I’m starting a series where I show how we are rebuilding Documentolog itself — almost from scratch — around AI. On our own processes, with the real pain points and how we solved them. The first case I want to start with is our technical support, because that’s where it hurt the most.
How it was
The more customers we had, the harder it became to keep support quality high — especially when the product itself is constantly evolving. Bugs and configuration errors became part of the daily routine, and that, frankly, did us no favours in our customers’ eyes.
When a problem came up, the request went down a chain: the user fills in a ticket through the support module in their account. Then that ticket has to be reviewed by the customer’s administrator — on average that took anywhere from 15 minutes to 8 hours. After that, if needed, the administrator escalated it to our second line, where it was handled by our technical specialists — which again took from 30 minutes to 7 days. On top of that, requests often bounced back for clarification.
Customers could wait days, sometimes weeks, for an answer. Meanwhile the team was drowning in routine, repetitive questions. So in most cases the customer simply called a manager directly and tried to solve the issue personally, bypassing the process. That much regulation and bureaucracy was a necessary evil — it was the only way to keep a minimum of quality and control.
What we did
We’ve been working deeply with AI for more than two years now, so we understand well that AI transformation is not “bolting AI on the side” and not a chatbot on claude or chatgpt that answers in generalities without the full context. It is a complete redesign of the process and embedding the agent inside it.
So we started by analysing the current process and designing a new one, using our d8n.ai platform to build AI agents. That’s how the d8n Support agent came to be — the one we announced this week and rolled out to all our SaaS customers who have already moved to d8n.ai.
Routine requests are now handled by the agent instantly: it knows the knowledge base, answers right in the chat, and turns a complex question into a structured ticket that it hands off to specialists, then monitors progress and notifies the user. We also upgraded our mobile app, which now has a corporate chat and the d8n Support agent, available 24/7.
On security and hallucinations
This is a critical question. Access to a customer’s internal information should belong neither to the agent nor to our own staff. That’s why d8n Support runs on local models in a closed contour: data never leaves the perimeter. It sees only service information — status, routing stage, registration number — but not the content of your documents. And it does nothing on your behalf: it doesn’t sign, approve or change settings.
To minimise hallucinations we built an architecture in which the AI agent for each customer instance runs on its own independent knowledge base, trained only on that customer’s documents and configurations. We also developed a golden-standard base that the agent continuously checks its answers against. Separately, we built a process in which this golden standard is constantly updated and enriched from our users’ feedback. As a result, the agent’s knowledge is not static — it lives alongside the company and the customer. A lot of work went into this, work that simply wasn’t possible before — without AI and the capabilities of d8n.ai.
The effect and ROI
Our goal is to make technical support through d8n Support fully self-sufficient. The agent should take the routine off the team entirely. We have already reduced the support team headcount by 70%. That does not mean AI replaced people: those employees were moved to other tasks. The routine has been automated, and the team shifted to complex work where a human is genuinely needed. And we deliberately do not delegate everything to the agent 100% — a human oversees its work. That’s the right thing to do, both for quality and for trust.
Takeaways
A mature methodology, technology and platform are very important for a successful AI transformation. But even more important is the role of AI champions inside the company — people with enough resources to drive genuinely difficult reforms. Real AI transformation has to change the processes themselves, so that AI doesn’t just answer nicely in a chat but is embedded in a real process, performs routine actions under tight control, and delivers a measurable result.
In the next issues I’ll keep sharing the approaches to AI transformation, the processes and the AI agents that we are living through first-hand. If you’re on this path too — tell me in the comments which process you started with, and what pains or questions you ran into.