For most of the history of business software, we’ve built technology around people.
Finance has its systems.
Sales has its CRM.
Operations has its platforms.
Customer service has its helpdesk.
Then we connect everything together and put people in the middle to make decisions, move information and keep the company operating.
AI changes something fundamental about that structure.
Software can now interpret information.
It can communicate.
It can make decisions.
It can use tools.
Increasingly, it can take action.
That makes me wonder whether we’re thinking about enterprise AI in the wrong way.
The conversation today is largely about giving employees AI assistants.
I think that’s probably an intermediate step.
My prediction is that companies will eventually develop something much more structural: entire divisions of specialised AI agents operating alongside their human organisation.
I think of them as AI sub-agent divisions.
And if that happens, we aren’t simply adding AI to the enterprise.
We’re introducing an entirely new organisational layer.
The current model is mostly one human, one AI
Most enterprise AI adoption today follows a fairly understandable pattern.
Give employees access to an AI assistant.
Let salespeople use AI for research.
Let customer service use it to draft responses.
Let finance analyse documents faster.
Let developers write code with it.
There’s enormous value in that.
But structurally, very little has changed.
The human still sits at the centre.
They receive information, ask the AI for help, interpret the response and decide what happens next.
The AI is essentially a very capable tool.
I don’t think that’s where this ends.
As agents become more reliable and gain controlled access to company systems, the relationship changes.
Instead of waiting for a person to ask a question, an agent can receive an event.
A customer replies.
An invoice becomes overdue.
A new lead enters the CRM.
A contract arrives.
A project moves stage.
Something unusual appears in operational data.
The agent understands what happened and determines what should happen next.
That is a very different role for software.
What is an AI sub-agent division?
The easiest way to explain the idea is to forget about AI for a moment and think about how a company already works.
A business doesn’t usually employ one extraordinarily capable person and give them responsibility for everything.
It creates specialisation.
Finance understands finance.
Sales understands sales.
Operations understands operations.
Legal understands legal.
Management coordinates responsibilities across those functions.
Within each department, responsibilities become more specialised again.
I think enterprise AI will increasingly develop in a similar direction.
Rather than one enormous AI agent having access to everything, a company could operate multiple specialised agents with clearly defined jobs.
A sales division might contain agents responsible for lead research, qualification, follow-up, pipeline management and reporting.
Finance could have agents responsible for reconciliation, invoice monitoring, forecasting or preparing information for human review.
Operations could have agents monitoring processes, identifying exceptions and coordinating work between systems.
Customer service could have specialised agents for different products, customer types or levels of escalation.
These agents don’t need to know everything.
They need to know enough to perform their job exceptionally well.
That distinction matters.
The organisational chart starts becoming digital
This is where the idea gets particularly interesting.
Imagine opening an organisational chart five years from now.
At the top, you still have leadership.
Below that are the traditional divisions of the company.
Finance.
Sales.
Operations.
Technology.
Customer service.
But underneath and between those human teams sits another organisational structure.
AI agents.
Some report information to humans.
Some work alongside humans.
Some coordinate other agents.
Some operate autonomously within narrow boundaries.
Some may never interact directly with a person at all.
An operational agent might supervise several specialised agents responsible for individual processes.
Those agents might communicate with agents belonging to other divisions.
Above them could sit supervisory systems responsible for monitoring performance, permissions or unusual behaviour.
Suddenly, the architecture starts looking less like a collection of software tools and more like an organisation.
That is why I use the word division.
The interesting question isn’t simply how many agents a company has.
It’s how responsibility is organised between them.
Agents need responsibilities before they need intelligence
There is a temptation with AI to start with capability.
We have this model.
It can reason.
It can use these tools.
What can we make it do?
I think enterprise architecture should approach the problem from the opposite direction.
Start with responsibility.
What job needs doing?
What information is required?
What systems are involved?
What decisions need to be made?
What authority does the role require?
What should happen when something falls outside that authority?
Only then should you decide what intelligence belongs inside it.
This is remarkably similar to designing a human organisation.
A company doesn’t usually hire someone and then invent their responsibilities afterwards.
There is a role.
The role has objectives.
It has access to certain information.
It has authority to make certain decisions.
And there are decisions it has absolutely no authority to make.
AI agents need the same clarity.
The more capable the models become, the more important those boundaries become.
Not every agent should be able to talk to every other agent
Once you create specialised agents, another problem appears.
