We measure the work first, then automate the part that costs the most

Most of an automation engagement is arithmetic, not modelling. We count how many times a thing happens in a month, how many minutes each instance takes, who touches it, and what share falls out as an exception. Until those four numbers exist on paper, every later claim about savings is a guess wearing a spreadsheet.

Then we build the integration that removes the re-typing: an API where one exists, a scheduled job against a database where it does not, a parser for the file format your supplier will never stop emailing. Most of it is unglamorous code — queues, retries, idempotency keys, an audit trail somebody can read during a dispute.

The last step is the one that gets cut everywhere else. We re-measure with the same instrument, on the same process families, and write down the delta including the cases the automation refused to handle. A number produced by a different method than the baseline is not a result; it is a coincidence.

We use language models for three jobs: pulling fields out of documents that have no schema, sorting free text into categories you already maintain, and drafting a reply a person approves before it sends. Everything else in an automation is deterministic code, because you cannot set a breakpoint inside a probability distribution.

That boundary is a cost decision as much as an engineering one. Human review of model output routinely costs several times the inference bill, so the design question is never whether the model can do it, but how many outputs a person has to read and what happens to the ones nobody reads.

MIT Project NANDA put roughly 95% of enterprise GenAI pilots at no measurable return to the P&L. We treat that as the base rate we are arguing against, not as a marketing problem. It is why every build we scope carries an acceptance threshold signed before the first result is seen.

A European car marketplace carrying over 300,000 listings. A digital dictionary for the Bulgarian Ministry of Education at beron.mon.bg, reachable from any school in the country. An NLP module that reads incoming email and Instagram messages and drafts the reply from the product and pricing data the team already maintains, which removed about 85% of the manual typing on repeat trading questions. Email-marketing automation for a beauty brand that removed about 60% of the time their team spent on outreach.

There is no logo wall on this site, because four logos is not a wall, and arranging four marks to look like twenty is the first small lie an agency tells itself. One of the four is open to anyone without asking us for anything: beron.mon.bg is published by the ministry and loads in your browser right now. The other three we will walk through in detail, and we make the client introduction before you sign rather than after.

We will not describe a fifth project until there is one. If a page here ever carries a client result, the client measured it, and the sentence names who measured it and over what period. Where the only reading is our own instrument — a listing count, a queue length — we name the instrument and the period it covers, so you can repeat it rather than believe it.

Published industry figures are fine as context and we attribute them every time — Eurostat has Bulgarian enterprise AI adoption at 8.55% against a 19.95% EU average; NANDA has the pilot failure rate. Those are other people’s measurements, and the sentence always says so.

What you will not find is a percentage with no baseline, no timeframe and no instrument attached. A round number with no method is our industry’s equivalent of a stock photograph, and the technical buyer we want reads it in about four seconds.

Where we state a target rather than a result, the sentence says target. Designing first-wave cases against a modelled payback under twelve months is a commitment about how we scope. It is not a claim that your payback will be eleven months.

The person on your first call writes the code. There is no account manager between the conversation and the build, and no discovery team handing a document to a delivery team that never sat in the interviews. On a six-to-twelve week project this is straightforwardly faster and cheaper.

The cost is capacity and bus factor. We run a small number of engagements at once, we say no when the calendar is full rather than staffing a project with whoever is free, and a serious illness is a real schedule risk on a team this size. Every proposal names who is accountable and what happens if that person is unavailable.

Specialists join by name for the parts that need them — a data engineer, a designer, a Bulgarian-language linguist, a DPO where a signature is not ours to give. They are contracted per engagement rather than listed on a team page to inflate headcount.

Everything here exists in Bulgarian because it was written in Bulgarian, not pushed through a translator once the English was signed off. Operations directors in Sofia read the Bulgarian version, and machine-translated technical Bulgarian is recognisable inside one paragraph.

There is a commercial reason as well as a courtesy. ERP, WMS, OEE and RAG stay in English inside a Bulgarian sentence because that is how the people who use them speak. A page that replaces them with invented Bulgarian equivalents tells the reader nobody in the room has done the work.