AI implementation for companies in Romania
WhatsApp AI agents, e-Factura checks in SPV and AI next to SmartBill or Saga for companies across Romania. Prices start from €1,260, with a working pilot 2–4 weeks after access
WhatsApp · customer
11:20Hi! We need 12 Ergo Pro chairs for our new office by the end of the month. Can you make it? Any discount for the volume?
AI
Reads and acts
Romanian or English, intent
- 12 × Ergo Pro
- due: end of month
- discount question
ERP · Stock and delivery
Stock and delivery
CRM · Customer record
Customer record
Access and rules
Reply to the customer
11:20Hi! Ergo Pro is in stock, and we can deliver 12 chairs in 3 business days. A manager approves volume discounts: Oliver will message you today by 5 pm.
Bitrix24 · Deal for a manager
- Deal:
- 12 × Ergo Pro, by month end
- Task:
- approve the discount
We connect your systems
-
Quality tested before launch
Criteria agreed before we start
-
Romanian and English
Tested on your own chats first
-
A decision after each stage
Every stage is paid for separately
Which business processes to hand to AI first
Four common scenarios, and the kind of solution that fits each one
Customers and candidates write on WhatsApp and Telegram, managers answer the same questions and fall behind
- Staff dig through policies, PDFs and chats, or just ask their manager because it is faster. Document search: An assistant answers from internal documents with a source link and respects who can see which documents.
- Supplier invoices from SPV are checked against orders by hand, and ANAF notices wait in the inbox. AI next to your accounting: AI works with SmartBill, Oblio, Saga or WinMentor: it matches incoming e-Factura invoices with orders, drafts replies to ANAF messages and fills fields in your CRM.
- The process lives in spreadsheets and messengers, and managers cannot see statuses, errors or workload. Internal tool: A workspace or panel for one process with AI inside: task queue, statuses, disputed cases, reports.
AI implementation case studies
Facts and numbers from published projects
Scenario transcript
Application
- Under the hood: New application. HR system, Vacancy «Cook», Vacancy: Cook. Applied for “Cook”
- Agent: Good afternoon, Daniyar! I’m a recruiter at Magnum. You applied for the cook vacancy. Are you still looking for work?
Interview
- Customer: Yes, I am
- Agent: Thank you! May I ask how old you are?
- Customer: 23
- Under the hood: Screening rules. Candidate’s answers: meet them. Meets the screening rules
- Agent: Tell me an address or district where you’d like to work, and I’ll find a store nearby.
Nearby stores
- Customer: Saina Shalyapin
- Under the hood: Stores with the vacancy. Yandex Geocoder, Almaty, Saina × Shalyapin, Priority: 0.7 km, Nearest: 1.8 km, Nearest: 3.4 km. Priority store first
- Agent: There is an open cook vacancy near the Saina and Shalyapin intersection: 1. Store, 0.7 km 2. Store, 1.8 km 3. Store, 3.4 km Which one suits you?
Trial shift
- Customer: The first one. I can come tomorrow
- Under the hood: Candidate’s answer. store 1, tomorrow. Store 1, tomorrow
- Under the hood: Trial shift booking. HR system, Status: Trial shift, When: tomorrow morning, Changed by: AI Bot. Status “Trial shift”
- Agent: You’re booked for a trial shift tomorrow morning. Please come to the store’s main entrance. Bring your ID; an electronic one from eGov or Kaspi is fine. Dress code: dark trousers or skirt, a plain top, closed shoes.
- System: Status written to Magnum’s HR system · after the shift the agent asks how the day went
AI agent · WhatsApp · MagnumMagnum: HR agent for store hiring
An application in Magnum’s internal HR system turns into a WhatsApp conversation. The agent runs the first screening interview, finds a store near the candidate’s address and books a trial shift. The status goes back to the HR system, so the recruiter sees the result.
- Offers priority stores first, then others nearby, each with the distance
- Sep 2026 the agent went live with Magnum’s store candidates, in Russian and Kazakh
- Understands Russian, Kazakh and mixed speech
- Screen recording from the live botDocument search · Magnum
Magnum: Olzhas knowledge base for staff
An employee asks in their own words: how to request leave, who takes the sick note, where to get an employment certificate. The assistant finds the answer in approved documents and shows which document it came from.
