AI Agents: The Moment Software Stops Being Software

AI agents are transforming software from tools people operate into systems that can act independently. Moving beyond chatbots and copilots, these agents can research, use applications, build workflows, and complete multi-step tasks. This shift could reshape how businesses buy technology, challenging traditional SaaS models based on user seats and interfaces. As agentic AI advances, software may become less visible, with humans increasingly assigning goals while autonomous systems handle the execution behind the scenes for businesses.

‍

For decades, software has worked according to a familiar formula: humans open an application, navigate an interface, enter information, click buttons and wait for the software to respond. Even the cloud and Software-as-a-Service (SaaS) revolutions largely preserved this relationship. Applications became easier to access and continuously updated, but people still had to operate them.

Artificial intelligence is beginning to challenge that model.

The technology industry is rapidly moving beyond chatbots that answer questions and copilots that suggest what a person should do next. The emerging generation of AI agents can potentially receive a goal, decide how to accomplish it, use tools, navigate applications, conduct research, execute workflows and return with a completed result.

That distinction could prove far more significant than another improvement in generative AI.

It could change what software actually is.

From answering questions to completing work

Traditional generative AI largely operates through a request-response model. A user asks for something, and the AI produces text, code, an image or an analysis.

An agent introduces an additional element: action.

Instead of asking an AI to explain how to research ten competitors, for example, a business could assign an agent to research the competitors, visit relevant sources, organize the information, analyse the findings, create a spreadsheet and prepare an executive summary.

OpenAI describes modern agents as systems capable of planning and completing tasks using tools, maintaining context across multiple steps and collaborating with other agents. Its Agents API, introduced in September 2026, is designed for long-running work and can provide agents with environments in which they can use files, run code and retain intermediate results.

The transition is also becoming visible elsewhere in enterprise technology.

In May 2026, Microsoft made computer-using agents generally available in Copilot Studio, enabling agents to interact directly with websites and desktop applications through their interfaces. Such agents can potentially operate software even when the underlying application does not provide the API integrations traditionally required for automation.

Anthropic has similarly developed computer-use capabilities that allow Claude-based systems to interpret what is displayed on a screen and interact with applications. In February 2026, Anthropic acquired Vercept as part of its effort to advance this technology for complex tasks across live applications.

The implication is profound: AI may no longer need software to be redesigned specifically for AI. It may increasingly learn to operate the same digital environments humans already use.

The emergence of the digital employee

This is why technology companies increasingly describe agents using workforce terminology.

Salesforce, for example, is positioning Agentforce around the concept of a “digital workforce.” Its sales agents can perform activities including prospect research, lead qualification, nurturing, meeting preparation and quote generation.

By September 2026, Salesforce said its platforms had processed 7 billion Agentic Work Units, including 3.2 billion in its second quarter alone, although these figures are company-reported rather than independently audited measurements of economic productivity.

ServiceNow offers another glimpse of what this could mean operationally. The company says its own AI agents support approximately 400,000 workflows annually, freeing an estimated 3 million hours of employee capacity and generating about $500 million in annualised value. Again, these are company-reported figures, but they illustrate the scale at which enterprise software vendors are beginning to think about agents.

The fundamental change is that companies may increasingly purchase digital execution rather than digital tools.

That difference could have enormous consequences for the software industry.

What happens when nobody opens the software?

The modern SaaS economy was largely built around users.

A company buys 500 CRM seats because 500 employees need to log into the CRM. It buys productivity subscriptions because thousands of employees require interfaces for documents, communication, analytics, accounting or project management.

But imagine that an employee tells an agent:

“Identify all customers whose contracts expire within 90 days, analyse their usage, flag accounts at risk, prepare personalised renewal recommendations and schedule follow-ups.”

The agent might interact with the CRM, analytics platform, email system and calendar without the employee directly opening any of them.

The software remains essential, but increasingly becomes infrastructure hidden behind the agent.

Gartner calls this phenomenon “agentic arbitrage.” In July 2026, it estimated that as much as $234 billion of enterprise application spending could be exposed to agentic AI by 2030, equivalent to roughly 20% of enterprise application SaaS spending. Gartner argues that agents capable of working across several applications could weaken the traditional connection between the number of human users and the amount companies spend on software licences.

That does not necessarily mean SaaS disappears.

It means SaaS may stop looking like SaaS.

From seats and screens to outcomes

A traditional software company sells access to features.

An agentic software company may increasingly sell an outcome.

Instead of paying for 100 customer-service software seats, businesses might pay according to cases successfully resolved. Instead of licensing prospecting tools to 200 salespeople, organisations might pay for qualified opportunities generated. Finance systems could ultimately charge according to invoices processed, reconciliations completed or compliance reviews executed.

Gartner predicts that by 2028, more than half of enterprises could move away from paying for purely assistive AI capabilities such as copilots and instead favour platforms focused on workflow outcomes.

This could move software economics away from per-seat pricing and toward usage, transactions, compute consumption, completed workflows or measurable business results.

The most valuable interface may therefore no longer be a dashboard.

It could simply be an instruction:

“Get this done.”

But autonomy remains far from solved

The agentic future is advancing quickly, but the hype can obscure an important reality: most enterprises are not yet running fully autonomous digital workforces.

Gartner's 2026 research found that only 17% of organisations had deployed AI agents, although more than 60% expected to do so within two years. Gartner also places agentic AI around the “Peak of Inflated Expectations,” warning that many current deployments remain narrowly scoped and that fully autonomous agents are not yet suitable for most enterprise situations.

ServiceNow's 2026 Enterprise AI Maturity Index similarly found that only 9% of surveyed organisations were using agentic AI for autonomous multi-step workflows.

The challenges are substantial.

Giving software the ability to act introduces questions involving security, permissions, identity, data privacy, hallucinations, accountability, audit trails and human approval. An AI that incorrectly summarizes a document creates inconvenience. An AI authorised to send money, modify customer records, purchase inventory or execute contracts can create significantly greater consequences.

For that reason, the near-term enterprise model is likely to combine automation with human supervision, particularly for consequential actions.

Software becomes the invisible layer

The history of computing has repeatedly moved users further away from underlying technical complexity. Command lines gave way to graphical interfaces. Installed software gave way to cloud applications. Search boxes and menus are now beginning to give way to natural-language instructions.

Agents could represent the next abstraction layer.

Tomorrow's employee may not need to know which application contains particular information or which sequence of buttons completes a workflow. The agent could determine that automatically.

That would not eliminate software. Databases, enterprise applications, APIs and SaaS platforms would still provide critical systems of record, business logic and infrastructure.

But the agent could become the layer through which humans interact with all of them.

And that is where the real disruption begins.

For the last generation of enterprise technology, the winning question was:

“Which software should our employees use?”

The next generation may ask something very different:

“Why are employees operating the software at all?”

When AI can understand the goal, coordinate the applications and complete the workflow itself, software does not disappear.

It becomes invisible and increasingly, it starts to look like work itself.

For questions or comments write to contactus@bostonbrandmedia.com