<- All cases

Security Code case

Security Code is a Russian developer of information-security products. We are building the company an AI agent on top of its records: 1C, warehouse, finance. The project is in progress: the technical part is ready, tuning on the company's data is under way.

The task

The records of a large company accumulate for years: 1C, warehouse, finance, and the processes around them. There is a lot of data in there, and an answer to an ordinary work question has to be assembled by hand — across the system, spreadsheets, and colleagues. The task sounds simple: give an employee an agent that answers such questions on its own, from the company's own records.

The catch is that records alone do not close it. Part of the process knowledge is written down nowhere and lives with the staff: which exceptions exist, how to read a particular entry, why something here is done off the instruction. Without that layer an agent answers formally correctly and misses the point, so a separate part of the work is figuring out how to collect and keep that knowledge.

What is done

  • Integration with 1C: the agent reaches the company's records.
  • Our own tools on top of low-level operations: the agent works with actions it understands rather than with the raw accounting system.
  • Installation and updates under corporate requirements — that part is handled by the DODVIR infrastructure.
  • A knowledge base and the pipelines that train it: answer accuracy depends on them.

The hard part turned out to be the data and the domain, not the integration. Years of records in 1C include entries that look alike and entries that contradict each other. On top of that the agent is no professional in the product range: where a specialist sees two different products, it sees two similar names. So the main work right now is the knowledge base and the tools that keep the agent from guessing.

Immediate steps

Within the first pilot:

  • Feedback forms on the agent's answers: an employee marks where an answer is wrong and what the right one is. Ships in the next release.
  • Pipelines where agents extend the knowledge base from accumulated experience, deciding themselves what is worth recording.
  • Multi-agent schemes for updating the knowledge base — an experiment for now.
  • Raise answer accuracy to a level people can lean on in daily work.

Second pilot: an agent for every employee

It is planned after the first one and works with people and their agreements rather than with records.

  • Every employee has one agent that knows their context well: tasks, agreements, the history of decisions. In effect a virtual copy of that person for work matters.
  • Alignment happens between agents rather than between people. What reaches a human is the actions and approvals that genuinely need a human.
  • The class of problem this addresses is familiar to any product company: sales promise what is hard to deliver in a sane timeframe, while engineering builds what the market does not need. Both sides act reasonably on their own terms; they simply have no place to meet on context — the agent loop becomes that place.

Where the project stands

Technically it all works: the integration, the agent, the install inside the company's perimeter. There is no business result yet — it arrives when answers are accurate enough to rely on. An analogy with language models: base training is behind us, the main effect comes from post-training, and that is the stage we are in. We start with one process and extend into neighbouring ones when that makes business sense. Numbers will be added once they exist.