Why Business Apps Are All Shipping AI Agents

You did not ask for it. One morning your CRM has a new button that offers to handle the follow-up sequence, your helpdesk software starts drafting replies, and your accounting tool proposes to reconcile the month for you.

None of those arrived through a procurement process. They shipped in an update.

Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of this year, up from under 5% the year before. That is one of the fastest feature rollouts in business software history, and most of it will reach you as a default setting rather than a purchase decision.

This is a guide to what that actually means for a buyer, and the questions that separate a useful agent from a liability.

Key Takeaways

• Gartner forecasts 40% of enterprise apps will feature task-specific AI agents, up from less than 5% a year earlier.
• In its best case, Gartner projects agentic AI driving about 30% of enterprise software revenue by 2035, above $450 billion, from 2% today.
• The difference between an assistant and an agent is task specialization: agents complete end-to-end work rather than answering questions.
• Governance is lagging badly. 60% of organizations cannot terminate a misbehaving agent and 67% lack audit trails covering everywhere agents operate.
• The single most valuable procurement question is whether permissions can be scoped per task, not per account.

What the Forecast Actually Says

Gartner's projection is specific. Forty percent of enterprise applications will be integrated with task-specific AI agents, up from less than 5% the previous year. The full press release is published on the Gartner newsroom.

Going from 5% to 40% in a single cycle is an eightfold increase. Software categories rarely move that fast, and when they do it is because the feature is being added to existing products rather than sold as new ones.

The longer projection is larger still. In its best case scenario, Gartner estimates agentic AI could drive roughly 30% of enterprise application software revenue by 2035, exceeding $450 billion, against 2% currently.

Numbers that size explain the urgency. Every vendor in the market has concluded that shipping an agent is not optional, which is why yours will ship one whether or not you asked.

Assistant Versus Agent: The Distinction That Matters

The vocabulary is used loosely, and the difference determines your risk.

An assistant responds. You ask a question, it produces an answer, and you decide what to do. The output is information.

An agent acts. It has a goal, access to systems, and the ability to take steps in sequence without checking in at each one. The output is a changed state in the world.

Gartner frames the transition as adding task specialization, which turns an assistant into something that performs complex end-to-end work. Their example is a cybersecurity response agent that scans network traffic, system logs, and user behaviour in real time, assesses what it finds, and initiates a response.

Notice the last clause. It does not recommend a response. It initiates one.

That is the whole distinction, and it is why the governance conversation is different for agents than it was for any previous software feature. We covered the organizational side of this in why businesses are racing to deploy AI agents.

The Governance Gap You Are Inheriting

Here is the problem with features arriving through updates: the controls do not arrive with them.

Research from the Cloud Security Alliance found that 65% of organizations had at least one security incident in the past year caused by AI agents on their networks. Among those affected, 61% saw sensitive data exposed and 41% saw unintended actions taken inside business processes.

The control figures are worse than the incident figures:

• 63% cannot enforce limits on what an agent uses its access for
• 60% cannot terminate an agent that is misbehaving
• 67% have no audit trail spanning all the systems agents touch
• Only 19% treat agents as the security equivalent of a human insider

Put that against the 40% adoption forecast and the shape of the next two years becomes clear. Capability is arriving roughly eight times faster than control. The detailed mechanics of how this goes wrong are covered in our piece on how AI agents breach companies on their own.

What a Rollout Actually Costs

Vendors present agents as included, which makes them feel free. Work the real number.

Take a 200 person company where an agent feature ships enabled by default across three tools:

• Licence uplift: commonly $15 to $30 per user per month for the tier that includes agents. At $20 across 200 users that is $48,000 a year.
• Review time: if agents handle 500 actions a month and you sample-check 10%, that is 50 reviews at five minutes each, about four hours monthly, or roughly $3,000 a year in loaded cost.
• Access audit: one engineer spending a week scoping permissions properly across three tools, around $4,000 once.

First year total lands near $55,000. Against that, if the agent genuinely removes two hours of repetitive work per week for 40 of those 200 people, you recover roughly 4,160 hours a year. At a loaded $40 an hour that is $166,400 of capacity.

The return is real. The point is that the $7,000 of review and audit work is the part everyone skips, and it is the part that determines whether the other $48,000 produces value or an incident.

Six Questions to Ask Before It Ships

Use these with any vendor that tells you an agent is coming.

• Can permissions be scoped per task rather than per account? This is the highest value answer in the list. An agent with account-level access has whatever the account has, including everything nobody remembered to remove.
• Is it off by default? Features that arrive enabled skip your review process entirely. Ask whether you control activation.
• What requires human approval? Specifically for anything irreversible: payments, sends, deletions, and anything customer facing.
• Where do the logs live and can I export them? An audit trail inside a vendor dashboard you cannot query is not an audit trail for your purposes.
• How do I stop it mid-task? Not whether a stop button exists. Whether you have tested it and watched the agent actually halt.
• Who is liable when it acts wrongly? Get the answer from the contract, not the sales call. Most terms place this squarely on you.

If AI systems are also becoming how customers find you, the discovery side is a related shift, and our guide to getting AI to recommend your business covers that half of the change.

Frequently Asked Questions

What is a task-specific AI agent?

It is software given a goal, access to systems, and the ability to complete a sequence of steps without checking in at each one. Unlike an assistant, which returns information for you to act on, an agent takes the action itself. Gartner's example is a security agent that assesses threats and initiates a response rather than recommending one.

How fast are companies actually adopting AI agents?

Gartner forecasts 40% of enterprise applications will include task-specific agents by the end of this year, up from less than 5% the year before. Most of that adoption arrives through product updates to software companies already use, rather than through separate purchases.

Are AI agents in business software safe?

Not automatically. Cloud Security Alliance research found 65% of organizations had at least one incident caused by AI agents in the past year, most commonly sensitive data exposure. The risk is manageable with scoped permissions, approval gates on irreversible actions, and exportable logs, but 60% of organizations currently cannot even terminate a misbehaving agent.

Should I turn off AI agent features in my business tools?

Turning them off wholesale usually costs more than it saves. The better approach is to enable them where the task is repetitive and reversible, require approval for anything involving money or customer communication, and audit what access each agent actually holds before it goes live.

How much do AI agent features cost?

Licensing commonly adds $15 to $30 per user per month for the tier that includes agents, but the licence is not the full cost. Budget separately for review time on agent output and for a one-time permissions audit, which together often run several thousand dollars in the first year and determine whether the rest produces value.

The Bottom Line

Going from 5% to 40% in one cycle means this is not a decision most companies will make deliberately. It will arrive in a release note.

That is exactly why the procurement questions matter now rather than later. Ask whether permissions scope per task, whether it ships off by default, and whether you have personally tested the stop button.

The productivity case is genuine and the numbers support it. The organizations that get burned will not be the ones that adopted agents. They will be the ones that adopted them without noticing.