AI for mid-market: too big for a chat tab, too small for a SAP programme

A company in the illustrative ₽50–₽500 million revenue range is often too operationally complex for scattered chat subscriptions and too resource-constrained for a multi-year transformation programme. It has real CRM, finance, documents, approvals, and customer commitments, but the same people may own process, operations, and change.

The useful middle path is a company OS: a small shared layer for roles, skills, memory, connectors, permissions, and evidence. It creates reuse without turning the rollout into an ERP replacement.

Why chat tabs stop scaling

Personal assistants can improve individual drafting, but instructions live in accounts, context is copied manually, and nobody knows which version is current. Sensitive documents move without a common policy. When a useful employee leaves, the procedure leaves too.

The problem is not that chat is bad. It is that private conversation history is not an operating asset. Promote repeatable procedures into versioned skills and confirmed facts into project memory.

Why a giant programme is also wrong

A broad consulting transformation can spend months on target architecture before one queue improves. Mid-market companies need controls, but they also need evidence within weeks. Customizing every system and creating a large governance body can exceed the value available.

Start with one cross-functional control plane and one role. Keep existing 1C, Bitrix24, amoCRM, document storage, and service systems where they work. Connect through official interfaces instead of rebuilding them.

Build a minimum company OS

Use a git repository as the source of truth for neuro.yaml, role skills, project memory, evaluation cases, and connector declarations. Run tasks in isolated sandboxes. Keep credentials in a broker outside execution. Choose cloud, VPC, or self-hosted deployment from data and operating needs.

Set Allow for bounded reads and local transformations, Ask for external writes, and Block for prohibited access. Human approval applies to the exact CRM change, send, posting, or publication.

Target margin through process

Margin does not appear because a model sounds intelligent. It improves when the company reduces rework, shortens cycle time, catches leakage, increases throughput without equivalent overhead, or improves service consistency.

Choose a process with volume and a measurable bottleneck: sales preparation, support triage, contract comparison, finance reconciliation, or onboarding coordination. Baseline labor, waiting, errors, and outcome. Include model, sandbox, connector, review, and operating costs.

Roll out in ninety-day increments

During the first month, define the role, clean critical source fields, and run shadow cases. During the second, promote reliable low-risk slices and keep writes behind Ask. During the third, establish a daily queue, weekly metric review, incident path, and one adjacent task.

The business owner reviews output and exceptions; a technical owner maintains runtime and connectors; security sets boundaries. This can be three responsibilities across fewer than three people, but none can be absent.

Preserve optionality

Use any approved model that passes the role’s evaluation. A foreign cloud model may fit difficult synthesis, a Russian provider may fit procurement or residency, and a local model may fit sensitive or high-volume tasks. Skills should survive the swap.

The mid-market advantage is speed of decision and proximity to process. Use it. Avoid both uncontrolled experimentation and heavyweight theatre. A compact company OS turns what the organization learns into files, checks, and roles that compound—while every expansion still has an owner, a metric, and a gate.

This work runs on Neuro OS. To scope a first role, get started.

Explore Neuro OS →