Forward-deployed engineering · by Sapio
AI that reaches production.
Most AI stops at the demo: the legacy system with no API, the security review nobody scheduled, the data nobody cleaned. We put a senior engineer inside your team who owns one use case until it runs every day — and leaves your people able to run it without us.
01 · The problem
You have tried AI, or been pitched it, and nothing reached daily use.
of AI initiatives have delivered the return expected of them, and only 16% have scaled across the business.
of senior executives name integrating AI with existing systems, tools and APIs as a blocker, and 77% name getting people to use it daily.
of European companies that considered AI and did not proceed say the reason is lack of relevant expertise. Cost is a distant second at 38.4%.
That last number is not a technology gap. It is a people-shaped gap, and it can be rented.
02 · Why it stops
The model was never the problem. Everything around it was.
What you were sold
- A model that answers well in a demo
- A platform licence and a pilot
- A slide deck of use cases
- A proof of concept nobody uses
What actually decides it
- The legacy system with no API
- The security review nobody scheduled
- The data nobody cleaned
- The team nobody asked
A demo runs once — in a room, on data someone tidied up beforehand.
Production runs again every working morning, on that day's real data, with nobody from us in the building.
Someone has to sit on your side of that gap and own what happens next. Not advise on it. Own it.
03 · The idea
A senior engineer who works inside your company, on your systems, with your people.
And owns one use case until it runs in production and your team actually uses it. It is how Palantir, OpenAI, Anthropic and Google deliver into their largest customers. It is not a new idea. It is simply not one that has reached the European mid-market.
Not a consultant
A consultant recommends and leaves. The system is still someone else's job, and the document ages faster than your stack.
Not an outsourcer
An outsourcer builds to the specification you wrote. If the specification was wrong, you get working software nobody uses.
Not a contractor by the hour
Hourly billing rewards duration. We are paid for something reaching production and being used.
04 · What we do differently
Three things a supplier billing you for hours cannot offer.
Your stack, your controls
Your environment, your cloud or on-premise, your security review, EU jurisdiction, a DPA on request, and the AI Act classification of the use case in writing before anything is built.
The outcome is in the contract
Before we start we agree one production milestone with a date and one adoption number measured 30 days after go-live. Part of our fee is paid only when both are true.
No dependency when we leave
One of your engineers pairs with ours from week one. The code sits in your repositories, the documentation is written for your team, and the evaluation set proves it still works after we go.
05 · How it runs
Three days to know. Ninety to have it running.
The tech call
You describe the process. We tell you whether it is worth doing, what the first step is and what it costs. If it is not worth doing, we say so — and the call still saved you a quarter.
Audit and roadmap
Two or three processes mapped with the people who run them · a data-readiness check on your real extracts · AI Act classification · a written roadmap · one use case demonstrated on your own data. Credited in full against the placement if it starts within 60 days.
The embedded engineer
One senior engineer in your team, with Vlad Tudor as deployment lead. Written plan on day one, the first workflow live in your environment by day ten, weekly written update, fortnightly sponsor check-in, monthly steering. Full time when the work justifies it.
The three months are one fixed price, with the milestone, the adoption number and the part of the fee held back until both are met. No hourly rate is quoted, because you are not buying hours.
06 · What it looks like
Ninety days, in the units you already measure.
Ninety days on a process you already measure
An illustrative engagement, not a delivered project. The numbers are round on purpose: the real ones come from your own baseline, measured in the audit.
- The situation
A 300-person distributor receives about 2,000 supplier documents a month — invoices, delivery notes, order confirmations — in six formats from ninety suppliers. Two people key them into the ERP. The finance lead knows the cost of an error downstream and cannot quantify the cost of the keying.
- Week one to day ten
The engineer is on site for the week. A copy of last month's documents, read-only access to the ERP, and two hours with the people who do the keying, which is where the real exception rules live. Baseline measured: documents per format, minutes per document, error rate found at reconciliation. By day ten the two highest-volume formats are read automatically and written to the ERP staging table, with everything below the confidence threshold routed to a person.
- Day thirty
The adoption number: the share of documents posted without manual keying. The team has folded the exception routing into their morning routine. One supplier's format, which broke everything in week two, is now handled by a rule the team wrote themselves.
