The Enterprise AI stack in 2027

Topic

Agentic Architecture

Kind

Perspective

Author

Dmitry Ermakov

Read time

4 min read

/

/

The Enterprise AI stack in 2027

/

/

The Enterprise AI stack in 2027

The enterprise AI stack is starting to look very different from the one companies were building two years ago.

The first architecture was straightforward:

Application โ†’ LLM API โ†’ answer.

Then came retrieval.

Application โ†’ vector database โ†’ LLM โ†’ answer.

Now enterprises are introducing agents, tools, local models, code execution, memory, permissions and dynamically generated interfaces.

AI is becoming an architectural layer across the organisation.

That requires a different stack.

Layer 1: Infrastructure

Everything starts with where the system runs.

Cloud.

Private cloud.

European cloud.

On-premise infrastructure.

Edge environments.

Most large organisations will probably use a mixture.

The important question is not whether one environment is universally better.

It is whether workloads can be placed according to their requirements.

Highly sensitive workloads may remain inside controlled infrastructure.

Other tasks can use commercial model APIs.

Architecture should make those decisions possible.

Layer 2: Enterprise data

AI becomes much more useful when it understands the organisation.

That knowledge already exists across:

  • ERP
  • CRM
  • databases
  • documents
  • email
  • knowledge bases
  • APIs
  • operational systems

The objective should not automatically be to copy all of it into an AI platform.

Sometimes retrieval makes sense.

Sometimes real-time queries make sense.

Sometimes APIs are better.

Sometimes data should remain exactly where it is.

The intelligence layer needs controlled ways to reach the appropriate source.

Layer 3: Models

There will not be one enterprise model.

There will be many.

Frontier models.

Small Language Models.

Domain-specific models.

Embedding models.

Vision models.

Speech models.

Potentially internally fine-tuned models.

Model selection will increasingly happen per workload rather than per organisation.

The architecture therefore needs routing.

Layer 4: Orchestration

This is where AI applications start becoming AI systems.

Orchestration determines:

  • what the user wants
  • which model should handle it
  • what information is required
  • which tools are available
  • whether another agent is needed
  • whether human approval is required
  • what happens if something fails

As systems become more agentic, this layer becomes increasingly important.

This is part of the problem SINAS is designed around.

Layer 5: Execution

AI becomes economically interesting when it can do work.

Query data.

Generate a document.

Execute code.

Update a CRM.

Prepare an order.

Trigger a workflow.

Call another service.

Execution introduces risk, so it requires controls.

Sandboxing.

Permissions.

Validation.

Transaction boundaries.

Human approval.

Logging.

The more capable the agent, the more important this layer becomes.

Layer 6: Memory and context

AI systems need context.

But "memory" can mean several things.

Conversation history.

User preferences.

Company knowledge.

Workflow state.

Previous agent actions.

Long-term organisational knowledge.

These should not all be treated as one giant context window.

Different forms of memory have different retention, security and retrieval requirements.

Layer 7: Governance

Governance cannot remain a document outside the architecture.

It increasingly needs technical expression.

Which models are approved?

Which data can leave the organisation?

Which tools can agents access?

Which actions require approval?

What is logged?

How long is information retained?

Who owns each AI system?

Governance becomes policy enforced through software.

Layer 8: Intelligence

This is where the system turns information into decisions.

Classification.

Extraction.

Reasoning.

Planning.

Generation.

Prediction.

Evaluation.

Different capabilities can come from different models.

The enterprise does not need to care which model wins a benchmark this month.

It needs an intelligence layer capable of using the right capability for the task.

Layer 9: Intent

This may become the most interesting layer.

Traditional enterprise software begins with structure.

Menus.

Forms.

Applications.

Modules.

AI begins with intent.

Find the customers most likely to churn.

Prepare me for tomorrow's meeting.

Explain why our margin fell.

Build a campaign for this product.

The system interprets the objective and determines which resources are required.

This is the idea behind ONE.

Layer 10: Interface

Finally, the result needs to reach a human.

Chat will remain one interface.

But it will not be the only one.

AI can generate:

  • dashboards
  • tables
  • forms
  • reports
  • workflows
  • visualisations
  • application states

This is where generative UI becomes interesting.

Instead of forcing every interaction through a chat window, software can assemble the interface appropriate to the task.

This is the territory we are exploring with IntentCX.

From applications to capabilities

The larger shift is that enterprise software may become less application-centric.

Today an employee thinks:

I need Salesforce.

I need SAP.

I need Power BI.

Tomorrow they may simply express an objective.

The intelligence layer determines which underlying systems are required.

Those applications do not disappear.

They become capabilities available to a broader operating layer.

Infrastructure, intelligence and intent

We increasingly think about this architecture through three broad questions.

Infrastructure

Where does everything run, and how is it controlled?

Intelligence

How does the organisation reason, automate and act?

Intent

How do people express what they want the organisation to accomplish?

SINAS, ONE and IntentCX each address different parts of that system.

WeAreBrain brings them together with the engineering, design and implementation teams required to put them into production.

The enterprise AI stack is therefore becoming less about adding an AI feature to existing software.

It is becoming the architecture through which people, software and intelligence work together.

Avatar-1
Avatar-2
Avatar-4
Avatar-4
Avatar-3

Fresh insights, straight to your inbox.

Engineering breakdowns, sovereign-AI perspectives and practical guides from the people building enterprise AI. One email when there is something worth your time.