There is a version of this product that would work differently. You connect your accounting software, an AI reads your data, and it tells you what it thinks is important. The AI decides. The AI narrates. You trust the AI.

We did not build that version. This is why.

The problem with asking an AI to decide

Large language models are extraordinarily good at processing information and generating fluent, plausible-sounding responses. What they are not good at is consistent, calibrated judgment on numerical thresholds.

Ask an AI whether a VAT liability of £8,200 due in eleven days is urgent, and it will probably say yes. Ask it the same question about £800 due in eleven days, and it will probably still say yes. Ask it about £80,000 due in eleven days for a business with a £200,000 bank balance, and it might say yes again. The AI doesn't know what urgent means for this specific business, in this specific context, against this specific set of obligations. It improvises.

Improvised financial judgment is not financial intelligence. It's pattern-matched language that sounds authoritative. The difference matters enormously when the subject is someone's actual cash position.

The signal engine

Strafi uses a different architecture. Before any language model is involved, a deterministic signal engine runs against your accounting data.

The signal engine is not AI. It is logic. A set of rules, thresholds, and conditions — written in code, calibrated by professional accountants — that evaluate your financial position against a defined set of patterns. Is VAT due within fourteen days? Has real headroom fallen below a threshold calibrated to your business type? Has there been no invoice raised in three weeks? Is the gap between your bank balance and your real available cash widening?

Each of these checks either fires or it doesn't. There is no improvisation. The signal engine produces a list of what is actually worth your attention — ranked by severity, grounded in professional knowledge of what matters for a UK limited company.

Only then does the language model come in.

What the LLM actually does

Once the signal engine has identified what matters, Claude receives a structured payload: which signals fired, at what severity, with what mechanism and consequence. Its job is narration, not judgment. It takes a finished picture and explains it in plain English.

This is a meaningful distinction. The intelligence — the decision about what to surface, what to suppress, what constitutes a material change — lives in the signal engine. The language model provides fluency. It is narration infrastructure, not the brain.

The result is a briefing that is specific, consistent, and defensible. Not because the AI is particularly good at financial analysis, but because the financial analysis was done before the AI was involved.

Why this matters for you

The practical consequence is that Strafi stays silent when nothing material is happening. This is intentional. A financial intelligence tool that sends you something every day, regardless of whether anything has changed, trains you to ignore it. The signal engine's job is not just to identify what fires — it is equally to identify what doesn't.

When Strafi does surface something, you can trust that it cleared a calibrated threshold. Not that an AI thought it sounded important. That a defined condition, grounded in professional judgment, was met.

The calibration question

The obvious follow-up is: who decides where the thresholds are?

The initial thresholds were set by judgment — informed by accounting knowledge, adjusted for UK limited company specifics, and deliberately conservative. Over time, they will be calibrated by two sources: structured interviews with professional accountants, and observed engagement patterns from real users.

The accountant interviews follow a structured questionnaire — specific numbers per area, composite pattern recognition, false positive identification. Every threshold is traceable to its source. Every change is auditable. This is not a black box that adjusts itself. It is a calibrated system that improves with input.

The user engagement data is newer. Every time Strafi fires a signal, we log it. Every time a user engages with that signal — asks a follow-up question, digs deeper — we record it. Over time, that data tells us which signals land and which are noise for specific business types. The threshold adjusts accordingly.

The architecture in one sentence

The signal engine decides what matters. The language model explains it. The accountant knowledge calibrates the former. The user data refines it over time.

That separation — between judgment and narration, between deterministic logic and generative language — is the architectural choice that makes Strafi something other than an AI wrapper around your accounting data.

It is also, we think, the right way to build a financial intelligence tool for businesses that depend on getting it right.