Over the past year, the conversation around AI in business has moved incredibly quickly.

First came copilots. Then assistants. Then agents.

Now the direction of travel is becoming much clearer: businesses are beginning to think about fleets of AI agents operating across real workflows, systems and teams.

And that creates an entirely new challenge.

It is no longer simply “Which AI model should we use?”

It becomes: How do we coordinate all of this intelligence safely, reliably and usefully across the organisation?

A number of developments over the past week point in the same direction. SAP is demonstrating agent orchestration within its Business Technology Platform. Microsoft continues to embed agents into operational functions. Palantir is emphasising orchestration, telemetry and evaluation as fundamental to enterprise AI. Meanwhile, emerging “digital coworker” products are increasingly focused on persistent, multi-step work rather than simply producing better answers. This is an important shift.

We are moving beyond the AI assistant

The first generation of enterprise AI largely sat beside the user.

Ask a question.

Draft an email.

Summarise a meeting.

Generate some content.

Useful? Absolutely.

Transformational? Not necessarily.

The next generation looks very different.

AI is increasingly moving inside the operating processes of the organisation.

An agent might detect an event in one system, understand its significance, gather context from several others, decide what needs to happen, initiate work, communicate with the relevant people, monitor the outcome and escalate when necessary.

The value is no longer in a single interaction with a model.

The value sits in the whole chain of action.

That is why orchestration, telemetry, observability, evaluation and governance are becoming so important.

And that creates the next enterprise problem: agent sprawl

Imagine a business a few years from now.

Microsoft agents operating across productivity tools.

Salesforce agents working inside CRM.

Finance agents.

HR agents.

Customer service agents.

Specialist industry agents.

Internal agents created by individual teams.

AI running inside workflows built by employees themselves.

Potentially several different models operating underneath all of them.

Suddenly the challenge is not a shortage of AI.

It is too much AI operating independently.

Businesses will need to know:

Who owns each agent?

What is it allowed to access?

What can it change?

What information does it retain?

Which other agents can it interact with?

What decisions has it made?

What has it cost?

How frequently does it fail?

When should a person intervene?

Can everything it has done be audited?

The growing conversation around “agent sprawl” is already leading organisations towards agent registries, ownership, lifecycle controls and risk management.

I think this will become one of the defining enterprise technology problems of the next few years.

The most valuable layer may sit above the agents

There is another important implication.

Businesses are unlikely to have one AI.

They will have many.

Different models will be better suited to different tasks.

One may be excellent at reasoning.

Another at communication.

Another may run locally because the information is sensitive.

A specialist model may understand a particular industry.

A computer-use agent might operate an application that has no modern API.

The winning architecture therefore may not be the organisation that chooses the “best model”.

It may be the organisation that can coordinate the right intelligence, with the right context, for the right job.

We are already seeing architectures emerge that allow models to be dispatched dynamically depending on the task.

That makes the orchestration layer increasingly important.

Context becomes infrastructure

There is another piece that is easy to underestimate: memory.

As agents undertake longer-running and increasingly complex work, they need more than access to today's data.

They need organisational context.

What was promised to this customer?

Why was this decision made?

Who approved it?

What happened previously?

What has already been tried?

What are the constraints?

What happened in the meeting that never made it into the CRM?

What is sitting in an email thread that changes the meaning of the task?

Research into multi-agent systems increasingly points towards long-running memory, planners and executors becoming fundamental components of these architectures.

That means organisational memory itself starts becoming infrastructure.

Models may change.

Agents may change.

Applications certainly will.

But the accumulated context of the organisation is incredibly valuable.

This is where we believe Autm fits

At Autm, we have been building around a simple premise:

Businesses already have systems. What they need is intelligence across them.

Email, meetings, calendars, CRM, documents, accounting systems, project tools and communications platforms already contain enormous amounts of organisational intelligence.

The problem is that the information is fragmented.

Decisions disappear into meetings.

Commitments remain buried in email.

Context gets lost during handovers.

Systems know what happened inside them, but rarely understand what is happening across the organisation.

Autm is designed to sit across that environment.

To connect systems.

Preserve context.

Coordinate workflows.

Enable AI agents to act.

Maintain permissions and auditability.

And turn the information flowing through a business into coordinated action.

As the number of agents inside organisations increases, we believe the need for that orchestration layer becomes greater, not smaller.

The next generation of AI needs management, not just intelligence

Businesses will also need to measure their AI workforce very differently.

Token usage tells a technology team how much computing was consumed.

It doesn't tell a CEO whether anything useful happened.

The metrics that matter will increasingly look more like:

Did the workflow complete?

Was the customer issue resolved?

How much human intervention was required?

What did the outcome cost?

How much time was saved?

How reliably does this agent perform?

That is why telemetry and business-specific evaluations are appearing repeatedly in discussions about production AI systems.

The future enterprise dashboard may not simply show revenue, pipeline and utilisation.

It may also show the performance of the organisation's digital workforce.

The opportunity is bigger than automation

For a long time, we have talked about automation as a way to remove repetitive work.

Agentic AI introduces something considerably bigger.

It creates the possibility of an organisation in which people, systems and AI continuously work together.

Information can become action.

Actions can produce signals.

Signals can trigger new decisions.

And the organisation can become progressively more responsive.

But that only works if there is coordination.

Without it, we risk creating hundreds of intelligent systems operating in isolation.

The enterprises that benefit most from AI may therefore not be those with the greatest number of agents.

They may be the ones that can orchestrate intelligence across data, applications, people and agents while maintaining context, governance and accountability.

That is the opportunity we are building Autm around.

The agentic enterprise is arriving.

The next question is who — or what — will orchestrate it.