Service
AI and process automation
The question is not “should we use AI” but “who does this work, and does it actually require a human”. Once we answer that, the solution is usually small and inexpensive.
Where it genuinely pays off
These are not trends — they are jobs where the time saved can be measured.
Document reading and extraction
Pulls fields out of invoices, waybills, contracts and forms and writes them into your system. Manual entry drops from hours to minutes.
First-line customer answers
An AI agent answers repeat questions from your own content and hands the rest to a person. The goal is not to remove humans, but to remove repetition.
Classification and routing
Incoming requests, emails and tickets reach the right team automatically. A request sitting in the wrong queue is a lost request.
Content production support
Product descriptions, category copy and translation drafts. Published only after human review — unreviewed output is a brand risk.
Search across your own data (RAG)
Ask questions in plain language over your internal documents. “Which document covers this procedure” takes seconds instead of an afternoon.
Process automation
Work done by hand that needs no AI at all: moving files, generating reports, shuttling data between systems. Often the highest-return step of the whole project.
First we ask whether AI is needed at all
Before handing a task to a model we ask three questions: how often does this repeat, what does an error cost and how will accuracy be measured. Without clear answers to all three, the project stops at the demo stage.
For most requests the answer is simpler: a rule-based automation or a proper integration between two systems does the same job more cheaply and far more predictably. Saying that is against our short-term interest — we say it anyway, because an AI project that does not work costs more than one that was never started.
When it genuinely fits, the setup looks like this: narrow scope, a measurable success criterion, a human in the loop, and a reversible architecture. If the model changes or costs rise, you must be able to switch provider.
Data security — put it in the contract
This is the most commonly skipped topic in AI projects, and the most expensive one to skip.
- Is your data used for model training? Enterprise providers let you disable this; get the fact that it is disabled in writing.
- In which country is the data processed? Cross-border transfer is a separate legal question under GDPR and Turkish KVKK.
- Which data actually reaches the model? Personal data should be masked; “just send everything” is the short path to a breach.
- What happens when it answers wrongly? Every automated reply that reaches a customer needs an owner and a way to roll it back.
- Are interactions logged? What was asked and what was answered must be auditable — for quality and for law alike.
Frequently asked
How long does an AI project take?
A narrow solution (document extraction, or answering frequent questions) is usually running within 2–5 weeks. What stretches the timeline is not the model but access to the data and agreeing what “accurate enough” means. With those two ready, it moves fast.
Does our data go to the AI provider?
It depends on the architecture, and we agree it in writing up front. There are three options: processing through an enterprise API (data is not used for training, guaranteed contractually), running an open-source model on your own server (data never leaves, hardware cost is higher), or a hybrid setup where sensitive fields are masked. Which one fits depends on the sensitivity of your data and your budget.
Can we drop our support team once we have a chatbot?
No — and read any proposal that suggests otherwise carefully. The realistic goal is that a share of repeat questions is answered automatically and the rest reaches the right person with context attached. Setups that cut humans out entirely lose the customer at the first difficult question.
How do we know it worked?
We set the metric before starting: manual data-entry time, first-response time, or the share of requests escalated to a human. We measure before and after. AI projects that start without a metric are projects that end with a pleasant demo.
Will it work with our existing systems?
It needs a connection point: an API, database access or a regular file transfer. If one of those exists, integration is possible. Closed systems can be worked around, and this is the first thing we check during discovery. The same team handles the systems integration side.
Which task should we automate?
Tell us where your team loses the most time on repetitive work. We will measure it: if there is a real gain we build it, and if there isn’t we say so.