Why we built it
FreeAgent is a good accounting tool. It holds your VAT position, your corporation tax liability, your invoices, your bank balance, your profit and loss — all of it. The problem is that it holds them separately, across different screens, with no single view of what you actually have.
Most business owners manage this by instinct. They check their bank balance and estimate from there. But the bank balance is not the answer — it includes money that belongs to HMRC, and excludes obligations accumulating quietly in the background.
Real headroom is bank balance minus VAT owed minus corporation tax due. FreeAgent has everything needed to calculate this. No single screen shows it. Strafi assembles it automatically, every time you ask.
The insight already exists in your accounting data. The problem is that nobody has assembled it into one place and explained it in plain English — until now.
Why WhatsApp
The original Strafi was a dashboard. A well-designed one — but a dashboard nonetheless. Usage data told a clear story: people checked it when something felt wrong, not as a regular habit. The insight had to go to them.
WhatsApp is where UK business owners already are. It requires no new app, no new login, no new habit to build. A 60-second voice note on the commute replaces fifteen minutes of screen-checking. A reply in plain English replaces navigating between four FreeAgent tabs.
Delivering financial intelligence through WhatsApp means it reaches people in the moments when they can actually think about it — not only when they have time to sit at a screen.
The signal engine — deterministic by design
When Strafi tells you that VAT is due in 14 days and your headroom covers it comfortably, that is not a language model making an inference. It is the output of a rules engine — precise, deterministic checks that run against your actual accounting data every time you interact with Strafi.
A signal either fires or it doesn't. VAT due within 14 days: yes or no. Headroom below threshold: yes or no. Invoice gap beyond the calibrated limit: yes or no. No probability, no confidence interval, no interpretation.
Why deterministic rules matter for financial intelligence
Probabilistic AI is appropriate for tasks where the answer is genuinely uncertain — summarising a document, writing a description, answering an open question. It is not appropriate for "is my VAT covered?" That question has a precise answer from the data. A rules engine finds it. A language model guesses at it.
The rules engine is what makes Strafi's output trustworthy and consistent. The same numbers, asked twice, produce the same answer — always.
The thresholds that determine when signals fire are not default values. They are calibrated by accountants who work with businesses like yours, through a structured knowledge pipeline described in section 06.
The role of AI — layered, not central
Strafi uses AI, but not in the way most AI products do. The language model does not reason about your finances. By the time it receives anything, the signal engine has already done the thinking.
Signal engine runs
Deterministic checks evaluate your live accounting data against calibrated thresholds. Signals fire or they don't. The result is a structured payload: what fired, why it matters, what the consequence is, how long the action window is.
Payload is assembled
The complete picture — signal data, financial figures, mechanism, consequence — is structured before the language model sees anything. The model receives conclusions, not raw data to reason over.
Language model narrates
A constrained language model turns the structured payload into plain English. Its role is narration, not analysis. It is given the answer and told to explain it — clearly, concisely, without advice.
Output is delivered
As a voice note, a text message, or a structured answer — depending on what you asked for. The same signal, the same figures, explained the same way every time.
This is what we mean by layered AI — deterministic rules that do the thinking, and a language model constrained to explaining the result. No interpretation, no inference.
The module approach — depth before breadth
Strafi is built in modules — one business type at a time, each deeply calibrated before the next begins. The first module is UK freelancers and small limited companies using FreeAgent. This is not a limitation — it is a deliberate choice.
A shallowly calibrated engine for ten business types is less useful than a deeply calibrated one for one. Generic thresholds produce generic answers. Expert-calibrated thresholds produce answers you can act on.
A freelancer and a seasonal retailer have fundamentally different financial patterns. A signal that fires correctly for one will cry wolf for the other. Serving both well requires separate calibration — separate thresholds, separate suppression rules, separate consequence language.
As each module is completed and validated, the next begins. The pipeline compounds — the cost of each additional module falls, and quality improves as the knowledge base grows.
The knowledge pipeline — how rules get calibrated
Every threshold in Strafi's signal engine has a source. Every consequence description has a professional origin. This is the result of a structured process for capturing, reviewing, and encoding expert accountant knowledge.
Expert interview or questionnaire
Accountants who work with the target business type complete a structured questionnaire covering thresholds, composite patterns, false positive conditions, and leading indicators. Each answer maps directly to a field in the rules engine.
Human review
Every item is reviewed individually — the source quote, the extracted value, the proposed rule. Approved, edited, or rejected with a reason. Ambiguous items queued for follow-up.
Promotion to the engine
Approved items are written to the rules engine. Every change is recorded in a provenance table — source quote, reviewer, reasoning. Six months later, when a threshold looks wrong, the reasoning behind it is readable.
Ongoing calibration
Rules are not set once. User feedback, false positive patterns, and additional expert interviews refine thresholds over time. The engine improves as it is used.
This process — documented, running, and traceable — is what distinguishes Strafi's signal engine from a prompt asking a language model to assess your finances. The moat is not the AI. The moat is the calibrated knowledge behind the rules.
The scaling challenge — and how we're solving it
Building financial intelligence across business types is genuinely hard. Not technically — the architecture is designed for it. The challenge is knowledge: you cannot calibrate thresholds for a seasonal retailer without accountants who work with seasonal retailers.
Freelancer
Drawing rate vs income
How much can be drawn safely given current revenue run rate and upcoming obligations?
Seasonal retailer
Stock-to-cash cycle
Cash drops before peak season are normal. The same drop in February is alarming.
Agency / studio
Project pipeline lag
Revenue recognises late. Invoice gaps don't mean no work — they mean work not yet billed.
Contractor
IR35 and payroll signals
PAYE obligations behave differently from VAT. Suppression rules differ entirely.
The solution is a tiered expert sourcing model. UK specialists for tax-specific signals where jurisdiction matters. Qualified practitioners from broader markets for universal signals where the underlying financial behaviour is the same regardless of geography.
This approach reduces the cost of each additional module significantly while maintaining the quality that makes the output trustworthy.
Where we're going — past, present, and future
FreeAgent tells you what happened. Stripe tells you what is happening. A CRM tells you what is about to happen. Together they give Strafi past, present, and future in one signal engine — a proposition no single-source tool can match.
The architecture is data-source agnostic. Adding a new data source means a new expert interview variant and a new module, not a new engine.
The cross-source composite vision
The most defensible signal Strafi will eventually produce is one no single-source tool can see. Stripe MRR declining 15% plus CRM pipeline below 2x coverage plus FreeAgent invoice gap widening fires 6–8 weeks before a cash problem appears in the bank balance. An accountant looking at FreeAgent alone cannot see it. Strafi can.
The sequencing is deliberate: FreeAgent first, deeply calibrated. Xero and QuickBooks next. Then Stripe, adding real-time revenue signals. Then CRM, completing the pipeline picture. Each layer compounds the value of the one before it.