For several years, the technology industry has enjoyed predicting the death of enterprise software.

Cloud computing was supposed to eliminate traditional platforms.

Composable architecture was supposed to dismantle the suite.

Low-code tools were supposed to remove the need for specialist systems.

Now AI agents are apparently going to replace the CRM, CMS, ERP and almost every other three-letter acronym in the enterprise technology estate.

The argument sounds plausible.

Why navigate a complex application when an agent can simply perform the task for you?

Why train an employee to use a CMS when they can ask an AI to create a campaign page?

Why open a CRM when an agent can summarise an account and update the opportunity automatically?

The interface may indeed change dramatically.

The underlying enterprise platforms are not disappearing.

AI agents are becoming another category of user.

The interface is not the system

Enterprise software is frequently judged by its user interface because that is the part people encounter.

A CRM appears to be a collection of contact records, forms and reports.

A CMS appears to be an editing interface.

An ERP appears to be an impenetrable collection of financial and operational screens.

But the user interface is only one layer.

Behind it sits the actual enterprise system:

• data models;
• business rules;
• permissions;
• transactions;
• workflows;
• audit trails;
• integrations;
• validation;
• lifecycle management;
• and operational history.

An AI agent may allow users to bypass parts of the interface.

It does not eliminate the need for those underlying capabilities.

When an employee asks an agent to issue a refund, something still needs to:

• verify that the order exists;
• determine whether the refund is permitted;
• calculate the correct amount;
• prevent duplicate transactions;
• update the financial record;
• notify downstream systems;
• and preserve an auditable history.

The agent is not the system of record.

It is a new way to interact with the system of record.

Agents need enterprise software more than enterprise software needs agents

A general-purpose AI model has broad reasoning capability but little organisational authority.

It does not inherently know:

• which customer record is accurate;
• which price is contractually agreed;
• which content is approved;
• which employee has access;
• or which workflow must be followed.

That knowledge resides in enterprise platforms.

The CRM knows the customer relationship.

The ERP knows the financial transaction.

The CMS knows the approved content.

The identity platform knows the user.

The analytics system knows the observed behaviour.

The agent becomes useful when it can work across those systems while respecting their rules.

This is why current enterprise agent platforms emphasise connections to systems of record, shared business context, permissions and operational boundaries.

The model provides reasoning.

The enterprise platforms provide truth, authority and execution.

The agent is a non-human identity

Architecturally, organisations should treat agents less like software features and more like non-human users.

That means an agent needs:

• an identity;
• assigned roles;
• explicit permissions;
• scoped credentials;
• approved tools;
• session controls;
• and an audit history.

This is a subtle but important change.

Traditional system integrations often use service accounts with broad privileges. The integration is assumed to execute a predictable set of predefined operations.

Agents are more dynamic.

They may determine which tools to call based on context. They may perform several steps in sequence. They may react differently when the data changes.

Giving that kind of system a highly privileged service account is dangerous.

A safer model is to evaluate every proposed action against the agent’s identity, the user’s delegated authority, the current context and the sensitivity of the operation.

An agent preparing a draft article may need read access to product information and write access to draft content.

It should not automatically receive permission to publish globally.

An agent analysing customer opportunities may need access to selected CRM records.

It should not automatically be able to export the entire customer database.

Four ways agents will interact with enterprise platforms

The relationship between agents and enterprise software will develop across four broad patterns.

1. The agent as an assistant

The agent helps a human complete a task but does not perform irreversible actions.

Examples include:

• summarising a customer account;
• suggesting improvements to a page;
• generating a campaign brief;
• explaining an analytics report;
• or identifying content that may be outdated.

This is the lowest-risk model and the natural starting point for most organisations.

2. The agent as an operator

The agent invokes approved tools to perform specific actions.

Examples include:

• creating a draft content item;
• updating a CRM field;
• opening a support case;
• translating approved copy;
• or scheduling an experiment.

The action is constrained, logged and usually reversible.

3. The agent as an orchestrator

The agent coordinates work across multiple platforms.

For example, it may:

1. identify a high-value customer segment;
2. analyse relevant products;
3. generate campaign content;
4. create assets in the CMS;
5. configure an audience;
6. initiate an experiment;
7. and notify the campaign owner.

The agent is not replacing those platforms. It is coordinating them.

