Igor Fokusov
Menu
Case study

Role-Aware AI Assistants for Enterprise Work

Active R&D on context-aware AI assistants that understand a user's role, work context, enterprise state and desired outcome while operating with bounded autonomy.

R&D2026
Contents

Active R&D on AI assistants that understand not only a user’s message, but also who the user is, what role they perform, what they are working on, what enterprise context applies, and what outcome they are trying to achieve.

The research began with natural-language access to 1C data, progressed through controlled business tools and the production Enterprise AI Assistant, and now asks a broader question:

Can an AI assistant understand enough of an employee’s work context to help complete real business tasks — without replacing the employee or bypassing enterprise controls?

The research path

01NL2QueryNatural-language enterprise data access
02Business toolsExplicit capabilities instead of raw ERP access
03Enterprise AI AssistantProduction · identity, permissions and controlled execution
04Role-aware assistantsWork context and bounded autonomy
AI Accountant · active R&DAI Manager · active R&D
NL2Query remains the first stage of the research lineage, not a standalone promoted case.

Each stage exposed a limitation of the previous one. The architecture evolved from interpreting a question to understanding both intent and work context.

Four stages of the work

Stage 1

Natural language to enterprise data

Early experiments used 1C metadata and controlled query generation to turn ordinary business questions into structured answers.

The user should not need to understand the internal ERP structure to ask a business question.

Stage 2

Business capabilities, not raw queries

Receivables, sales, inventory and reconciliations became explicit business tools. The model reasons about capabilities rather than inventing unrestricted access to the ERP model.

Stage 3

A production enterprise assistant

Identity propagation, deterministic authorization, controlled tool execution, auditability, observability and confirmation for consequential actions became the production Enterprise AI Assistant.

Stage 4 · current R&D

Understanding the employee's work

The current direction adds role, task continuity, enterprise state and desired outcome. Understanding one request is not the same as understanding the user's work.

Context model

A useful work assistant needs more context than a chatbot.

01User identityWho is this person?
02Role and permissionsWhat are they responsible for and allowed to do?
03Work contextWhat are they working on now?
04Enterprise contextWhat is happening in 1C?
05History and desired outcomeWhat happened before, and what are they trying to achieve?
Together, these layers provide context for a role-aware AI assistant.

The goal is not merely to classify a sentence. The assistant should understand enough surrounding work to infer a useful and safe next step.

From answering questions to assisting work

Consider an accountant saying:

The reconciliation with Romashka is wrong again.

A role-aware assistant can connect the user’s role with the current period-close task, identify the counterparty, inspect supported documents, payments and settlements, explain the discrepancy, and prepare the next supported action. Confirmation remains required where the action is consequential.

The transition is from “What answer should I return?” to “What outcome is this employee trying to achieve, and what safe next step can I help with?”

Current R&D directions

Active R&D · not a production product

AI Accountant

A context-aware work assistant for investigation and preparation around accounting tasks.

  • investigate reconciliation discrepancies;
  • collect relevant documents and amounts;
  • explain likely causes;
  • prepare supported corrective actions;
  • return ambiguous situations to the accountant.

The accountant remains the responsible professional.

Active R&D · not a production product

AI Manager

An exploration of the same architecture for operational work and task continuity.

  • find overdue or blocked work;
  • identify where a process is stuck;
  • collect business context from 1C;
  • propose a next action or follow-up;
  • return to work when relevant enterprise state changes.

The manager retains responsibility for decisions and outcomes.

Persistent context and proactive assistance

Enterprise work is task-centred rather than message-centred. A longer-lived assistant may need the active task, relevant business objects, previous findings, unresolved questions, user decisions, pending actions and changes in enterprise state.

The exact persistence model is still under active development. The design goal is to preserve enough task context to continue work without treating every message as unrelated.

Proactive assistance means recognizing when help may be useful — for example when an investigated issue changes state or a supported workflow is waiting for a decision. It does not mean allowing an agent to act freely in the ERP. Irrelevant events should produce no action; relevant events may surface context or propose a next step.

Bounded autonomy

The target is not unrestricted autonomous execution. Autonomy increases only for low-risk, well-defined work.

01Observe
02Investigate
03Reason and prepare
04Evaluate action risk
Low risk + permitted → continueConsequential → user review and confirmation
Boundaries come from enterprise permissions, tool contracts, business policy, action risk and approval requirements.

Proven foundations and active development

Implemented or validated

Proven foundations

  • natural-language access to enterprise data;
  • 1C metadata and business context;
  • MCP / explicit business tools;
  • role-aware identity and permissions;
  • multi-step agent reasoning;
  • production experience from the Enterprise AI Assistant.
Active R&D

Being developed

  • deeper role context;
  • persistent task context;
  • AI Accountant and AI Manager;
  • proactive assistance;
  • bounded autonomous continuation.

These capabilities should not be interpreted as fully deployed production functionality.

Research principles

  • Context is part of the product. A strong model with weak work context still behaves like a chatbot.
  • User intent is larger than the message. Employees often describe symptoms rather than formal goals.
  • Enterprise semantics matter more than raw access. Explicit business capabilities remain the safer interface.
  • Autonomy should grow with confidence and control. Investigation can be more autonomous than consequential actions.
  • The employee remains responsible. The purpose is augmentation, not substitution.
  • Long-running work requires task memory. Conversation history alone is not enough to represent enterprise work.

Why this matters

The opportunity is not another chat interface on top of ERP data. It is a class of enterprise software that understands an employee’s role, current work and business context well enough to help move real work forward.

For 1C environments, the central research question is how to bridge safely between what the employee wants to accomplish and what the enterprise system can do.

Scope

Research lineage

  • NL2Query;
  • controlled business tools;
  • Enterprise AI Assistant;
  • role-aware work assistants.

Current R&D directions

  • AI Accountant;
  • AI Manager;
  • persistent work context;
  • proactive assistance;
  • bounded autonomy.

Enterprise foundations

  • identity and role-aware permissions;
  • business tools and MCP;
  • 1C integration;
  • policy and risk boundaries.

Status

  • foundational capabilities: implemented / proven;
  • Enterprise AI Assistant: production;
  • AI Accountant / AI Manager: active R&D.