- Day ninety
Five formats live, covering most of the volume. The milestone is met. The distributor keeps the code, the documentation, the evaluation set, and the engineer who paired. The saving is their own arithmetic: documents per month × minutes saved × their loaded cost, against our fee and the run cost — and they can compute it without us, which is the point.
07 · What we have shipped
Systems we have built, inside other people's companies.
Tender screening. An agent that reads new public tenders and flags the ones that fit a company, judged against its past projects. In testing it found the relevant tenders; it is not yet in daily use.
Support assistant over a retailer's own documentation. From first conversation to a system the support team used daily.
Speech recognition. Delivered with the training pipeline, so the client's own team could retrain it without us. That is the test of a handover.
ai-aflat.ro. A free AI assistant over Romanian legislation, updated daily and used by 15,000+ people. It led to an AI session for 100 lawyers at Baroul Cluj.
08 · What you keep
When we leave, the capability stays in the building.
- The code, in your repositories
- Documentation written for your team, not for us
- The evaluation set that proves it still works
- One of your engineers, trained to run and extend it
- A written review of what shipped and what it would take to go further
Why we publish this
The fear with any outside AI partner is lock-in: that you end up renting your own system back. Pairing your engineer in from week one and handing over a working evaluation suite is how we make that impossible. It is also why clients come back for the second use case instead of being trapped in the first.
09 · The alternatives
What else you could do, honestly compared.
| Option | What you get | What it costs you |
|---|---|---|
| Hire an AI engineer | Permanent capability, if you find one | It takes 7.7 months on average to fill an IT position in Germany, and 57% of EU firms with open IT vacancies could not fill them at all. Bitkom Research, 855 German companies · Eurofound, IT sector in focus |
| A consultancy | A strategy and a roadmap | The build is still nobody's job, and the document ages faster than your stack. |
| An outsourcer | Capacity you direct, by the day | You carry the specification risk and the adoption risk. It is your outcome, not theirs. |
| The platform vendor | The licence and a pilot | Their engineers are sized for their largest accounts, not for yours. |
| An embedded engineer | One use case in production and a team that can run it | Starts in weeks, one milestone at a time, part of the fee at risk on the result. |
10 · Questions
What people ask before the call.
Staff augmentation gives you capacity that you direct: you write the specification, you own the result. We take responsibility for one named outcome, we direct the engineer's work, and part of our fee depends on that outcome happening. The difference is written into the contract, not just the pitch.
Access and a baseline. We agree the metric with your sponsor, get the engineer into your systems, map the current process with the people who actually run it, and publish a written plan on day one. The first workflow runs in your environment by day ten.
No, and you should not choose one before the audit. The audit tells you which parts of the process are worth automating at all, what your data can support, and which stack fits your constraints. Our audits never name a vendor as a foregone conclusion, and any partner relationship we hold is disclosed in writing.
Inside your company. The audit is three days on site. During a placement the engineer is on site for the first week, for go-live and for handover, plus the days agreed at kickoff. Everything else is remote in your time zone. Travel outside your city is at cost.
The audit classifies the use case under the AI Act and puts it in writing. We work inside your environment, on-premise where you require it, under an EU entity, with a DPA on request. For regulated clients we arrive with the register fields, third-party clauses and exit plan already drafted.
One named senior engineer, whose profile you see before you sign anything, with Vlad Tudor as deployment lead on every engagement: he sets the target with your sponsor and owns adoption. Sapio is a Bucharest AI engineering company; the people are ours, not subcontracted out of sight.
Who does the work
Vlad Tudor
Founder of Sapio AI, and the deployment lead on every engagement: he sets the target with your sponsor and owns adoption while the engineer writes the code. He also teaches this work — corporate workshops, the AI Leaders network, conference talks, TEDx, Forbes. Most AI projects do not fail on code; they fail because nobody senior owned the change inside the business.
One call, 30 minutes, no deck
Bring the process that annoys you most.
We will tell you whether AI belongs in it, what the first step is, and what it costs. If the answer is no, you will get that straight — it is the cheapest thing we can give you.
Request the tech call