Where AI actually belongs in your business (and where plain code wins)

Most businesses I talk to get AI wrong in one of two directions. Half are using it for nothing beyond the occasional email draft, convinced it is not ready for real work. The other half want to push everything through it: calculations, lookups, routing, the whole workflow, because AI is the future and the future should presumably do everything. Both camps waste money, just in different currencies. One pays in missed leverage, the other in unreliable systems and API bills.

The rule that sorts it out fits in one line: use AI only for judgment calls. If plain code can answer, code answers.

What counts as a judgment call

A judgment call is any decision where the input is messy and the answer requires reading meaning rather than following rules. Four families cover almost all of it:

• Classification. Is this enquiry a hot lead, a support request, or spam? Is this review angry or just blunt? Which category does this expense belong to?
• Drafting. A first-pass reply to a customer, a product description, a proposal outline. The human edits; the blank page is gone.
• Summarising. Turn a forty-minute call recording, a long email thread, or a month of feedback into the three things that matter.
• Extraction. Pull the vendor, date, amount, and line items out of a photographed receipt or a messy PDF invoice that follows no template.

What these share: the input is unstructured, no fixed rule covers every case, and a wrong answer is cheap to catch and correct. That last property matters more than people think. Judgment calls are where models shine, and also where their occasional mistakes do the least damage, because a human or a downstream check is naturally positioned to catch them.

What AI should never be doing

Now the other side. Adding up invoice totals. Calculating tax. Checking whether a payment came in. Deciding whether to retry a failed job. Looking up a customer's order status. Converting a date format. None of these are judgment calls. They have exactly one correct answer, and code produces that answer instantly, for free, every single time.

Push them through a language model and you pay three times. You pay in money, because tokens cost more than arithmetic. You pay in speed, because a model round-trip is slow next to a database query. And you pay in reliability, because a model does deterministic work correctly almost every time, and "almost" is a word that has no place near your books. A calculator that is right 99% of the time is not a calculator.

We build accounting software, so this line is not theoretical for us. In Lejar, AI reads the receipt you snap a photo of, because deciphering a crumpled receipt is a judgment call. But the ledger math, the tax logic, the double-entry rules? Plain, boring, deterministic code. The AI is never allowed to do arithmetic that matters.

The pattern that holds up: AI at the edges, code at the core

Put those two halves together and you get the architecture that actually survives production. AI sits at the messy edge of the system, where the real world comes in: photos, emails, voice notes, free-text forms. It turns mess into structure: a category, a draft, a set of extracted fields. Then plain code validates that structure and takes over, and everything downstream is deterministic, testable, and fast.

Two rules keep the seam honest. First, validate everything the model hands over, the same way you would validate user input, because that is what it is: a confident guess about messy data. If the extracted amount does not parse as a number, the pipeline stops rather than guessing. Second, keep a human on anything that leaves the building. AI drafts the reply; a person sends it. AI flags the anomaly; a person acts on it. This is the same principle as treating done as a claim rather than a fact: the model proposes, something stricter disposes.

The economics, briefly

Used this way, AI is cheap. A classification or extraction call costs a fraction of a sen, and even thousands of them a month cost less than an hour of anyone's time. What gets expensive is the everything-machine: agents looping on tasks code should own, whole workflows re-reasoned from scratch on every run, retries burning tokens on problems a queue would handle. When an AI bill surprises someone, the cause is almost always deterministic work being done expensively in a model rather than cheaply in code.

There is a compounding benefit, too. A model call you make today behaves slightly differently after the provider updates the model next year. Code does not drift. The smaller the judgment surface, the less of your system can quietly change underneath you.

How to pick your first AI project

If nothing in your business uses AI yet, do not start with a moonshot. Look for a task with four properties: it happens often, the input is messy, a human currently eyeballs it, and a wrong answer is annoying rather than catastrophic. Sorting incoming enquiries. Drafting the first reply. Summarising calls into your CRM. Extracting fields from supplier invoices for a human to confirm.

Each of these is high-volume enough to matter, low-stakes enough to be safe, and checkable enough that you will know within a week whether it is earning its keep. Ship one, watch it, then add the next. The businesses getting real value from AI right now are not the ones with the most ambitious pilots. They are the ones with three or four boring judgment calls quietly automated and verified.

The bottom line

AI is not a workforce replacement and it is not a toy. It is a judgment engine, and judgment is only one ingredient in a working system. Give it the classification, the drafting, the summarising, the extraction. Give everything with one right answer to plain code. Validate the seam between them, keep a human on anything outward-facing, and both halves stay cheap and trustworthy.

This sorting exercise, deciding which parts of your operation are judgment calls and which are just rules nobody wrote down yet, is most of the value of mapping a business for AI. It is exactly the first step of how we build AI systems for clients. If you want help running it on your business, send a brief.