- 150+ checks built from real employee questions in Russian and Kazakh, run against every update
- Understands Russian, Kazakh and a mix of both
- Magnum’s new documents go live after the eval set has been run
Scenario transcript
Request
- Customer: Hello! I’d like to become a supplier
- Under the hood: Site from the QR code. Compass, QR · Compass site, City: Almaty, Site: Compass station. Almaty, Compass station
- Agent: Hello! You’re at a Compass site in Almaty. What is your company called, and what is its BIN?
BIN check
- Customer: Dala Snab LLP, BIN 1809•••••12
- Under the hood: BIN check. Company registry, BIN 1809•••••12, Company: Dala Snab LLP, Status: active. BIN is active
- Agent: BIN verified. What do you supply?
Category
- Customer: Household goods and cleaning supplies
- Under the hood: Supplier category. household goods, cleaning supplies. Household goods
- Agent: Category: household goods. Please send your price list: PDF, Excel or a photo.
Bitrix24 card
- Customer: price_list_2026.xlsx (XLSX · 38 KB)
- Under the hood: Supplier card. Bitrix24, Company details: Dala Snab LLP, Category: household goods, Price list: price_list_2026.xlsx. Card created
- System: Bitrix24 card · company details, category, price list
AI agent · WhatsApp and Bitrix24 · CompassCompass: supplier onboarding into Bitrix24
A supplier scans a QR code at a Compass station and lands in WhatsApp. The agent collects company details, validates the tax ID, picks the category, accepts a price list in any format and creates a card in Bitrix24.
- A QR code at every station, so the agent knows the city and site
- Russian and Kazakh, language detected automatically
- The supplier card goes straight to Bitrix24
Scenario transcript
Question
- Customer: Hello! I’d like to sign up for the marathon
- Under the hood: Race calendar. Knowledge base, races 2026, Almaty Marathon 2026: registration, Astana Half Marathon 2026: registration, Tau Jarys 2026: registration. Three races on the calendar
- Agent: Good afternoon! I’m the Almaty Marathon assistant. Which race would you like to know about? The calendar now has Almaty Marathon 2026, Astana Half Marathon 2026 and Tau Jarys 2026.
Race kit
- Customer: Almaty Marathon. Where do I pick up my race kit?
- Under the hood: Race kit pickup. Knowledge base, Almaty Marathon · race kit, Where: pickup point, How: QR code from the email. Pickup point, QR code
- Agent: Race kits are handed out at the pickup point before the race. Show the QR code from your registration email.
Refund to a manager
- Customer: And if I can’t run, will I get my money back?
- Under the hood: Agent permissions. Refunds: manager. A person decides on refunds
- Under the hood: Handoff to a manager. WhatsApp, To: manager, Summary: race, kit, refund. The manager has the summary
- System: Refund → manager · chat summary attached
- Agent: I’ve passed your question to a manager. They’ll reply right here.
Support · WhatsApp, Instagram · Almaty MarathonAlmaty Marathon: runner support agent
The agent answers participants on rules, schedule, registration and race packs. Refunds, complaints and VIP requests go to a manager with a summary of the chat.
- 9 in 10 conversations resolved without an operator
- ~2,130 conversations with race participants
- WhatsApp and Instagram around the clock
A working version in 2–4 weeks
Five stages. After each one you decide whether to go on. Each stage is paid separately
Pick one process
2–3 working daysWe find where the team loses hours and pick one process with an owner on your side and a result that can be counted.
A short memo: the process, the type of solution, what success means, and what stays out of the first stage.
Process, metric and owner are agreed.
Timelines start once access is granted: without chat exports and a test API the audit cannot start. WhatsApp number verification with Meta takes from a few days to three weeks, so we start it on day one. Projects with several processes or heavy integrations with 1C and custom systems take 3–4 months, but even there a working version ships in the first month.
