Service

AI automation

Language models are good at a narrow, specific thing: reading messy input and producing structured output. That is most of what makes a task tedious. We find the tasks in your week where that applies, build the system that handles them, and hand it over.

The interesting question is never whether a model can do something. It usually can, in a demo. The question is whether it does it correctly often enough to be left alone, what happens the times it is wrong, and whether anyone would notice. That is the part we build for.

What this actually looks like

Not a chatbot bolted onto your website. The work is almost always invisible: something that runs when an email arrives, when a file lands in a folder, or on a schedule, and puts the result where a person was previously typing it.

Typical work

Inbound triage. Reading email or form submissions, classifying them, routing them, and drafting a first reply for a human to approve.
Document extraction. Pulling structured fields out of invoices, statements, contracts or scanned forms — the ones that arrive in twelve different layouts.
Summarising. Long threads, call transcripts, or reports condensed into the two paragraphs someone actually needs before a meeting.
Enrichment and matching. Reconciling records that nearly match, where the rules are fuzzy enough that a spreadsheet formula was never going to do it.
Drafting. First-pass content that a person edits rather than writes — quotes, descriptions, responses that follow a house pattern.

How we build it

The model is the smallest part. Most of the engineering is the plumbing around it: getting the input reliably, giving the model only what it needs, checking the output before anything acts on it, and making failure visible.

RetrievalThe model sees the relevant records, not your whole database. Narrower context is cheaper, faster and more accurate.

EvaluationA test set of real examples with known-correct answers, so accuracy is a number you can watch rather than a feeling.

Human in the loopAnything irreversible — sending, paying, deleting — waits for approval. Reversible things run unattended.

Audit trailEvery decision logged with its input, so when something is wrong you can see why rather than guess.

Cost controlModel choice per task. Most steps do not need the expensive model, and the bill is what makes automations quietly stop being worth it.

What it costs

A first automation is scoped to ship inside four weeks at a fixed price, quoted before anything starts. If you would rather test the idea before committing, a pilot states up front what result would make us recommend stopping.

If you are not sure which task is worth automating first — most people are not — the Automation Audit answers that specific question, and its fee is credited back against a build.

What you end up owning

On handover

The source code, in your repository
Systems running on your accounts and API keys, not ours
Documentation written for a human, not generated from function names
Thirty days of fixes included after launch

Nothing stops working if you stop paying us. That is deliberate, and it is the main thing that separates this from a subscription dressed up as a build.

When we will tell you not to

Some tasks are not worth automating: they happen twice a month, they take four minutes, or the cost of being wrong is far higher than the time saved. Some need a rules engine rather than a model, which is cheaper and more predictable. And some need the underlying data cleaned up first — that is what the AI readiness review is for.

You will get that answer plainly, including when it costs us the work.

Not sure if your task qualifies? Tell us what it is and roughly how often it happens. You will get a straight answer on whether it is worth automating — and a rough number — usually within one business day.

Get your free review

All services Home otr-world@pm.me