Custom AI development for your business — chatbots, smart agents and automations that understand text
Your dispatcher gets the same 100 messages a day: 'where's my parcel?', 'do you still have stock?', 'how much does it cost?', 'dental appointment Saturday'. Your support team loses 4 hours daily on repetitive WhatsApp and email replies. An AI assistant trained on your company's knowledge can take over 60-80% of these intelligently — not with stupid menus like old chatbots, but with real answers, in Romanian, connected to your data (stock, orders, status, bookings). I build custom AI agents — WhatsApp and on-site chatbots, email assistants, automations that extract data from invoices or requests, Claude and OpenAI integrations into your existing software. The most concrete result: the WhatsApp AI agent in colet.app, based on Claude Haiku, that turns customer messages into real orders without involving the dispatcher — part of the ecosystem with 1000+ orders processed weekly.
The real problem: generic AI knows nothing about your business, and old chatbots are stupid
Pause for a second and think about your work day. How many hours does your team waste answering the same 50 repeated questions? Customers asking what time you open, whether a product is still in stock, whether you accept card payment, whether you do home visits, whether tomorrow's appointment is available. Simple questions, with answers somewhere in your company — on the site, in a file, in your head — but consuming real people for hours. At 5-10 people, it's bearable. At 20-30 with constant volume, it becomes the team's main complaint: 'I'd rather do anything else than answer messages'.
The solutions you find on the market are split into two categories, both flawed for a small or mid-size business. The first are old chatbots — Manychat, Tidio, Intercom basic — which work on menus (press 1 for hours, 2 for products, 3 for contact). They frustrate customers, don't understand real free-form questions, and if someone writes 'do you still have the red chair I saw last week?', the bot says 'didn't understand, choose from menu'. The customer closes and calls a competitor.
The second category is generic AI platforms like public ChatGPT or Microsoft Copilot. They sound impressive — 'artificial intelligence, talks naturally' — but have a huge problem: they know nothing about your business. Public ChatGPT can tell you the capital of France but doesn't know how many rooms your guesthouse has, doesn't know that you ship free over 200 RON, doesn't know your holiday hours. And if you upload all company data into a public ChatGPT account, you violate GDPR and lose control over who sees what. There has to be something else.
That something else is a custom AI agent — a smart assistant (based on Claude, GPT-4 or another large language model) built around your company's specific knowledge and connected to your real data (management software, databases, calendar, inventory). It answers naturally in Romanian, understands free-form questions, has access to up-to-date info (not a static FAQ from 2023), and the conversation feels like talking to an informed employee. If you ask 'do you still have the red chair for the weekend?', the agent checks real stock and answers concretely: 'Yes, 3 left, free delivery by Friday'. For colet.app I built exactly this on WhatsApp — the agent turns customer messages directly into structured orders in the software, without involving the dispatcher.
A good AI agent doesn't replace the team — it frees the team from repetitive work so they can focus on what matters (complex cases, sales, big customer relationships). Everything that follows on this page is about how such solutions are built concretely for your business, with real proof (colet.app's WhatsApp agent), pricing tiers by complexity, and what you can ask before deciding.
What types of custom AI solutions I build
The list below covers the main types of AI projects I get. Almost no project is just 'chatbot' or just 'automation' — most are combinations reflecting your company's processes. For colet.app I built a WhatsApp agent that understands free-form delivery addresses in Romanian and creates orders automatically in the software.
Smart WhatsApp chatbot (like colet.app's)
AI assistant that talks on WhatsApp Business with your customers. Answers frequent questions with real data, takes orders or bookings, hands complex cases to a real person. For colet.app, the agent turns delivery address messages directly into orders in the software — with dispatcher confirmation before execution.
On-site chatbot with your company's knowledge
AI assistant integrated into your website that answers questions about products, services, prices, delivery, warranties. Knows your catalog, hours, policies, FAQ. Unlike old menu-based chatbots, it understands free-form questions and answers naturally. Technically uses RAG (knowledge retrieval from your database) for fresh data.
AI email assistant — automated replies for repetitive requests
For companies receiving 30-100 emails daily with the same 10 question types (standard quote requests, order status, product questions), I build an assistant that reads incoming emails, classifies them, generates replies in your style and sends them for approval in 30 seconds. The team approves with a click, AI never sends alone — you keep control.
Automated data extraction from documents (invoices, orders, forms)
For companies receiving PDF invoices from suppliers, free-form email orders, or scanned forms, I build an automation that automatically extracts structured data (supplier, amount, VAT, items) and uploads it to your management or accounting software. A secretary spending 2 hours/day on data entry now spends 15 minutes, and errors drop by 80%.
AI-generated SEO and marketing content
For companies needing constant content publishing (blog, product descriptions, Facebook/Google ad copy), I build a semi-automated system using Claude or GPT-4 to generate drafts in your company's style, with real keyword research. You review and approve, AI produces 5-10x more in the same time.