Communication.
This is one of the areas my team and I find particularly interesting.
If a sales agent needs information from finance, what happens?
The obvious answer is to give the sales agent access to the financial system.
But that may be completely unnecessary.
The sales agent might only need one piece of information.
Instead, it could ask an authorised finance agent.
That finance agent determines whether the request is legitimate, retrieves what is permitted and returns only the information required.
The sales agent continues its work.
That architecture is much closer to how a well-controlled organisation already operates.
Employees don’t necessarily receive unrestricted access to another department’s systems every time they need something.
Information moves through boundaries.
AI should be no different.
This means agent-to-agent communication needs its own architecture.
Who can initiate communication?
Which agents can communicate?
What information can move between them?
Can an agent delegate work?
Can another agent reject that request?
How is identity established?
What gets recorded?
What happens when an agent attempts something outside its authority?
These aren’t theoretical details.
They are what separates a clever multi-agent demonstration from something I’d actually want operating inside an enterprise.
There will probably be managers in the AI organisation too
If you have enough specialised agents, somebody or something has to coordinate them.
That creates another interesting possibility.
Manager agents.
Imagine an operational process involving six different specialised agents.
One receives the initial event.
Another retrieves information.
Another performs analysis.
Another interacts with an external system.
Another verifies the result.
Another communicates the outcome.
You could hard-code every possible sequence.
Or you could introduce an orchestration layer capable of understanding the objective, assigning work and monitoring whether it was completed correctly.
That begins to resemble management.
Not management in the human sense.
The agent isn’t motivating employees or dealing with office politics.
It is coordinating responsibilities.
It knows which agents exist.
It understands what they are allowed to do.
It can delegate tasks.
It can evaluate whether something has been completed.
And when something falls outside the architecture, it can escalate to a person.
Eventually, I think we’ll see multiple levels of this.
Specialist agents.
Coordinating agents.
Supervisory agents.
Human managers.
The organisational chart becomes considerably more interesting.
Humans don’t disappear from this model
Whenever you talk about autonomous AI, the conversation inevitably turns to replacing people.
I think that oversimplifies what is happening.
Some work will absolutely disappear.
Some roles will change significantly.
New roles will appear.
But I don’t think the most sophisticated enterprise AI systems will simply remove humans from the architecture.
They’ll become much more deliberate about where humans belong.
There are decisions where judgement matters.
There are situations involving relationships.
There are unusual circumstances where historical data provides very little help.
There are decisions carrying legal, financial or reputational consequences that an organisation may deliberately reserve for a person.
And ultimately, somebody has to remain accountable for the system.
The interesting architecture isn’t fully human or fully autonomous.
It’s deciding where responsibility should sit.
An AI agent may perform 95% of a process independently and escalate the remaining 5%.
Another may only prepare information and never have permission to execute anything.
Another might operate autonomously until a financial threshold is reached.
The point isn’t maximum autonomy.
It’s appropriate autonomy.
Permission architecture becomes part of organisational design
This is where enterprise AI becomes much more serious.
A model being capable of doing something is irrelevant if the organisation hasn’t authorised it to do it.
Capability and authority are different things.
A finance agent might be capable of issuing a $1 million payment.
That doesn’t mean anyone should give it permission to.
A customer service agent may technically be capable of accessing an entire customer record.
That doesn’t mean it needs to.
A sales agent might be capable of changing commercial terms.
That doesn’t mean it should.
Every agent needs an identity.
That identity needs permissions.
Those permissions should correspond to the agent’s responsibility.
And ideally, the architecture should operate on the principle of least privilege.
Give the agent what it needs.
Nothing more.
As sub-agent divisions become larger, I think this becomes one of the defining characteristics of good enterprise AI architecture.
Not how intelligent the agents are.
How well controlled they are.
The same applies to memory
Memory creates a similar problem.
An agent responsible for sales may need to remember previous conversations with a prospect.
A support agent may need historical information about a customer’s issue.
A finance agent may need access to transactional history.
That doesn’t mean those three agents should share one enormous pool of memory.
Enterprise information already has boundaries.
AI memory should respect them.
Some memory should be temporary.
Some should persist.
Some should belong to a particular customer.
Some should belong to an agent.
Some should be available across a division.
Some should never leave a particular system.
And some should probably never be stored by the AI architecture at all.
The question isn’t how much an agent can remember.
