Blog// ai

Why 84% of Companies Get No Return from AI: The Missing Context Layer

Why 84% of companies see no ROI from AI: no context layer linking models to business processes. How to build a semantic bus and avoid small business pitfalls.

29-10-2026 · 7 min read
Why 84% of Companies Get No Return from AI: The Missing Context Layer

According to MIT Technology Review, virtually every organization today uses artificial intelligence in some form. But only 16% of surveyed leaders report transformational value from implementations. The rest get stuck at the pilot stage or local improvements, failing to scale results across business processes. The reason is not model immaturity — in the words of Atlan co-founder Prukalpa Sankar, general intelligence has already been achieved for most tasks. The problem is that AI does not know how your company works.

Numbers That Explain the Disappointment

The MIT Technology Review Insights survey revealed a clear hierarchy of results. 40% of respondents record moderate benefits in specific tasks or teams. 21% achieved substantial improvements. Nearly a quarter — 23% — see no return on investment or a minimal one. Only 16% reached the coveted zone of transformational value, where AI changes business economics rather than just accelerating individual operations.

Meanwhile, the implementation structure looks like this: 38% of companies are still piloting use cases, 35% use AI as an employee assistant — copilots, chat assistants. Only 12% have autonomous agents in production, and a mere 3% have entrusted AI with core work in key processes. The overwhelming majority remain stuck at the “assistant” stage, not advancing to the “executor” stage.

Where the Gap Between Capabilities and Reality Lies

“There is a huge gap between what models are capable of and how they are actually useful inside organizations,” says Prukalpa Sankar. She compares the current situation to hiring the smartest intern in the world, blindfolding them, and dropping them in an office without an onboarding briefing. The model may know everything about quantum physics and Roman history, but it does not know how your company calculates margin, why customers drop off at the payment stage, or what “urgent order” means under your internal rules.

In Sankar's view, a decade ago AlphaGo beat the world Go champion because it had access to the full game board and rules. Today corporate AI often plays Go without seeing the board and without knowing how the knight moves. Models have become smarter, but the architecture connecting them to business has not changed.

What Is a Context Layer and Why Is It Needed

A context layer is an architectural middleware between a company's systems of record (ERP, CRM, knowledge bases, logistics) and AI platforms. Its job is to translate corporate intelligence into a managed, machine-readable format that agents and models can use for decision-making.

Without this layer, every tool operates in a vacuum. The support chatbot does not see order history in 1C. The procurement agent does not know informal agreements with suppliers recorded only in commercial directors' correspondence. The marketing automation system does not account for current warehouse stock. The result — hallucinations in business logic, calculation errors, and employee distrust of AI recommendations.

The context layer solves three tasks:

  • Glossary and metrics — unified definitions of terms (“Lead,” “Margin,” “Shipment”) and calculation formulas understandable to the model.
  • Permissions and gates — who can request which data, which actions an agent is allowed to initiate without human confirmation.
  • Feedback — a mechanism for accumulating experience: AI work results return to the knowledge base, improving subsequent iterations.

Comparison: AI With Context and Without It

Without a context layer, answer quality is generic, templated, and often irrelevant to the process. With a context layer, answers are tied to current company data and rules.

Without a context layer, agent autonomy is low — constant human control is required. With a context layer, autonomy is high — the agent acts within defined mandate boundaries.

Without a context layer, speed of deploying a new use case takes a week for prompt tuning and integrations per task. With a context layer, it takes hours — the agent connects to a ready semantic bus.

Without a context layer, production error risk is high — “hallucinations” of business logic. With a context layer, risk is controlled — gates and validation at the infrastructure level.

Without a context layer, knowledge accumulation is fragmented, in logs of separate tools. With a context layer, it is centralized, improving all connected systems at once.

What This Means for Small Business in Russia and the CIS

Owners of shops, services, and small manufacturers often think: “A context layer is for enterprise with hundreds of systems. We have 1C, a website, and amoCRM — we'll manage with prompts.” In practice, the gap hits harder. Small businesses lack a staff of data engineers who manually extract context for each pilot. An AI error in cost calculation or an auto-reply to a key client costs more relative to turnover.

Concrete steps for this week:

  1. Audit definitions. Collect 15–20 key business terms with formulas in one document (Google Sheets or Notion): how LTV is calculated, what goes into CAC, when an order is considered “problematic.” This is the draft glossary for AI.
  2. Close accesses on a “need to know” basis. Do not give the neural network a full 1C database dump. Export only product catalogs, counterparties, and stock balances in structured form (JSON/CSV) and connect via RAG or API function calling.
  3. Require mandatory confirmation for financial actions. Any operation where AI initiates a write-off, refund, or price change passes through a “Confirm” button with a manager. This is your primitive safety gate.
  4. Start collecting feedback into structure. Every time a manager corrects a bot response or rejects a system recommendation — log it in an “Exceptions and Rules” table. In a month you will have a dataset for fine-tuning or rule configuration.

If you lack resources to build this infrastructure yourself, you can connect a ready-made solution. Our team builds exactly such a context layer — a unified semantic bus between your systems and AI agents. More on the approach and implementation examples in the AI agents for business section.

Frequently Asked Questions

Do we need to buy an expensive platform like Atlan or Palantir?

No. For small business, a context layer starts with structured CSV/JSON exports, a simple vector store (such as Qdrant or PGVector), and a set of Python scripts that translate AI queries into SQL for your 1C or CRM. Buying a boxed product is justified when data sources exceed ten and there is a data team.

Can we just use long prompts with instructions?

Prompts work until the first price change, new manager, or non-standard situation. Context in a prompt is static, becomes obsolete instantly, and does not scale across multiple agents. A context layer is a living, versioned infrastructure, not a text hint.

How to protect trade secrets when connecting external models?

Use local models (Llama, Qwen, GigaChat On-premise) or private clouds with FSTEC certification. Only metadata and reference data enter the context layer — no client personal data or full contract texts. The models themselves are not trained on your data; they only read context at query time.

Where to start if there is no internal technical team?

Hire a freelance data engineer for project work: the task is to set up daily reference data exports from 1C/CRM to cloud storage and deploy a simple RAG system over them. This costs comparable to a junior analyst intern's salary for 2–3 months, but provides a foundation for any AI tool for years ahead.

Based on: MIT Technology Review — Creating AI value with enterprise context

Analysis & review by
Alexander K.

Web development, digital advertising and marketing expert since 2012. Founder of AIFORSALE.

How we can help
Online storesMarketingAI agentsContact us
← All articles