Automate the work nobody should be doing.

The first week goes on counting rather than coding. We sit with the person doing the task and time it with a stopwatch. A four-minute process that runs 180 times a week is 12 hours of salary, every week, forever. A nine-minute process that runs twice a month isn’t worth automating, and we’ll say so out loud. Plenty of candidate workflows fail that test, and hearing it in week one is cheaper for both of us than finding out in week six.

Then we resist the model for as long as we can. A typed parser, a lookup table and a regex beat a language model on cost, latency and reliability for anything with a fixed structure. A deterministic step costs nothing per run and fails the same way twice, which means you can write a test for it. In a typical document pipeline about 60 percent of the logic ends up as plain Python. The model comes in where the input is genuinely unstructured. Free-text notes, inconsistent supplier layouts, an email that could mean three different things.

For retrieval we start with hybrid search, and we almost never fine-tune. Fine-tuning costs real money to maintain, and the week your policy documents change you own a stale model. Retrieval augmented generation over pgvector, with a BM25 pass for exact terms like part numbers, gets you most of the way, and updating it means updating a document. Fine-tuning earns its place for narrow-format output at very high volume. It doesn’t earn its place for knowledge. If you’d rather settle that decision before anyone writes code, our AI consulting and readiness assessment does exactly that.

Every automation ships with a confidence threshold and a human queue behind it. A system that’s right 92 percent of the time and silent about the other 8 is worse than no system, because it quietly poisons your database. So we build a labelled evaluation set of 100 to 300 of your real documents before the first prompt reaches production, and we report accuracy per field. When a new model version lands, we rerun the suite instead of guessing. Working this way costs you something. The first two weeks look slow next to a demo someone can build in an afternoon. You see the difference in month three, when nobody has quietly gone back to doing it by hand.

Most of these systems need a face. A review queue, an exception dashboard, an admin screen for the ops lead. We build those in Next.js inside the same engagement, using the same practice behind our Next.js web development work. You can see how that comes together in our published case studies.

The furthest we have taken this is Fiona, a desktop AI voice agent we built for a client, which bridges a live phone call to a realtime model and prices every call from billed usage rather than an estimate.