What does Veles Digital build?

Custom AI tools, automation pipelines, conversational chatbots, and analytics dashboards. We engineer practical AI systems that eliminate waste, accelerate growth, and ship. Not demos. Every project is shaped around the specific workflows of a client's business.

How long does a custom AI project take?

A typical first-version delivery is two to eight weeks. Discovery and specification take about a week, a focused MVP runs two to four weeks, and larger integrations or analytics dashboards extend to six to eight weeks. We ship working software in short cycles rather than long-lead waterfalls.

Do you work with non-technical clients?

Yes. Most of our clients are not engineers. We translate business problems into AI solutions, handle the technical complexity end-to-end, and deliver tools your team can actually use. Onboarding includes training and clear documentation.

What industries does Veles Digital serve?

We're industry-agnostic. The common thread across our clients is teams who want concrete AI-driven outcomes like automation, better analytics, and customer support efficiency, rather than experimental AI projects. Each engagement is shaped around your specific workflows and operational reality.

How does a project start?

Book a free 30-minute discovery call via calendly.com/velesdigitalsolutions/30min, message us on WhatsApp at +381 69 251 1196, or email contact@veles.digital. We discuss your business, identify which workflows AI can improve, and propose a scoped first project with a fixed price and timeline. We respond within 24 hours.

Where is Veles Digital based and who do you serve?

Veles Digital is headquartered in Serbia and serves clients across Europe and worldwide. We operate fully remotely with English and Serbian language support, and most engagements happen via async tooling and scheduled video calls.

How much does custom AI development cost?

We don't publish fixed prices because every project differs in scope, integrations, and timeline. The discovery call is free, runs 30 minutes, and ends with a scoped quote based on your specific needs. Typical engagements range from focused two-to-four-week MVPs to six-to-eight-week multi-system integrations, with fixed scope and fixed price for the agreed phase.

What's the difference between Veles Digital and a ChatGPT subscription?

ChatGPT is a general-purpose assistant; Veles Digital builds production AI shaped around your specific workflows, integrated with your data, deployed on your own stack. ChatGPT helps individuals draft emails. We build systems that automate your support inbox, qualify your inbound leads, surface insights from your sales calls, or pull together a board-ready dashboard from your real-time data. Owned by you, not rented per seat from OpenAI.

Which AI platforms and models do you build on?

We're model-agnostic. We work with OpenAI (GPT-4o, GPT-4.1), Anthropic Claude (Sonnet, Opus, Haiku), Google Gemini, Mistral, and open-source models (Llama, Qwen) depending on the use case. Selection is driven by cost, latency, accuracy, and data-residency requirements, not vendor loyalty. We also use vector databases (Qdrant, pgvector, Pinecone) and orchestration frameworks (LangChain, custom agentic patterns) where they make sense.

Can you integrate AI with our existing CRM, ERP, or data warehouse?

Yes. Integration is the core of every project. We connect AI capabilities to Salesforce, HubSpot, Pipedrive, Notion, Airtable, Postgres, MySQL, BigQuery, Snowflake, and most modern SaaS APIs. If your stack has a documented API or webhook surface, we can wire AI into it. We don't ask you to migrate your data to a new platform.

Do you provide ongoing AI maintenance and updates?

Yes. After delivery, we offer optional retainers covering production monitoring, model upgrades, prompt tuning, and feature evolution. Most clients move to retainer after the initial build to keep their AI current as models and APIs evolve. What worked on GPT-4o today may be obsolete or expensive six months from now, and a small ongoing engagement keeps the system competitive.

How do you measure AI ROI for our business?

Every project starts with a measurable target like hours saved per week, response time reduction, conversion lift, support ticket deflection, or qualified leads per month. We instrument the AI system to report on those metrics from day one. If we can't measure the win, we don't ship it. Discovery includes a baseline measurement so the impact is provable, not anecdotal.

How we built a WhatsApp AI agent for a B2B wholesaler

In December 2025 we shipped a WhatsApp AI agent for a B2B mobile-parts wholesaler in Serbia. The owner was personally answering 80-120 WhatsApp messages a day from buyers across the country. He could not scale the business because he was the bottleneck. Four weeks later, the AI was reading every inbound message, proposing a daily promo, taking orders from approved buyers, and routing only the genuinely complex conversations to a human. The owner went from 5 hours a day in WhatsApp to 30 minutes.

This is the story of how we built it, what we got wrong on the way, and what a custom WhatsApp AI agent actually does that off-the-shelf chatbot platforms cannot.

The problem the owner actually had

Most "we need a chatbot" pitches turn into Intercom installations. This one was different. When we ran Discovery, three things stood out:

The buyers were repeat B2B customers. They knew what they wanted, used Cyrillic and Latin Serbian interchangeably, and expected fast pricing for irregular SKU combinations.
The owner was making one strategic decision per day: which slow-moving inventory to promote. The rest of the work was repetitive. Quote, confirm, hand-off, repeat.
WhatsApp was non-negotiable. Buyers wanted WhatsApp. The owner wanted WhatsApp. Migrating them to a web app or a "customer portal" would have lost the relationship, not improved it.

