AI and workflow automation, with the evaluation harness built in week one.

A step a person does the same way every time, or a rule someone is supposed to remember. That's less glamorous than an agent that runs the company. It's also where the hours and the mistakes live.

The clearest example we've shipped has no model in it. Medicileaf sells CBD in the US, where what a product page can claim changes by state and the list keeps moving. Manual review stopped scaling once a small team was publishing product copy, education content, and promotions on a normal cadence. We moved the rule out of people's memory and into the system: every claim lives in a shared library with a review status, and a claim that hasn't passed review can't ship. Structurally, not by reminder.

Same shape on HospiHealth. 'Email your CV' became a structured intake with a custom admin, and forty-three applications across three roles have gone through it without a spreadsheet.

Neither needed AI. Both replaced a step that was quietly failing.

A model earns its cost when the input resists structure: free text, documents, support tickets, a catalog that changes every week. On LaunchProd we built retrieval over a manufacturer catalog that updated weekly, with a refuse-to-answer threshold so the system returns nothing rather than a confident wrong answer. A fine-tuned model would have been stale before the first cohort signed in.

The model is one component. The work is the system around it: integrations, failure handling, human approval gates, an audit trail. An agent without those is a demo with an invoice.

On LaunchProd the harness came in month three. Before that we judged retrieval quality by eye, reviewing samples manually. That isn't evaluation. That's vibes.

We made three parameter changes in weeks three through five with no reliable signal on any of them, and only learned which ones helped in week eight, once the harness existed and we ran them backward. It cost us five weeks. Every AI build we take now scopes the harness in the discovery document.

If we can't write down what a correct output looks like, we don't ship the feature. Define correct, build the check, then ship the automation.

Integrations, more than the model. Every system an automation reads from or writes to adds a connector, auth, and new ways to fail. Evaluation is a line item in every quote we send, not an afterthought, and a customer-facing workflow needs a clean handoff to a human, which is product work, not prompt work.

Describe the workflow: what triggers it, what systems it touches, and what happens when it's wrong. We reply with pricing within 24 hours.

If a SQL query and a rules engine answer the question, ship those. They don't hallucinate, they don't drift, and they cost almost nothing to run. Most AI requests we see are filtering problems, not inference problems.

And if the step works fine by hand, leave it by hand. On Joshua Trees the founder wanted the storefront synced live with the yard's paper inventory. We shipped a manual 'sold' toggle and said: automate it once you know how often a listing flips. It flips a few times a week. It's still manual, and nobody has asked for more.

Automate the step that breaks under manual load, not the step that feels unfinished.