What AI implementation costs
The cost of implementing artificial intelligence depends on the process and the integrations. These are starting prices, and we name the exact figure after the audit, once we have seen your data and systems
- Timeline: 1–2 weeks
What we build: AI in a CRM or 1C
Price: from €1,260What is included: One feature inside the system: request parsing, draft replies, document sorting
- Timeline: 2–4 weeks
What we build: AI agent
Price: from €1,800What is included: A scenario, answers from your materials, handoff to a manager
- Timeline: 3–5 weeks
What we build: Document search
Price: from €2,700What is included: An assistant over your policies and knowledge base, answers with a source link, role-based access, a test set
- Timeline: 3–6 weeks
What we build: Internal tool
Price: from €3,600What is included: A staff workspace with AI features: summaries, document checks, reports
Model usage (OpenAI, Anthropic and others) and the WhatsApp API are billed separately at the providers’ rates.
Get an estimate for your case
Send the task for an estimateAI quality testing before launch
We agree the threshold before we start and test it on your chats, Romanian and English measured separately
- A set built from your chats
- 100–200 real requests with reference answers from your staff. Romanian messages with and without diacritics are always in the set, and so are questions that switch to English halfway.
- A different model grades
- A separate judge model, not the one that answered, scores the answers. We check its scores against your staff’s own grades.
- Every miss gets reviewed
- An overall score can hide individual errors, so every answer below the threshold is reviewed on its own and fixed before launch.
- Acceptance on new questions
- We sign off on questions the system never saw during tuning, so the result can’t be fitted to familiar ones.
- RO Când ajunge comanda 4187? 0.92
- RO cand ajunge comanda 4187 0.88
- EN Can you invoice our company? correctly handed to accounting 0.90
- RO Cum returnez produsul? judge: right answer, but in English 0.74
- RO Factura nu apare în SPV, ce fac? 0.83
- EN I want to talk to a person handed to a person right away 0.97
- RO Aveti vreo reducere? judge: promised a discount outside the rules 0.62
- EN Can I change the delivery address? 0.85
- RO Pot anula comanda? 0.81
- RO Cât durează garanția? 0.86
Illustration of the mechanics with sample data. The real set is built from your requests during the audit.
What AI implementation in Romania has to handle
We check all of this during the audit, before building. Companies working beyond Romania can start from our general AI implementation page
Romanian, with or without diacritics
Customers often type without ș, ț or ă and switch to English mid-chat. The agent has to understand both and reply in the customer’s language. We pick the model by testing it on your chats, and measure Romanian and English separately.
Buna ziua, cand ajunge comanda 4187?
-
WhatsApp first, Messenger second
Since July 2026 Meta prices Romania as its own WhatsApp market, with higher utility rates. From 1 October 2026 service replies above 1,000 a month per number are paid too. We put both into the running-cost estimate.
Business API · Messenger · 1,000 free replies a month
-
RO e-Factura and SPV
AI pulls incoming e-Factura invoices from SPV, matches them with orders and receipts, and flags mismatches before payment. It drafts replies to ANAF messages and compliance notices for your accountant to approve.
SPV · incoming invoice · purchase order
-
SAF-T (D406) checks
Before filing, AI checks the D406 export for empty codes and turnovers that do not add up, and lists the lines for your accountant. Filing stays with the accountant.
D406 · empty codes · turnover that does not add up
-
SmartBill, Oblio, Saga, WinMentor
We connect through the API where there is one, as with SmartBill and Oblio, and through exports for Saga and WinMentor. AI reads first and writes only through a route you confirm.
SmartBill API · Oblio · Saga · WinMentor
-
GDPR, ANSPDCP and the AI Act
Kazakhstan has no EU adequacy decision, so we sign a DPA and standard contractual clauses. Model keys and EU hosting sit on your account. Since 2 August 2026 the bot says it is AI (Article 50). Outbound voice calls go only to people who agreed: in 2026 ANSPDCP fined an online shop about €54,000 for calls without consent.
DPA · SCC · AI Act Article 50
-
Contract and tax from Kazakhstan
A contract in euros with our Kazakhstan company. Under the Romania–Kazakhstan tax treaty the 16% withholding tax does not apply once you hold our tax residence certificate with a certified Romanian translation, renewed every year. You self-account 21% VAT under the reverse charge.