A scripted chatbot was wrong for this. A web app was wrong for this. What the owner actually needed was a senior salesperson who lived in WhatsApp, knew the catalog cold, and could make pricing decisions inside owner-approved limits.

The design: AI proposes, humans approve

We built around a constraint that has become a Veles signature: AI proposes, humans approve any revenue-critical decision. Specifically:

The AI reads every inbound WhatsApp message in real time. It classifies intent (price quote, order, support question, off-topic), looks up the buyer's history, and drafts a response.
The AI proposes the daily promo: which slow-moving SKUs to push, at what discount, with what margin. The owner reviews this once a day and approves or edits. Total time: 5-10 minutes in the morning.
The AI takes orders within owner-approved limits: standard pricing, standard delivery, known buyers, in-stock SKUs. These flow straight to the back-office system without human review.
The AI escalates the rest: new buyer signups, custom pricing requests, low-stock substitutions, complaints. The owner sees a summarized conversation, not a transcript dump, and can reply in one tap.

In short, the system inverts the time budget. The owner used to spend 5 hours answering WhatsApp and 30 minutes deciding strategy. Now he spends 30 minutes deciding strategy and the system answers WhatsApp.

The technology choices, briefly

We used the official WhatsApp Business Cloud API (not third-party scrapers or unofficial wrappers, which both fail at scale and risk number bans). The AI layer is a custom orchestration of frontier language models with structured tool calls into the wholesaler's existing order system. Inventory and pricing data flow through a small in-house API we built. Nothing exotic, nothing rented from a no-code stack with a markup.

The two technology decisions that mattered most:

Latin and Cyrillic Serbian, plus English. The buyers code-switch mid-message. The AI handles all three natively because the underlying models do. We did not have to train anything custom for this.
Tool use over pure prompting. The AI does not "decide" how much to charge. It calls a pricing function with structured arguments, gets back a number, and quotes that number. Same for inventory checks. This makes the system auditable and stops the model from hallucinating prices.

What we got wrong (and fixed)

Two things broke in production that we did not catch in dry-runs.

The 24-hour customer service window. WhatsApp Business API distinguishes between freeform messages (allowed inside a 24-hour rolling window from the customer's last message) and template messages (allowed outside it). We initially built the agent to send freeform messages always, and the API started rejecting after-hours follow-ups. Fix: classify whether we are inside or outside the window, and route to approved templates when outside. Cost us a day, no customer impact because we caught it during the soft-launch phase.

Buyer disambiguation. Several buyers shared phone numbers across multiple businesses, or used personal numbers for their wholesale ordering. The AI initially attached orders to whichever business name first matched, which was sometimes wrong. Fix: explicit buyer confirmation on the first message of every new conversation. Adds 5 seconds at the start, eliminates the ambiguity. The owner had been doing this implicitly for years; we just had to encode it.

Both of these are typical for production AI. The reason we caught them in week one rather than month three is that the scoping process had a two-week post-launch monitoring period built in. Custom AI is never "done at launch" — it is launched and then tuned for a sprint.

The result, in numbers we can share

Daily WhatsApp time for the owner: from ~5 hours to ~30 minutes
Orders handled without human review: ~80% (the easy ones)
Orders routed to the owner with full context: ~20% (the high-value or unusual ones)
New buyer signups handled by the agent: previously zero, now most of them
Project timeline: 4 weeks from kickoff to production, plus 2 weeks of close monitoring

The full case study lives on the Mobilni Market portfolio page. It has the screenshots and a more detailed breakdown of the agent behavior.

When a WhatsApp AI agent is actually the right move

A custom WhatsApp AI agent makes sense if all three of these are true:

WhatsApp is where your customers already are. Forcing them to a web portal will lose them. Doubly true for B2B in Serbia, the Balkans, and most of Europe outside the Nordics.
The bottleneck is a person, not the tool. If a no-code platform would solve it, use no-code. Custom agents are for problems where judgement, multi-language understanding, or system integration are required.
There is a clear human-in-the-loop checkpoint. Pure full-automation in WhatsApp customer-facing work is risky and usually wrong. The agent should propose, the human should approve revenue-critical actions.

If those three are true, custom is meaningfully better than off-the-shelf. If only one or two are true, you can probably get there with AI automation or a focused AI agent in a different channel.

What this kind of build typically costs

We do not publish tiered pricing because every WhatsApp AI agent is different. A focused agent on one channel for one buyer segment with light integration is at the lower end of our range; a multi-tenant agent across channels with on-premise inference and a back-office CRM rebuild is at the higher end. After a 30-minute discovery call you get a real fixed-price estimate, not a "starting from" number. The full methodology lives on our scoping page.

If your business runs on WhatsApp and your bottleneck is you, we should probably talk. Book a 30-minute discovery call and we will tell you on the call whether a custom WhatsApp AI agent is the right fit, or whether something simpler will save you money.