"We don't know where to start."
There are dozens of ideas, and it is unclear which one would actually work.
"Our data can't leave the building."
Customer contracts, audit obligations, trade secrets. The constraint is real.
"Departments are already using their own tools."
This is what happens when no decision is made. Enterprise data leaves with no record of it.
We don't sell a large transformation program. We start with a two-week discovery. Through department interviews, we identify which task is worth solving with AI, and write down the conditions under which your data can be used.
You end up with one recommendation, with reasoning behind it. Moving forward is your call — the report and all output stay with you either way.
When this requirement is handled as a single yes/no decision, you're left with two bad options: everything goes, or nothing goes. We split the data into four layers instead.
Layer 0
Prices, costs, contracts, technical drawings
Never leaves — search happens inside the organization
Layer 1
Customer name, part number, order number
Goes out encoded; the external service sees a meaningless label
Layer 2
Monthly and shift-level summaries
No individual records — only aggregated data leaves
Layer 3
Catalogs, procedures, standard text
No constraint needed
Which of your data falls into which layer gets written into a table, in your own terms, during the discovery. That table becomes a document you can show an auditor.
The model never accesses the database directly. It only calls read-only, parameterized, pre-defined functions. Which function was called with which parameters is always on record.
Enterprise knowledge doesn't live in one place. Alongside the ERP there is accounting software, SharePoint and file servers, email, spreadsheets and documents, the CRM, internal communication, and third-party data. In most companies, the most critical information is still sitting in a spreadsheet.
The discovery covers all of it. What matters is not whether a source is the "official system" — it's whether the work actually happens there.
2 weeks
Department interviews, data inventory and layering, a regulatory assessment, scenario prioritization, and a written report.
4 weeks
On top of the above, we test whether the chosen scenario actually works on your own data, through an example your own team tries out.
7 weeks
The chosen scenario is handed over as an application your users can run in their daily work, including security documentation and training.
The discovery is a complete piece of work on its own. You are not obligated to continue — the report and output stay with you.
I spent 13 years working on Oracle and enterprise data — at Uzmar Shipyard on IFS and Oracle, and later at Vodafone on a multi-country CRM and revenue platform. On the AI side, I hold Anthropic certifications and have published work in the field.
The hard part of this work was never the model — it is accessing enterprise data safely and in an auditable way. That is what I have been doing for years.
Barış Akverdi — View on LinkedInIf a Sales Rep Can't See the Cost, Can They See It by Asking AI?
A sales rep who can't open the cost screen in the ERP can still get that number from the AI assistant. The problem isn't the AI — it's how it was set up. Here's the fix.
Read articleWhere Should You Start With AI? First, See Where You Stand Today
The first question about AI isn't "what should we do?" but "where are we today?". How to assess your company's current position across six dimensions, from usage to ownership.
Read articleA discovery service that identifies which task genuinely has a case for AI without your enterprise data leaving the organization. The chosen scenario is then made to work while keeping your data in-house.
There are three options: Discovery (2 weeks), Discovery and Proof of Concept (4 weeks), and Discovery and Pilot (7 weeks). These are calendar durations that leave room for scheduling delays on the client side; for example, the smallest package, Discovery, is about 5 person-days of work. Discovery is a complete engagement on its own; you do not have to continue, and the report and all outputs stay with you either way.
We split data into four layers. Prices, costs, contracts and technical drawings never leave; customer names and part codes go out encoded; monthly and shift summaries go out aggregated; catalogs and standard text need no restriction. During discovery we map your own items to these layers in a table.
No. The model never touches the database directly; it can only call read-only, parameterized, predefined functions. Every call is logged and no free-form queries are generated.
No. Besides the ERP it covers accounting software, SharePoint and file servers, email, Excel and Word files, CRM and internal messaging. What matters is not whether a source is an "official system" but that the work actually happens there.
In the half-hour call we do not present. We ask which task keeps repeating, which report is prepared by hand, which data can never leave, and what has been tried so far.
We don't do a pitch. We ask four questions: which task keeps repeating, which report is still built by hand, which data can never leave under any condition, and what has been tried on this so far.
If something concrete comes out of it, we talk about the next step. If it doesn't, you've lost half an hour.