4. The agent as an autonomous colleague

The agent continuously monitors a business objective and takes action within defined boundaries.

It might optimise a content portfolio, adjust campaign activity or investigate operational exceptions.

This is the most ambitious model and the one requiring the strongest governance.

Why enterprise platforms may become more important

Agents increase the value of structured, governed platforms.

A human can often compensate for a poor system.

They can interpret an ambiguous field.

They can recognise that two customer records refer to the same company.

They can notice that a workflow document is outdated.

Agents are less reliable when context is inconsistent.

The more organisations delegate to AI, the more they need:

• clean data;
• explicit schemas;
• reliable APIs;
• well-defined permissions;
• documented business rules;
• and controlled workflows.

In other words, AI does not eliminate the need for enterprise architecture.

It exposes the consequences of weak enterprise architecture.

What happens to the user interface?

Enterprise interfaces will change.

Many tasks currently performed through menus, forms and dashboards will become conversational or automated.

However, visual interfaces will continue to matter for:

• reviewing complex information;
• comparing options;
• handling exceptions;
• configuring permissions;
• investigating incidents;
• approving sensitive changes;
• and understanding system state.

The future is unlikely to be a blank chat box replacing every application.

It will be a combination of conversational interaction, generated interfaces, dashboards, forms and automated workflows.

The interface will adapt to the task.

The underlying system will continue to enforce consistency.

CMS as an example

Consider the apparently simple request:

Create a new product page for our German market.

An agent could generate the words almost instantly.

But creating the actual enterprise experience may require the system to determine:

• which product content is approved;
• which content model to use;
• which components are allowed;
• which German terminology is mandated;
• which claims require legal approval;
• which assets are licensed for Germany;
• which metadata is required;
• which accessibility rules apply;
• which audience can see the page;
• and when it may be published.

This is not merely content generation.

It is governed content operations.

The CMS remains central because it provides structure, lifecycle and control.

The agent becomes a new user of those capabilities.

The risk of letting agents bypass platforms

Some organisations will be tempted to allow agents to connect directly to databases or perform actions outside established workflows.

This may appear faster.

It is also likely to create serious problems.

Bypassing the platform can mean bypassing:

• validation;
• permissions;
• event handling;
• version management;
• business rules;
• approval processes;
• audit logging;
• and integration contracts.

The agent may successfully update the primary table while leaving search indexes, downstream services or compliance records inconsistent.

Enterprise platforms are often frustrating precisely because they encapsulate years of operational requirements.

Those requirements do not disappear when the request comes from an AI.

What vendors should build

Enterprise software vendors should stop thinking of AI solely as a feature inside the existing interface.

They need to design their platforms for agent consumption.

That means offering:

• well-described tools;
• structured context;
• machine-readable schemas;
• fine-grained permissions;
• delegated authorisation;
• human approval checkpoints;
• observable execution;
• safe rollback;
• and predictable error handling.

The best enterprise platform will not necessarily be the one with the most impressive embedded chatbot.

It may be the one that allows many different agents to use its capabilities safely.

The Optimizely perspective

Optimizely’s direction with Opal reflects this broader shift.

Opal is increasingly positioned not just as a text-generation assistant but as an agent and orchestration layer spanning content, experimentation and other digital-experience operations. Optimizely has introduced pre-built and custom agents, workflow orchestration and tools embedded across its products.

The opportunity is significant because digital experience management is inherently cross-functional.

Creating a page is only the beginning.

The organisation may also need to:

• target it;
• personalise it;
• experiment on it;
• measure it;
• optimise it;
• and govern it.

An agent that understands those connected capabilities becomes more valuable than a generic content generator.

The CMS, experimentation platform and analytics tools remain the operational foundation.

Opal becomes another user of that foundation—albeit a potentially very capable one.

Final thoughts

AI agents will undoubtedly change how people experience enterprise software.

Some interfaces will become less important.

Some manual tasks will disappear.

Some workflows will be compressed from hours into seconds.

But enterprise software exists because organisations need reliable systems for managing truth, authority and process.

Agents do not remove that need.

They depend on it.

The winning architecture will not replace the enterprise stack with a collection of autonomous models.

It will make the enterprise stack legible, accessible and safely actionable by those models.

AI agents are not replacing enterprise platforms.

They are becoming another user.

And, like every other user, they need identity, permissions, context, tools and supervision.