When the agent does the work, what are you paying for?

Picture a new customer signing up. An agent reads the contract, the sales notes and the email thread, then writes the kickoff plan. Nobody opens the tool that used to do that. The company holding the customer records still gets paid, but over a few months the agent has learned everything nobody wrote down: this customer needs security approval before kickoff, that team works a little differently.

Three things happen at once. The tool that used to assemble the plan gets easier to leave. The agent you trained gets harder to replace, because those lessons live in its instructions and memory. And the moment someone decides to switch AI vendors, nobody at the table is counting the same costs.

Lunch is the clue

On September 30, DoorDash launched ordering by text message and a limited beta of an MCP server, the standard way for agents to plug into a service. It wants you to use its own assistant, but it also wants to be there when you use anyone else's. Either way DoorDash gets the order, even if you never open its app again. Losing the interface doesn't mean losing the customer, and an agent may even bring in orders that were too much hassle for a person to arrange.

That's the pattern for software in general: the screen matters less, what sits behind it matters more.

What are you still paying for?

When an agent does most of the finding, sorting and acting on information, think about when you still open an application yourself. Usually it's for one of two things: information you can't get anywhere else, or something that has to be carried out correctly.

Much of SaaS sold a blurry mix of storing, finding, arranging and acting on information. Once an agent gets good at part of that mix, the whole relationship is exposed, because the value was never clearly separated.

Three layers

Data. Records and structured information an agent needs and can't produce itself. Agents make good data more valuable, not less.
Workflow. Configured rules, approvals and dependable execution on top of the data. The interface may change or disappear; the workflow underneath can be the most valuable part.
Agent. The layer that reads, decides and acts across everything else, and increasingly the one people get attached to.

Being reachable by agents, through an API or an MCP server, is table stakes. It gets you considered. It doesn't give anyone a reason to keep paying you; that still has to come from the data or the workflow.

How the big players are positioning

Meta launched Muse as a personal agent, then followed within weeks with Muse for small business, adding connections to business accounts like Shopify and QuickBooks. The path from personal assistant to business tool turns out to be very short.
Salesforce announced Koa, a reasoning model tuned for CRM work and trained on synthetic data, now in pilot. It's positioning on all three layers: the customer data, a control panel where any company's agents can work with it, and its own agent for the jobs it knows best.
Microsoft faces the same question with Outlook and Excel, as work moves from clicking through interfaces to asking an agent. Autopilot, a persistent agent now in preview, is part of its pitch: data storage, development tools, productivity apps and AI from one trusted vendor.
The AI labs start from the other side. They don't hold your data or your distribution, so their pitch is intelligence: the best models, and help applying them to your data. Partnerships between labs and data holders give buyers what they want most, which is not being locked in.

Build on data, or buy the workflow

Two smaller examples show the split.

Recruiting. Crustdata describes a two-person recruiting firm that built its own Claude skill on Crustdata's APIs instead of buying another recruiting tool. The firm supplied its own selection criteria; Crustdata supplied structured data on people and their work histories. It kept paying for the data and built the workflow itself, which is now cheap to do.
Payroll. Explaining a paycheck is very different from running the system that pays people. Workday's payroll agent uses each company's configured data and rules to spot issues and suggest fixes for review. Hardly anyone wants to hand-code payroll rules, given the compliance risk, so the agent makes that workflow more valuable, not less.

The common thread: agents are best at cutting through complexity. Where the complexity is the product, as in payroll, the workflow holds its value. Where it was mostly an interface over someone else's data, the workflow is easy to rebuild.

What to do about it

If you use AI at work, talk about how you do it: which data you use, which workflows your agent handles, and where you still open the application. Keeping a clever setup to yourself makes it fragile; if the company switches vendors without knowing about it, weeks of trained behaviour can disappear. Sharing it also helps your whole team.
If you buy software, look at how teams actually use AI before deciding. Usage is different in every team, often down to individuals, and simplifying from the top can quietly remove productivity nobody knew was there. Ask vendors about all three layers, not just whether they have AI.
If you build or sell software, you're selling the future. Make the data easy for any agent to read, keep workflows easy to plug AI into, and if you ship an agent, let other agents work alongside it. Customers increasingly buy outcomes, not features.

Software isn't going away; it's coming apart into layers that can be judged separately. That makes it a more interesting time to build it, and a more demanding one to buy it.