For operations, finance and support teams
Automate the typing, not the judgement
We build AI automation for Greek companies - document extraction, assistants grounded in your own material, and the chains of steps a person currently babysits between two systems. Every one ships with an accuracy number measured on your data, and a person keeps the expensive decisions.
The bench below is running in your browser. Feed it a bad scan and watch what it refuses to guess.
In plain language
What is AI automation?
AI automation is using a language model to do the reading, typing and matching inside a business process that a person currently does by hand - extracting fields from an invoice, answering a support question from your documentation, routing a case to the right team - while the process itself, including who approves what, stays under your control. It differs from traditional automation in that it handles unstructured input: scans, emails, free text, documents that arrive in a different shape every time. It differs from simply giving staff a chatbot in that it is wired into your systems, measured against your own data, and built to escalate rather than guess.
Words that get used interchangeably and should not be
- AI agent
- A model given tools and a goal rather than a single question - it can look something up, write to a system, and decide the next step. Useful where the path varies. Risky where it can act without review, which is why ours ask before anything irreversible.
- RAG (retrieval-augmented generation)
- The assistant searches your own documents first and answers from what it found, with citations. It is what makes the difference between an assistant that knows your returns policy and one that invents a plausible one.
- Self-hosted LLM
- An open-weight model - Llama, Mistral, Qwen - running on your own server or a Greek or EU data centre instead of a vendor's API. Slower to set up, more to maintain, and the only honest answer when the data legally cannot leave.
- Human in the loop
- The model prepares, a person confirms. Not a compromise on ambition - the correct design for anything that moves money, signs something, or contacts a customer.
The extraction bench, running here
Pick a document and watch the fields come out with a confidence score each. Anything the model is unsure about routes to review instead of being written through - which is the whole design, and the part vendors leave out of the demo. Degrade the scan quality and the confidence drops honestly rather than staying green.
Document
Scan quality
Fields under 90% go to a person. Drag it and watch the queue.
Choose a document and run it.
Runs entirely in your browser. Nothing you click here is sent anywhere, and the confidence model is a simulation of the real pipeline's behaviour, not a live model call.
Why now, in Greece specifically
Not because the technology is new - because the gap between what Greek companies use and what the rest of the EU uses is wide enough to be a commercial advantage for a couple of years.
- 8.9%
- of Greek enterprises with 10 or more employees used AI in 2025, against an EU average of 20% Source: State of SMB Automation in Cyprus and Greece 2026
- 50–60%
- subsidy on software and equipment under ΕΣΠΑ Ψηφιακός Μετασχηματισμός ΜμΕ, from a €135m budget Source: espa.gr
- 4 regimes
- GDPR, NIS2, DORA and the EU AI Act now stack audit and residency duties on the same automation Source: Enterprise guide to GDPR-compliant LLM deployment
What we actually build
Four shapes cover almost every enquiry. If yours is none of them we will say so on the first call rather than reshaping your problem to fit what we already know how to do.
Documents in, structured data out
Invoices, delivery notes, CVs, contracts, customs paperwork - whatever arrives as a PDF or a phone photo. We extract the fields you need, check them against what you already hold, and attach a confidence score to every row so a person reviews the uncertain ones and nothing else. myDATA-bound documents get validated before they are transmitted, not after.
- OCR
- Extraction
- Validation
- myDATA
- Human review
Support that answers from your own material
An assistant grounded in your documentation, past tickets and product data rather than the open internet. It cites the source of every answer, it hands over to a person when it does not know, and it works in Greek - which most off-the-shelf assistants do measurably worse than they do in English.
- RAG
- Citations
- Greek
- Escalation
Processes that run themselves
The chain of steps between two systems that somebody currently watches: triage, routing, reconciliation, chasing. We automate the path and leave the approval exactly where a mistake would be expensive. Usually this is the cheapest project with the largest number attached to it.
- Workflows
- Integrations
- Approvals
- Reconciliation
Self-hosted models and private deployments
Open-weight models on your own infrastructure or in an EU data centre, when the data cannot leave or a regulator will ask where it went. We will also tell you when you do not need this: below roughly two million tokens a day an API is cheaper and easier to run, and pretending otherwise would be selling you a server rack.
- Llama
- Mistral
- Qwen
- vLLM
- On-premise
What it connects to
An automation that cannot reach your systems is a demo. These are the connections that come up in Greek engagements over and over.
- myDATA / ΑΑΔΕ
- Documents validated against the schema before transmission, so a rejection is caught at extraction rather than at the tax authority.