16% withholding · residence certificate · 21% VAT
Guides for your industry
Each guide covers what to hand to AI first, what stays with people and how to run the pilot
Frequently asked questions
With the process. In the first meeting we look at where the team loses time: customer chats, searching through policies, retyping data, or the lack of a shared picture. That decides the type of solution. If there are several processes, we pick one with a clear owner and a result that can be counted.
AI next to your CRM or accounting starts at €1,260, an AI agent at €1,800, document search at €2,700, an internal tool at €3,600. A full launch with an audit, testing in Romanian and English and integrations usually costs €7.2–10.8K, several tasks on a shared knowledge base €18–27K. Prices exclude Romanian VAT, which you self-account. The exact figure comes after the audit.
Support costs €180–900 a month: every month we run the test set, update the knowledge base and review disputed answers. Model usage, hosting and the WhatsApp API are billed to your company at the providers’ rates. Since July 2026 Meta prices Romania as its own market, and from 1 October 2026 service replies above 1,000 a month per number are paid, so we include messaging in the running-cost estimate.
AI in a CRM or 1C takes 1–2 weeks, an AI agent 2–4, document search 3–5, an internal tool 3–6 weeks. The full path with process selection, an audit and a pilot gets you a working version in 2–4 weeks after access is granted. Projects with several processes and heavy integrations take 3–4 months.
A process owner who makes decisions, 30–100 real requests, conversations, or documents, a list of systems with access to their test APIs, handoff rules, and examples of good team answers. The knowledge base does not have to be perfect: we build the test set together.
It has to, because many customers type “cand ajunge comanda” instead of “când ajunge comanda”. Romanian is a mid-resource language for current models: good, but not as solid as English. We do not promise quality upfront. The test set includes messages with and without diacritics and chats that switch to English, and we measure each language against its own threshold before the pilot goes live.
Yes. You are the controller and we are your processor, so we sign a DPA under Article 28 and the EU standard contractual clauses, because Kazakhstan has no adequacy decision. Model keys and hosting sit on your account in the EU, and we test on anonymised samples instead of pulling data to Kazakhstan. Under the OpenAI and Anthropic API terms, requests are not used for training by default. Since 2 August 2026 the bot tells people they are talking to AI, as Article 50 of the AI Act requires.
Yes, through the SPV access your accountant already has. AI downloads incoming e-Factura invoices, matches them against orders, receipts and the bank statement, and marks mismatches before anyone pays. It can draft a rejection message or a reply to an ANAF notice, and your accountant approves it. Filing D406 and replying to ANAF stay with people.
We agree on the success criterion before the start. Usually it is the share of requests handled without an operator, accuracy per language and response time. After the pilot there are three options: launch, rework or stop. Each stage is paid separately, and the next one only if you continue. When the AI is unsure, a person decides the disputed case.
The code, the knowledge base, the test set with reference answers, and documentation. Model keys and accounts are in your company’s name, so the system can keep evolving with us or without us.
Not if you have our tax residence certificate before the first payment. The Romania–Kazakhstan treaty taxes our service profits only in Kazakhstan. ANAF asks for the original or a notarised copy with a certified Romanian translation, valid for the calendar year, so we renew it every January. The contract is for services, not a licence, which keeps the payment out of the royalties article.
Read before the call
- How an AI agent checks documents Matching invoices and contracts against your rules, and what stays with a person
- How AI answers customers in WhatsApp Spam filtering, routine answers, estimates, and handing real buyers to staff
- AI pilot in 30 days A day-by-day plan and the three decisions a pilot ends with
- AI business process automation Where workflow rules are enough, where AI earns its place, and how to do the ROI math
- What to prepare before implementing AI A company checklist and the minimum package for a vendor
Let’s talk
Tell us which process you want AI to take over, which systems hold the data (SmartBill, Saga, your CRM) and how customers reach you. Attach a few real chats or invoices if you have them. We will tell you which solution fits and where to start.
I read every request myself.