AI analysis of customer reviews and feedback
If you have reviews on Google, Booking, Trustpilot, Facebook, or internal feedback forms, I build an AI analysis that processes all texts monthly/weekly and tells you: what customers say about you, the top 3 complaints, the trend over the last 90 days, what to actually do. That's more valuable than a 4.7-star rating without context.
Voicebot for phone (limited in RO currently, but feasible)
Voice bot answering phone calls in Romanian — for companies with high volumes of repetitive calls (medical clinic receptions, booking services, retail support). Technically harder than a text chatbot (Romanian voice recognition still imperfect, but viable for specific cases), but I build it if the scope justifies.
LLM integration into your existing software (Claude API, OpenAI)
If you already have a custom management system or website but want to add AI intelligence at specific points (quote summarization, proposal drafting, contract analysis, internal team assistant), I integrate a language model directly (Claude via Anthropic API, GPT-4 via OpenAI API, or open-source alternative) as a module in the existing software.
Bonus on demand: Light fine-tuning for specific cases (your company's terminology, industry jargon), conversation monitoring with monthly reports (what customers ask, where AI fails), token rate limiting for cost control, smart fallback to a human operator when AI isn't sure, GDPR compliance with data isolation on your servers or in the EU.
Pricing and AI costs
Custom AI solutions have two cost components: initial development (fixed price, paid once, set after requirements analysis) and monthly usage cost (AI tokens consumed per conversation — variable but predictable). I'm transparent about both before we start. For orientation, token costs for a chatbot with 1000 conversations/month are under 100 RON/month using Claude Haiku or GPT-4o-mini.
Why me, not an AI agency or a SaaS chatbot platform
There are agencies selling 'AI transformation' with 50-200k RON budgets. There are also generic platforms (Manychat, Tidio, Drift) with 100-500 EUR monthly subscriptions. Here's what you get working with me instead, for a custom AI project that actually works.
I've already delivered an AI agent in production processing real orders
The colet.app WhatsApp AI agent is real proof, not marketing portfolio. Based on Claude Haiku (Anthropic's fast and cheap model), connected to the courier management software, it turns customer WhatsApp messages into structured orders (pickup address, delivery address, weight, contact) with dispatcher confirmation. Processes real volume in an ecosystem with 1000+ weekly orders. I can run a live demo, not just static screenshots.
Solo plus AI — no AI agency overhead
Romanian AI agencies often hire juniors who learn on your project, plus consultants selling 'AI strategy' with 60-slide presentations. I'm a single person who has already built AI agents in production, accelerated by Claude Code. You talk to whoever's building, get working code, pay once for development. The cost stays 30-60% lower than agencies on the same scope.
Custom around your company, not SaaS platform constraints
Manychat, Tidio and other chatbot platforms ask you to fit your processes into their predefined menus. I build the agent exactly around your processes (like colet.app, where the agent understands free-text Romanian delivery addresses — impossible on generic platforms). Plus the code is yours at the end, no 200-500 EUR monthly subscription to an American vendor.
Transparent about AI costs and limits
I tell you straight what AI can and can't do today. Voicebot in Romanian? Still limited, but feasible for narrow cases. Agent that understands any email? Yes, but with 5-10% error rate on ambiguous cases, so I recommend human approval. Unpredictable token costs? No — I add rate limiting and alerts so you don't get a 2000 EUR invoice out of nowhere. No 'revolutionary AI breakthrough' — just real tech applied to your real processes.
Real case: colet.app's WhatsApp AI agent — Claude Haiku connected to the software
The most concrete example I can show is the WhatsApp AI agent integrated into colet.app — the courier management software for Romania-UK routes. The agent runs in production, processes real customer messages daily, I can run a demo with permission from the user companies.
RO-UK courier companies received transport requests on WhatsApp in free-form text, like 'hi, I want to send a 15 kg parcel from Suceava to Birmingham, room this week?'. The dispatcher read each message, asked for additional details (exact address, recipient contact, dimensions), noted in Excel, transferred to the software. At 200 messages/day, the dispatcher became a total bottleneck — replied late, missed requests, made data entry errors. Generic WhatsApp chatbot solutions didn't understand free-text Romanian delivery addresses.
WhatsApp AI agent based on Claude Haiku (Anthropic's fast and cheap model), trained to understand free-form Romanian transport requests. The agent reads the customer message, automatically extracts structured data (pickup address, delivery address with Google Maps validation, weight, approximate dimensions, contact), asks if critical data is missing, generates an order summary, sends it to the dispatcher for one-click confirmation. Connected directly to the colet.app software — confirmation creates the order in the system with all structured data, ready for driver assignment. Full GDPR compliance (data isolated per company), conversation monitoring for monthly reports.