It’s what it is allowed to remember.
Compliance cannot be something added afterwards
This becomes particularly important in regulated industries.
Healthcare.
Financial services.
Insurance.
Legal.
Government.
Large enterprises operating across multiple jurisdictions.
In those environments, an AI system cannot simply be judged by whether it produces useful outputs.
You need to understand how it reached an action operationally.
Which systems were accessed?
Which agent performed the work?
What permissions did it have?
What information moved between systems?
Was human approval required?
Was that approval received?
Which model processed the information?
Where was the data processed?
What happened when something failed?
If an architecture can’t answer those questions, adding a compliance policy document afterwards doesn’t fix it.
Security, compliance, governance and liability have to influence how the system is designed from the beginning.
That may slow down some decisions during development.
I think that’s a good thing.
The objective is not to deploy the maximum possible amount of AI.
The objective is to deploy AI that an organisation can actually trust.
Every action should leave a trail
As these systems become more autonomous, observability becomes critical.
Imagine something goes wrong.
A customer receives incorrect information.
A financial record changes unexpectedly.
An agent takes an action it shouldn’t have.
A process fails halfway through.
You need to be able to reconstruct what happened.
Not the model’s private internal reasoning.
The operational chain.
Agent A received an event.
It accessed these systems.
It requested information from Agent B.
Agent B returned this data.
Agent A proposed an action.
A human approved it.
Agent C executed it.
This was the resulting system state.
That level of visibility turns an opaque AI system into something an engineering team can actually operate.
It also means the architecture can improve.
Patterns emerge.
You see where agents struggle.
You see which escalations happen repeatedly.
You discover where permissions are too broad.
You identify processes that were badly designed in the first place.
The system starts teaching you about the organisation around it.
Legacy systems aren’t going anywhere
There is another reason I think the sub-agent model makes sense.
Enterprises already have enormous technology stacks.
CRMs.
ERPs.
Data warehouses.
Internal applications.
Communication systems.
Databases built fifteen years ago.
Software nobody particularly likes but everybody is terrified to replace.
And, inevitably, at least one spreadsheet holding together something far more important than anyone wants to admit.
AI isn’t going to replace all of that overnight.
It doesn’t need to.
Agents can increasingly become an intelligent layer between existing systems.
One agent might interact with the CRM.
Another understands an internal database.
Another communicates with the ERP.
Another knows how to retrieve information from company documentation.
Instead of replacing every system, the agentic layer helps coordinate them.
That is potentially much more practical than asking a large enterprise to rebuild twenty years of infrastructure because AI arrived.
The hardest part may not be building the agents
Models will continue improving.
Agent frameworks will improve.
Tool use will improve.
Memory will improve.
Infrastructure will improve.
Building an individual agent will probably become easier and easier.
I think the difficult part will increasingly be organisational architecture.
Deciding what the agents should do.
Deciding where responsibilities begin and end.
Designing permissions.
Managing communication.
Creating escalation paths.
Maintaining security.
Understanding liability.
Monitoring performance.
Integrating everything with systems that already exist.
And making the entire thing understandable enough that the humans responsible for the organisation can trust it.
That’s a much bigger problem than writing a prompt.
It’s also why I think enterprise AI will increasingly become an architectural discipline rather than simply a software implementation exercise.
I think companies will eventually have two organisational charts
This is the prediction I keep coming back to.
Today, an organisational chart shows people.
Tomorrow, I think large companies may effectively have two.
The first shows the human organisation.
The second shows the intelligent systems operating around it.
Agents.
Responsibilities.
Divisions.
Supervisors.
Permissions.
Communication pathways.
Systems.
Human escalation points.
Some of those agents may effectively belong to departments.
Others may operate across the entire company.
Some may perform millions of small actions without anybody thinking about them.
Others may exist purely to watch the rest.
The two organisational charts won’t be separate in practice.
They’ll overlap.
Humans will manage AI systems.
AI systems will support humans.
Agents will coordinate other agents.
And increasingly, work will move between all three.
That is why I don’t think the future of enterprise AI is simply giving everyone a better chatbot.
We’re beginning to design an entirely new operating layer for the company.
I call that idea AI sub-agent divisions.
The terminology may change.
It probably will.
The architecture behind it is what matters.
And I think we’re going to be building a lot more of it.
You can follow what we’re building, or explore working with us, at RayneAI.com.