- ERP and accounting
- SoftOne, Entersoft, Epsilon, SAP, or the bespoke thing built in 2009 that everything still depends on.
- Email and shared inboxes
- Where most Greek business processes actually live. Triage, extraction and routing straight out of the inbox.
- Viber and WhatsApp
- Viber is where Greek customers reply. An assistant that only lives on your website is answering the wrong channel.
- Ticketing and CRM
- Zendesk, Intercom, HubSpot, Jira - grounded answers written back as drafts a person can send.
- Your own database
- Postgres, MySQL, Firestore, or a read replica we are allowed nowhere near production with.
What we measure
Every page on this site argues that a number measured on your own data beats a vendor benchmark. That is cheap to say, so here are the four we would actually put in front of you.
- Field-level accuracy
- Scored on an evaluation set built from your documents, not ours. You get the number and the failures together, because the failures are what tell you whether the number is usable.
- Share going to review
- How much a person still has to look at. Watch it fall too fast and it usually means the threshold got loose, not that the model got better.
- Errors that got through
- The expensive ones. Found by sampling what was auto-accepted, because by definition nobody flagged them.
- Cost per document
- Tokens and human minutes in the same figure. Leave the second term out and the number flatters itself.
Reference builds
Things we built ourselves to show the approach. Not client work - where a real engagement is publishable it says so, with a name on it.
- Wholesale and distribution
The supplier invoice queue, read once by a machine
A reference build: supplier invoices arrive by email, are read, validated against the ledger and the ΑΦΜ registry, and land in the ERP with only the uncertain lines waiting for a person.
Python · Claude API · Postgres · myDATA · Next.js
- Customer support
A support assistant that cites where it got that
A reference build: a Greek-language assistant answering from a company's own documentation and past tickets, with a citation on every answer and a hard rule about what it does when it does not know.
Python · pgvector · Claude API · Viber API · Next.js
Questions we get asked about this
- How much does AI automation cost?
- We quote per engagement rather than publishing a list, because the same words describe a two-week pilot and a year of work. The shape is what is fixed: a scoping engagement first, priced on its own and yours to keep, ending in a written scope and a fixed price for phase one before any code exists. You will have a band from us on the first call, once we know how many documents a month, which systems they have to reach, and whether the data is allowed to leave your infrastructure.
- Should we self-host a model or use an API?
- Use an API unless you have a reason not to. The honest thresholds: self-host when the data legally cannot leave your infrastructure, when a regulator will ask where it went, or above roughly two million tokens a day where the economics turn. Below that an API is cheaper, faster to ship and less to maintain. We will tell you which side of that line you are on during scoping.
- How do you know the automation is accurate enough?
- We measure it on your real data before you trust it, not on a vendor benchmark. That means building an evaluation set from your own documents or tickets, scoring against it, and showing you the number and the failures. If it is not good enough you find out during the pilot rather than in production.
- Do the models work properly in Greek?
- Better than they did, and worse than in English - which matters more than most vendors admit. Extraction from Greek documents and Greek-language support both need testing on Greek material specifically, and model choice changes on the results. We evaluate in Greek rather than assuming English performance transfers.
- Will our data be used to train someone's model?
- Not with the defaults we use. We choose providers and configurations that do not train on inputs, and where the data cannot leave your infrastructure at all we run open-weight models on it instead. Which of the two applies to you is a decision we make in scoping, in writing.
- What happens when the model gets it wrong?
- It routes to a person rather than guessing. Low confidence is a first-class outcome in everything we build, not an error state - an automation that fails loudly is worth far more than one that is quietly wrong. Anything that moves money, signs something or contacts a customer stays behind an approval until you decide otherwise.
- Can you work with our existing ERP?
- Usually. SoftOne, Entersoft, Epsilon and SAP all come up regularly, as does bespoke software nobody has documentation for. Where there is no API we read the database or the export, and we say so plainly rather than describing it as an integration.
- Is this eligible for ΕΣΠΑ funding?
- Software and equipment spend under Ψηφιακός Μετασχηματισμός ΜμΕ is subsidised at 50–60%, and custom software development generally qualifies. We are not ΕΣΠΑ consultants and will not advise on your application, but we will write the technical specification and costing your consultant needs.
Bring us a process, not a brief
The most useful first call is about one thing somebody on your team does every week that they should not have to. Forty-five minutes, and you will leave knowing whether it is worth automating and roughly what it costs.
Or write to hello@appcrafters.gr.