Stop Forcing AI Agents Into Human-Shaped Software

Most business software was designed around a human sitting in front of it.

The employee reads the screen, understands the surrounding situation and decides what to do next. The application records information and presents options, but the human supplies much of the intelligence required to use it.

Businesses are now trying to plug AI agents into existing software applications and expecting them to work like experienced employees. Give the agent access to the CRM, ERP, email inbox and document repository, and it should be able to handle the work, right?

Sometimes, but not always.

An agent may be able to click the buttons, but it doesn’t automatically inherit the context and informal knowledge that make those buttons meaningful. Even if it can technically operate the application, it may not have the context needed to make sound business decisions.

Intelligence alone doesn’t make an agent operationally useful. The company also needs to provide an environment in which that intelligence can be applied reliably.

Competing in the agentic era will require more than connecting AI to existing applications. Businesses need to rethink their software stack so it presents their operations in a way agents can understand, act on and navigate safely.

Humans Can Ask a Colleague. We Expect AI to Just Know.

When an employee doesn’t understand something in a system, they can ask.

From casual pantry chats to formal onboarding handovers, human employees have the advantage of being able to gather context naturally. When they need help, they can message the team on Teams, walk over to finance or simply ask questions over lunch.

Without being purposeful about it, humans naturally operate inside a cloud of shared context.

This is why old, clumsy systems remain usable for years. Employees quietly compensate for their weaknesses, coming up with creative workarounds and passing them on as tribal knowledge.

AI agents don’t have this same privilege.

They don’t overhear a conversation between two managers. They don’t know that the warehouse supervisor stopped trusting one report three months ago. They don’t automatically understand that “confirmed” means something different to sales than it does to operations.

So when a business puts an AI agent in front of existing software and expects it to be proficient, the agent often underperforms. The business chalks it up to “AI isn’t smart enough”, but in reality, the agent was never given the information necessary to do its job well.

At best, the agent does a poor job and someone catches it. At worst, it fails to realise that context is missing and proceeds confidently with the wrong answer.

You Cannot Teach an Agent Every Workaround in the Business

Once this gap becomes visible, the obvious response is to teach the agent everything employees already know. The agent is given instructions explaining what each field means, which records to trust, which customers are exceptions, who normally approves what and what to do when the system disagrees with reality.

This can work for one narrow workflow. The problems begin when companies try to scale the same approach across the business.

Every process requires another set of instructions. Every exception turns into another rule. Every policy change means someone has to update the agent. Trying to narrate every edge case and special treatment becomes an infinite game of whack-a-mole (we’ve been there ourselves).

The company’s software stack was built on the assumption that its users could fill in the gaps by asking someone. Agents cannot.

The problem isn’t only the difficulty of describing every edge case. The knowledge an agent needs to do its job well is rarely stored in one place, it’s scattered across different systems, past conversations and the experience of individual employees.

Giving the agent access to more applications so it can “build its own context” doesn’t automatically bring that knowledge together. It may simply expose the agent to more incomplete records, conflicting information and undocumented exceptions, increasing its chances of reaching the wrong conclusion.

Eventually, the company is no longer building an intelligent worker. It is building an enormous operating manual describing how humans compensate for software that was never designed to explain itself.

That is a Sisyphean task. The alternative is to change what the software exposes.

Agentic Software Must Make the Business Legible and Operable

What companies need isn’t smarter agents or longer instruction manuals. They need to rethink how their software presents the business. Traditional applications are designed to store records, display information and let humans work out what those records mean. Agentic software must make that meaning more explicit.

It starts with context.

A useful system shouldn’t simply return a customer record. It should surface the active request, previous commitments, unresolved issues, relevant policies and any reason the underlying information may be unreliable.

The agent shouldn’t have to search through five applications and reconstruct the situation from fragments. The software should present the situation in an agent-optimised shape.

Once the business is legible, it must also become operable.

An agent shouldn’t need to imitate a person clicking through five screens if the actual business action is “reserve stock for this order” or “request approval for revised payment terms”. The system should expose the action itself, not merely the human interface used to reach it.

Finally, that operation needs control.

The system must define what the agent may do independently, what requires approval and when it must stop or escalate. An agent might answer routine status questions, but not change a contractual date. It might reorder common items within an approved threshold, but require human approval before engaging a new supplier.

Together, context makes the business understandable, actions make it operable, and controls make that operation safe.

This is what makes agentic architecture fundamentally different from placing AI on top of existing applications. The goal isn’t to give an agent a digital mouse and keyboard. It’s to expose the business in a form the agent can understand and operate reliably.

Instead of teaching the agent every workaround humans have accumulated, the software makes the operation itself explicit.

The Advantage Will Come From Operationalising AI Better

In time, most companies will have access to similarly capable AI models. The best models may remain expensive or differentiated for a brief period after release, but the broad direction is clear: access to intelligence will become more widely available.

The competitive difference will come from what each company enables that intelligence to do.

One company’s agent will spend its time opening screens, searching message histories, reconciling contradictory records and asking employees what information means. Another company’s agent will receive the relevant context, perform approved actions and escalate only the decisions that genuinely require human judgment.

Both companies may be using the same underlying model, but one has simply purchased intelligence, while the other has operationalised it.

This is similar to the difference between buying powerful machinery and redesigning a factory around it. The machine may be impressive, but its value depends on how materials reach it, what processes surround it and whether the rest of the operation can keep up.

AI agents are no different.

A highly capable model connected to fragmented operations may produce impressive demonstrations without materially changing how the company works. It can draft emails, summarise documents and answer questions, but if the actual business still depends on employees manually moving information and coordinating decisions, it is unlikely to produce material business returns.

The real leverage appears when the agent can participate in the operation itself.

Designed properly, an AI agent should be able to identify that an order is at risk, gather the relevant context, check the available options, recommend a response, complete permitted actions and escalate the final decision with a clear explanation.

At this point, the company isn’t just using AI. It’s redesigning part of the business around it.

That is likely to become the dividing line in the agentic era.

The winners won’t simply be the companies that adopted the strongest models first. They’ll be the companies that made their operations usable by those models.

Your Existing Software May Stay, but It Cannot Remain the Whole System

This doesn’t mean every company should throw away its current applications.

Existing systems can continue storing customer records, inventory, transactions and financial information. Replacing them completely may be unnecessarily disruptive and expensive. However, they cannot remain the complete environment through which agents understand and operate the business.

Ideally, existing software will evolve to expose its context, actions and controls in an AI-native way.

Where existing applications cannot evolve, businesses will need an agentic layer between those systems and the agents using them. That layer gathers relevant context, turns fragmented application functions into clear business actions and enforces the boundaries within which agents may operate.

This is why companies shouldn’t evaluate agentic readiness by asking whether their current applications have an AI feature or an API.

The more important question is whether an agent can understand what is happening, take the right action and know when it must stop.

Your existing software was built for employees who learned how to work around it.

Agentic software must be built so agents don’t have to.