Where the data lives
The first step is a list of places where company information lives. Usually these are the CRM, the 1C accounting system, email, messengers, telephony with call recordings, shared folders and the personal spreadsheets of staff.
For each place we note three things: what is stored there, who owns it and whether you can get an export. If you can’t get an export, AI can’t connect to that information.
We also look for important things outside the systems: a price list in a manager’s spreadsheet, agreements with a client in a messenger, return rules in the head of a senior employee. This has to move into a system before launch.
In AI due diligence, a full review of the company before AI decisions, this list of sources becomes the basis for every other step.
Completeness and quality
We check completeness on a sample: we take a sample of client, deal or request records and see how many fields are filled in, whether there are duplicates and whether the numbers match between systems.
What we look for: one client entered three times under different names, deal statuses not updated for months, stock in the accounting system that doesn’t match the warehouse, calls with no recording.
Quality matters more than volume. An AI agent that answers clients needs a current knowledge base and the history of correspondence, while a purchase forecast needs a sales history without gaps. The requirements depend on the task, so we check the data for a specific process.
We also check freshness: a knowledge base that nobody updated for a year will teach an AI agent to answer with old prices and terms.
The result of this step: for each process, a mark of “ready”, “needs cleaning” or “collect from scratch”.
Access
We check who has access to what today: how many employees see the whole client base, who can export it, and whether accounts of former staff and former contractors are still active.
For AI, the rule is the same as for an employee: access only to what its task needs. A support AI agent sees the history of requests, but not finances or salaries.
Accounts and access must belong to the company, not to the contractor. In our projects, the data and the access stay with the client.
We write AI access into a separate list: which agent, which system, which rights and who granted them. When you change contractors, this list shows what to switch off.
How AI can lead to a leak and how to close it is covered on the page about data leaks when working with AI.
What you may give AI under 152-FZ
Clients’ personal data: names, phone numbers, addresses, purchase history. Under 152-FZ, Russia’s personal data law, it is processed only on servers in Russia and with a lawful basis, such as the client’s consent.
We split all company information into three groups.
1. You may give it to AI without limits: price lists, catalogs, company rules, the knowledge base, anonymized statistics.
2. You may give it only with processing in Russia: personal data of clients and staff, correspondence, call recordings.
3. Not given to outside services without a separate decision by management: trade secrets, financial statements before publication, passwords and access keys.
The check shows where information already goes today: which employees paste client correspondence into foreign chatbots, and through which services call recordings pass. The legal requirements are covered in the article on 152-FZ and AI agents, and how we protect client information is described on the page about security.
What to do if you have little data
Little data is no reason to give up on AI. Some tasks run on the model’s general knowledge plus your rules and price list: answers to common questions, draft emails, sorting of incoming requests.
If a process lacks information, we start collecting it before the rollout: we add required fields to the CRM, turn on call recording, and move rules from employees’ heads into a knowledge base.
Sometimes the first AI solution is the collection itself: an AI agent sorts incoming emails and calls and enters them into the CRM. In a few months, the company builds up a history on which the next solutions rest.
In the map of AI solutions, such processes don’t come first: what works on the existing information goes first, and collection for the next steps runs in parallel.
We decide by the numbers: how long collection will take and how much the process loses while the data is missing.