The dispatcher recovered 60-70% of the time previously lost on replies and data entry. The agent answers messages in under 10 seconds, anytime (nights, weekends), without blocking anyone. New requests turn into structured orders without manual steps. Token cost for real volume is under 200 RON/month using Claude Haiku — compared to the cost of 4 daily dispatcher hours previously doing the same work manually. I can run a live demo with permission from user companies.
How I work — the process from first conversation to launch
The process is clear and predictable. Here's how a typical custom AI solution project unfolds, from the first call to the moment your team enters the system daily and forgets manual replies existed.
Discovery call (60-90 minutes, free)
On Google Meet or in person. You tell me what your company does, what questions/messages you get daily that consume your time, what processes you have today, what you've tried before (old chatbots, public ChatGPT). I ask specific questions to understand if AI is the answer (sometimes it isn't — I tell you straight if classic automation fits better).
Analysis and written quote (3-5 working days)
After the call I take 3-5 days to think through the structure: which language model fits (Claude Haiku for speed, Sonnet for complexity, GPT-4 for specific cases), which data sources need connecting, what architecture. You get a written quote with: agent structure, clear stages with concrete deliverables, firm pricing per stage, monthly token cost estimate for your volume.
Confirmation and 50% deposit — start in 24-48h
If the quote works, we sign a simple contract (1-2 pages, no corporate lawyer pages), you pay 50% deposit, I start effectively in 24-48h. Repository setup, cloud infrastructure, first conversation prototype visible in the first week.
Iterative development with weekly testing
I build in 1-2 week stages with testing at the end of each stage. You and your team interact with the agent at a temporary URL, see where it works well and where it fails, we tune prompts and logic together. That's critical for AI — real quality only shows from real testing.
Gradual launch with close monitoring
Public launch isn't on/off. We start the agent on 10-20% of real traffic, monitor conversations the first 1-2 weeks, adjust based on real errors, then ramp to 100%. 30 days of free post-launch support for adjustments. After that, optional maintenance adapted to complexity, with monthly reports on conversations, token costs and improvement recommendations.
Frequently asked questions
How much does a custom AI solution cost — including monthly costs, not just development?
Two components: development (fixed price set after requirements analysis — simple chatbot smaller than a complex website, multi-channel agent larger, full system like colet.app the largest category) and monthly token costs (variable but predictable). For orientation: a chatbot with 1000 conversations/month costs 30-100 RON/month using Claude Haiku or GPT-4o-mini. A complex agent with multiple actions, 100-500 RON/month. Full quote with both components in 3-5 working days after the first call.
What happens when AI gets it wrong?
It does get it wrong. No language model is perfect — Claude, GPT-4, any of them. My strategy: on reversible actions (conversational reply, summarization), AI replies autonomously. On irreversible or expensive actions (creating an order with payment, sending an official email, modifying critical data), AI proposes but a human approves with a click. Plus conversation monitoring for monthly reports with all cases where AI was uncertain — that's how we identify improvement areas continuously.
Is it GDPR compliant? Customer data doesn't end up in public ChatGPT?
Yes, GDPR compliant. I use professional APIs (Claude via Anthropic API, GPT-4 via Azure OpenAI or OpenAI enterprise API) which have data protection contractual clauses — your data isn't used to train public models. For sensitive data cases (medical, financial), I can run open-source models on isolated EU servers. Privacy policy adapted to your sector is part of delivery.
Can I replace an employee with the AI agent?
Honestly? No. The AI agent doesn't replace a real employee — it frees the employee from repetitive work. The colet.app dispatcher who used 4 hours/day on WhatsApp replies now uses that time for complex cases (new clients, real problems, emergencies). Team efficiency rises significantly, but you still need real people for cases where AI isn't sure. AI is a tool, not a replacement.
How long does development take?
Depends on scope: simple chatbot on site or WhatsApp 4-6 weeks, multi-channel AI agent with integrations 6-12 weeks, full AI system with admin portal and automations 8-16 weeks. We start within 24-48h after I receive the deposit. Every 1-2 weeks you test a real version with your team, not just a mockup.
Can I see real proof before deciding?
Yes. I can run a live demo of colet.app's WhatsApp AI agent with permission from user companies — see how it receives a free-form Romanian message ('hi, I want to send a 15 kg parcel from Suceava to Birmingham'), how it automatically extracts structured data, how it validates the address with Google Maps, how it sends to the dispatcher for confirmation. Plus screenshots from other integrated AI automations. The demo is part of the initial discovery, free.
Read more
Let's talk — initial conversation is free
Tell me what questions/messages your team gets daily and what consumes your time. In 60-90 minutes I can tell you directly if AI is the right answer, what a concrete solution would look like and at what budget. No obligations, no aggressive follow-up.