JK LogicWorks

The Right-Tool Review

Use the model once. Then own the machine.

A short audit of where large language models sit in your operation, or are about to: what's earning its keep, what's quietly costing you, and what to build once instead.

From the field

Renting what you could own

If you're paying a model to do the same basic jobs every day or every week, you're renting what you could own. The logic never changes, but the invoices never end. We'd rather use the model once, to write a built-for-purpose program that solves the problem directly and efficiently, and hand it to you.

Two we keep seeing

How it goes wrong

The report that never stops costing

A distributor has a chat assistant compile the same stock report every morning: open the sheet, sum the columns, send the summary. It works. It also bills for every run, takes most of a minute, and phrases the totals a little differently each day. One afternoon of ordinary programming produces the same report in milliseconds, for nothing, identically, for as long as the business stands.

The plausible wrong number

A shop drafts invoices from job sheets through a metered agent. Months go by, fine. Then one line item comes out plausible and wrong: wrong in the way that sails through review because it looks right, with nothing underneath to inspect. LLMs are non-deterministic, and if you aren't careful about how you use them, their outputs can be too. A properly built program is strictly deterministic, and can be rigorously tested and proven to return the same answers for the same inputs every time.

The line we draw

The honest test

When to build it

The logic is knowable and stable. The same input must give the same output. It runs at high volume, or for years.

When not to build it

The task is fuzzy or language-shaped. It's a low-volume one-off. A human checks the output anyway. You're prototyping to learn what to build.

Fair's fair

Where models belong

Plenty of places, and we use them every day in our own work. The review tells you which side of the line each of your processes sits on, with "do nothing" as a live option.

(You'll notice we don't say "AI". Long story. Ask us over coffee.)

What it involves

The review

A walkthrough of your operation, then a written report that makes a call on each process, "leave it alone" included. The fee is fixed and agreed before we start, and the report is yours whatever you do with it. It's just as useful before you sign up for the agent platform someone is pitching you. It sits where our Discovery stage sits, so a build follows only if the report says one should.