cash flow reliably, the pattern extends to receivables, payables, forecasting, and beyond.

What this means for finance teams evaluating AI tools

If you're looking at AI-native FP&A platforms, three questions separate the serious tools from the demo-wear:

• How does it behave on your full transaction volume? Ask for a test on real data, not a curated sandbox.
• What happens when the model is wrong? Look for traceable outputs, not just confident ones.
• Is it one model doing everything, or specialised agents doing one thing well? Architecture predicts reliability more than model choice.

Cortena's next beta ships in mid-September with the rebuilt core. Xero users can resume testing. Non-Xero teams can explore via demo data and share feedback directly.

FAQ

What is agent-based reasoning in finance AI?

Agent-based reasoning splits finance automation across multiple specialised AI agents: one handles data cleaning, another applies accounting logic, another reconciles, another narrates. Each agent has a narrow job, clear inputs, and traceable outputs, so the system behaves more like a finance team than a single black-box model.

Why do most AI finance tools fail in production?

They're built on single-prompt LLM architectures tuned for fluency, not finance-grade rigour. Real finance data is high-volume, exception-heavy, and requires deterministic logic, conditions that expose the limits of generic models the moment demo data is replaced by live ledgers.

Is Cortena live?

Cortena's first beta ran in August 2025 with five early Xero adopters. After learning from that round, the team is shipping a rebuilt core in mid-September focused on cash flow. Read more about where finance automation should actually start.

What is Cortena?

Cortena is an AI-native FP&A platform that turns real finance numbers and market signals into foresight, surfacing risks, opportunities, and answers before the business has to ask. Built for lean finance teams who need clarity, not more dashboards.