AI for retail, e-commerce and customer support
Vlad Tudor is an AI consultant and engineer in Bucharest who helps retailers, online shops and support teams decide what AI should answer and what should stay with a person. At the start of the LLM era he built one of the first production customer-support systems whose generated answers cite their internal source, for a startup serving retail brands.
Reply within 24 hours · no sales pitch
What I hear from support teams.
“Half the tickets are ‘where is my order’ and ‘how do I return this’, and my best people spend the day on them.”
“We tried a chatbot. It answered confidently and got the returns window wrong.”
“It takes months before a new agent knows the policies well enough to work alone.”
What the evidence actually says.
The only peer-reviewed field study of AI in customer support found that an assistant raised productivity by 14% on average, and by 34% for the least experienced agents, with little effect on the best. That is a more honest promise than “the bot will handle 70% of tickets”: AI lifts the bottom of the team towards the top.
The opposite lesson comes from Klarna, which announced in 2024 that AI was doing the work of hundreds of agents and in 2025 said, through its own CEO, that it had focused too much on cost and quality had suffered. It began hiring people again. The volume target was met; the customer wasn't.
From the REWIRED podcast.
What I'd do with you.
- 01
Answers from your own policies.
A retrieval system over your returns, delivery, warranty and product information, so every answer is traceable to the paragraph it came from and changes when the policy changes.
- 02
An assistant for your agents first.
Drafted replies an agent checks and sends. It's the fastest way to see where the answers are right before any of them go to a customer unsupervised.
- 03
Routing and triage.
Recognising what a message is about, in several languages, and sending it to the right queue or the right standard resolution.
- 04
Onboarding.
New agents asking the policy base instead of the colleague next to them.
- 05
A clear line for escalation.
Complaints, refunds above a threshold and angry customers go to a person, with the conversation summarised.
What I've built.
Multilingual routing, per brand
In 2022–23 I worked on a customer-support system for a startup serving retail brands, each with its own catalogue and policies. The first phase was a multilingual model that recognised intent and routed conversations to resolutions adapted per brand.
Retrieval, in under four weeks
When ChatGPT launched, the architecture was redesigned and a first retrieval prototype built in under four weeks, at a time with no dedicated vector-database tooling. It became one of the first production support services answering with generated text that cites its internal source, roughly two years before that became the norm.
Access to an AI assistant raised support productivity by 14% on average and 34% for the least experienced agents, with minimal effect on the most experienced.
“We focused too much on cost. The result was lower quality.”
Send me the three questions your customers ask most. I'll tell you which ones AI should answer, and which it shouldn't.
A few short questions. The first conversation is free.
Reply within 24 hours · no sales pitch
What an engagement can cover.
A first conversation
About the problem, before any technology. Free.
An audit of processes, data and tools
How the work runs today, what data exists, which tools are already paid for.
A prioritised roadmap
What to build, what to buy and what to skip, in order, with the reasoning written down.
The build, or a clean hand-off
Hands-on build through Sapio AI, or a roadmap your own team can take from there.
Team enablement
A workshop for the people who will use the result, so it gets used.
Follow-up
A check on what changed once the work is in place, and what comes next.
For support teams: the audit reads a sample of real tickets and the policies the answers depend on (returns, delivery, warranty); follow-up measures on the support team's own numbers.
Related pages.
AI consulting for companies in Romania
How an engagement runs, step by step, and what it can cover.
AI process automation for operations teams
Manufacturing, energy, distribution and back office: fix the process, then automate it.
AI consulting for startup and SaaS founders
What to build, what to buy and what to leave out, before money is spent.
Questions that come up.
Will an AI chatbot replace my support team?
Unlikely, and the companies that tried it publicly have walked it back. What works is taking the repetitive part off the team and giving every agent the same access to the policies your most experienced agent has.
Do I need a custom system or an off-the-shelf chatbot?
It depends on how many policies, brands and languages you run, and on how much a wrong answer costs you. A small catalogue with a clean FAQ can live on a ready-made tool. Several brands, systems and exceptions usually can't. The audit answers this before anything is bought.
How do you stop it from inventing answers?
By making it answer only from your documents, cite the passage it used and hand over to a person when the documents don't cover the question. Then testing it on a set of real past tickets before it talks to a customer.
Does it work in Romanian and other languages?
Yes. The system I described above was multilingual from the first phase.
What has to be ready on our side?
Policies that are written down and current. If the returns window lives in three places with three different numbers, that gets fixed first.
Tell me what you're working on.
A few short questions, two or three minutes. The reply comes from me, within 24 hours on working days.
Reply within 24 hours · no sales pitch
Help with
What would you like help with?
Pick all that apply.