AI Software, Built to Be Relied On
The useful question is rarely “can AI do this?” It is “should it?” We build AI into software where it genuinely changes the outcome, keep it out where plain code does the job better, and make every answer accountable either way.
Most of what business software does is deterministic: move the record, apply the rate, send the document. Code does this perfectly, millions of times, for almost nothing. A language model doing the same job is slower, dearer, and occasionally creative in ways ledgers should never be. So our default is procedural logic, and AI has to earn its way into a build the same way any expensive component does: by doing something code cannot.
What code cannot do is read the world the way people speak it. That is the doorway. The moment your input is a voice memo, a scrawled note, a supplier email, or a question phrased six different ways by six different staff, deterministic parsing breaks and AI becomes the only honest tool for the job.
Where AI Pays Its Way
Understanding speech and free text. Our product FlowSpec listens to a tradesman describe a job in plain words on site, and turns it into a structured, priced quote in Xero. The AI does exactly one job: reading intent from human language. Item matching runs against the user's own price list, quantities obey written rules, and the arithmetic is plain code. The model never invents a price.
Reading documents at volume. Invoices, tenders, compliance paperwork, inbox triage. Work where a person currently skims hundreds of pages to extract twelve fields. AI does the reading; validation does the trusting.
Drafting for human review. Replies, summaries, first-pass classifications. The person stays the decision-maker and the software removes the blank page. This is the least glamorous use of AI and routinely the highest-value one.
Where It Does Not
Anywhere the rules are known, the answer must be exact, or the result moves money. Pricing, tax, payroll, scheduling constraints, permissions. A model that is right 98% of the time is wrong 2% of the time at scale, and a business runs on the 2%. We are direct with clients about this, because an AI feature sold where a database query belongs is an expensive way to make software less reliable.
The Rubric We Actually Use
Every proposed AI feature in a Flow-Through build gets put through the same five-signal test before it earns a line of code.
| The signal | The call | Why |
|---|---|---|
| The input is messy human language or judgement | AI earns its place | Speech, free-text notes, documents written for people. Code alone reads none of these well. |
| The rules are known and must hold every time | Procedural code | Pricing, tax, payroll, compliance. A model that is right 98% of the time is wrong 2% of the time, forever. |
| The output feeds a decision a person reviews | AI, with the working shown | Drafts, summaries, classifications. The person stays the authority; the AI removes the typing. |
| The output moves money or commits the business | Procedural code, or AI behind a verify step | Anything irreversible gets checked deterministically before it executes. Always. |
| The task is rare and high-stakes | A person | Some work should stay human. Honest software says so. |
How It Stays Accountable
AI in our builds runs once, at a defined moment, inside a frame. Inputs are constrained to what the step needs. Outputs come back structured, so they can be checked by code rather than by hope. Anything that commits the business, a price, a payment, a record written to your accounting system, passes a deterministic verify step first. And the working is kept: when the software makes a call, you can see what it read and why it answered the way it did.
The same discipline that governs the rest of our engineering applies here, including the security practices that protect every input and secret. Our security page covers that line in full.
The Short Version
Procedural first. Deterministic code wherever the rules are known. It is faster, cheaper, and always right.
AI where it changes the outcome. Reading language, documents, and judgement-shaped input that code cannot touch.
Accountable always. Structured outputs, deterministic verification, a person in the loop where it matters, and the working shown.
Common Questions
Will AI make the software cheaper to build?
Usually the opposite, modestly. AI adds a model, an evaluation harness, and a verification layer that plain code skips. What it changes is capability: the software can accept inputs that were previously impossible to handle, like speech on a job site. You pay for that capability once, where it earns its keep, and the procedural majority of the build stays lean.
What happens when the model gets something wrong?
The build assumes it will. Every AI step in our software runs inside a frame: constrained inputs, structured outputs, deterministic checks on anything that matters, and a person in the loop wherever the result commits the business to something. A wrong guess surfaces as a flagged item to review, never as a silent error in an invoice.
Which models do you use?
The boring answer is whichever model passes our evaluation for your task at the best cost, revisited as the field moves. The architecture keeps the model swappable, so the software improves as models do, without a rebuild.
Our industry is specialised. Will AI understand it?
Out of the box, partially. Built properly, yes: the system is grounded in your catalogue, your terminology, and your documents, so answers come from your business reality rather than the open internet. FlowSpec understands plumbing supply codes because it reads the tradesman’s own price list, not because the model memorised plumbing.
If you are weighing AI for your business, bring us the problem rather than the technology. On one 15-minute call you get an honest read on what deserves a model, what deserves plain code, and what either one costs.
Book your 15-minute call