AI Agents vs SaaS: What Changes for Businesses by 2028

"SaaS is dead" is a headline, not a plan. The more useful question for a business is: which of my workflows are really just a person moving information between tools, and what happens when an agent does that instead? Here is what the forecasts say and how I approach it when building agents for clients.

What the forecasts actually say

Gartner has published a series of predictions on this, and they add up to a clear direction.

• Agents inside apps, fast. Gartner predicts that up to 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. In a best case, it expects agentic AI to drive about 30% of enterprise application software revenue by 2035, over $450 billion.
• Agentic AI in a third of enterprise software by 2028. Gartner also expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024, with at least 15% of day-to-day work decisions made autonomously.
• Spend at risk. In July 2026 Gartner said up to $234 billion of enterprise application software spending is exposed to "agentic arbitrage" through 2030: agents completing tasks across several systems, so people no longer need to work through each application's interface.
• Buyers shifting to outcomes. Gartner expects that by 2028 most enterprises will stop paying for assistive AI such as copilots in favour of platforms that commit to workflow results.

Note the wording: "exposed" and "at risk" describe spending that could move, not a forecast of losses. The money does not vanish. It moves from seats and interfaces toward agents, integration and outcomes.

Why agents threaten the seat, not the software

Most SaaS is priced per user because value used to scale with the number of people logging in. An agent breaks that link. If one agent handles the triage that ten people used to do in a helpdesk, the vendor sells fewer seats even if the helpdesk is doing more work.

The underlying system still matters. The agent still needs somewhere to store tickets, invoices or contracts, with permissions, history and an audit trail. What changes is who operates it. Systems of record stay; the screens people click through become less central.

What changes inside a business

Which workflows move first

Gartner's April 2026 release points to approval-heavy, time-sensitive workflows as the first to change. In my experience, the best early candidates share three traits:

• High volume, structured inputs. Invoices, purchase orders, intake forms, support tickets. At Avenuor, the platform I built as CTO, a photographed invoice becomes structured data through an AI extraction step, validated against a schema before anything is saved. Nobody retypes it into another tool.
• Clear rules for "done." If two people would agree on whether the task was completed correctly, an agent can be evaluated against it.
• Reversible or approvable actions. Drafting a reply, proposing a categorisation, preparing a purchase order for sign-off. Irreversible actions, such as payments, come later and always behind approval.

Workflows that depend on relationships, negotiation or ambiguous judgement stay with people longer, with agents preparing the groundwork.

The risk: building agents on top of tools you cannot reach

The catch is that agents can only work with what they can reach. A SaaS tool with no API, or an API that cannot do what the interface can, blocks automation. Before you renew a major contract in 2027, ask the vendor:

• Is every action in the interface also available through the API?
• Do you offer an MCP server, or plan to? MCP is now governed by the Agentic AI Foundation under the Linux Foundation, so it is a reasonable thing to ask for. I explain the protocol layer in MCP vs A2A.
• Can an agent act with a specific user's permissions, rather than one all-powerful API key?
• How will pricing work if agents, not people, do most of the work?

Expect costs to move, not disappear

Agent-based work is not automatically cheaper. Gartner predicts inference costs per agentic workflow will increase more than fivefold through 2028, because agents make many more model calls per task. Seat savings can be eaten by token bills if no one sets step limits, caches results or routes simple tasks to smaller models.

And many agent projects will not make it. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value and weak risk controls. The fix is the same as for any agent: measure it. See why AI agent projects fail.

How to prepare before 2028

• Map your swivel-chair work. List every task where a person reads from one system and types into another. That list is your agent backlog.
• Clean up permissions and data first. An agent inherits your mess. Role-based access enforced on the server, clear ownership of records, and an audit trail make agents safer and easier to debug.
• Start with one workflow and a number. Pick one process, measure today's time or error rate, and build an agent with human approval on the risky steps. I cover measurement in how to measure AI automation ROI.
• Negotiate for the transition. Ask for API parity, MCP support and usage-based options in renewals, so you are not locked into seats you will not need.
• Keep systems of record. Replace interfaces with agents where it helps; do not replace your source of truth with an agent's memory.

Work with me

I build AI agents that work across the tools a business already uses, with permissions and approvals designed in. If you want to find the first workflow worth automating, see AI agent development and workflow automation, or book a scoping call.