The Next Customer for Your Software Is an AI Agent

For about twenty years, the software business has been a contest for human attention. A company builds a product, puts it behind a login, and then spends enormous amounts of money getting people to come back and click around. Engineering teams pour their time into dashboards, onboarding flows, color schemes, and page load speeds. Marketing teams spend millions driving traffic. The whole model rests on one assumption: that a person will sit in front of the screen and use the thing.

That assumption is starting to break, and the reason is a new kind of user. AI agents, the software assistants that can take actions across your tools instead of only answering questions, are becoming a meaningful share of the "people" using business software. An agent doesn't care about your onboarding flow. It doesn't notice your brand colors. It wants one thing from a piece of software, which is a clear, reliable way to use it to get a job done.

I think this is one of the biggest openings in the software market in a generation, for builders and for the businesses that buy software. So let me explain what's changing, show you how some of the best-known companies are responding, and offer some practical advice depending on which side of the table you sit on.

A plain-English guide to MCPs, CLIs, and skills

There are three terms you'll hear a lot in this conversation, and they're simpler than they sound.

An API, or application programming interface, is the way one piece of software talks to another. APIs have been around for decades. They work well for developers who write code against them, but they're rigid. There's a fixed set of doors, and you can only walk through the ones that were built.

An MCP, short for Model Context Protocol, is a newer standard designed specifically for AI. You can think of it as a universal adapter. When an AI agent connects to a platform through an MCP server, the platform tells the agent what it does, what actions are available, and how to use them properly. The agent can then operate the software much the way a trained employee would, without anyone writing custom code for each task. Anthropic introduced MCP in late 2024, and adoption has been remarkable. By the time Anthropic donated the standard to the Linux Foundation last December, there were more than 10,000 active public MCP servers and more than 97 million monthly downloads of its developer kits, and it had been adopted by ChatGPT, Gemini, Microsoft Copilot, and most other major AI products. The new foundation that oversees it was co-founded by Anthropic, OpenAI, and Block, which tells you how much the industry has rallied around it.

A CLI, or command-line interface, is a text-based way to control software. Developers have used them forever, and it turns out they're a natural fit for agents, which work in text anyway.

Skills are the final piece. They're instruction files that teach an agent when to use a tool, in what order to do things, and what to do with the results. A good way to picture it is the difference between handing a new hire the keys to a system and handing them the keys plus a well-written training manual.

Stripe: letting agents run the money side

Stripe, the payment processing company, is a good example of why this matters. Once agents got capable, people immediately wanted to connect them to Stripe so the agent could send invoices, follow up on payments, and log everything in the CRM. For a while, the only way to do that was to build your own custom connection between the two, which was a lot of work, and the few people who pulled it off would post about it online while everyone else wondered how they found the time.

So Stripe built the connection itself. Today it offers an official MCP server that lets an agent create customers, send invoices, manage subscriptions, issue payment links, and pull financial reports, along with a command-line tool and a set of agent skills that can be installed with a single command. In practical terms, an agent can now use Stripe much the way a person on your finance team would.

The detail I find most instructive is a safety feature. Stripe requires a person to confirm certain actions before the agent can complete them, including refunds and outbound payments. The agent sends you a link, you review the request, and only then does the money move. One of the most agent-friendly platforms on the market decided that some decisions still belong to a human, and designed that checkpoint into the product. I think that's exactly the right instinct.

Resend: software built for machine users

Resend is a newer company that provides email sending for other software products. We use it inside our own CRM at Avolis, so our team can send a client email with one click without switching over to Gmail.

Resend has leaned hard into agents. Its command-line tool for AI agents is designed around the way machines work. It automatically switches to structured, machine-readable output when an agent is using it. It can send up to 100 emails in a single request. It has a built-in safeguard that prevents an agent from accidentally sending the same email twice if it retries after an error. It can stream back replies and bounces in real time so the agent knows what happened after it hit send. And it ships with agent skills that teach an agent how to use all of it well.

None of those features would matter much to a person clicking around a dashboard. Every one of them matters to an agent running a follow-up sequence at two in the morning. That's what it looks like to design a product for a different kind of customer.

Salesforce: an incumbent protecting its moat

The most telling move came from one of the largest software companies in the world. In April, at its TrailblazerDX developer conference, Salesforce announced Headless 360, an initiative to make its entire platform usable by agents through APIs, MCP tools, and command-line access. It launched with more than 60 new MCP tools and 30 preconfigured skills, and Salesforce described the goal as letting agents work with its data, workflows, and business logic "without touching a UI." It also positioned Slack as the front door where people and agents work together.

Why would a company that has spent decades perfecting its interface make it optional? Think about why most companies stay on Salesforce for ten years. For most of them, the reason is that all their data lives there, and moving it somewhere else would cost tens of thousands of dollars and months of disruption. Agents change that math. An agent can help move data from one system to another in a fraction of the time and cost it used to take. When switching gets cheap, the old lock-in stops working.

Salesforce's answer is to make itself the most useful place for agents to do their work. If your agents can read your data, update records, and run your workflows directly inside Salesforce, there's much less reason to leave. The CRM becomes a tool for the agents, and that usefulness becomes the new reason to stay. I think it's a smart move, and I expect most large software companies to follow.

Wall Street is already pricing this in

If this all sounds speculative, the stock market didn't treat it that way. In January, Anthropic released Claude Cowork, an agent for everyday professional work, and followed it with a set of plugins aimed at functions like sales, legal review, and document handling. Investors quickly started asking what would happen to software companies whose customers could get the same work done through an agent. Between January 12 and February 23, the S&P software and services index fell about 25 percent, erasing more than $285 billion in market value. The press nicknamed it the "SaaSpocalypse."

What happened next is the part I'd pay attention to. When Anthropic announced integrations with companies like Salesforce, DocuSign, and Intuit, those stocks recovered some ground. The market seemed to reach the same conclusion I have: software companies that make themselves useful to agents have a future, and the ones that only serve humans clicking through screens are in a much tougher spot.

Pricing is changing too

When an agent is the one using your product, charging per seat stops making sense. A customer can simply give one agent access and let it do the work of ten people. So the industry is moving toward two alternatives.

The first is paying for outcomes. Intercom's customer service agent, Fin, is one of the clearest examples. It charges $0.99 per outcome, which usually means a customer's question got resolved with no further help needed. If Fin hands a conversation to a human without resolving anything, the customer isn't charged. The vendor only gets paid when the work actually gets done, which is a very easy pitch to say yes to. You can imagine the same model elsewhere: a small base fee plus a set price for every qualified lead an agent generates, every invoice it collects, or every appointment it books.

The second is paying for usage. The agent runs on a budget, and the business pays for what it uses, such as each batch of data processed or each piece of content generated, with limits the owner sets inside the platform. In some cases, the pricing page may disappear entirely, and the agent will simply pay as it goes.

What to do about it

The right move depends on where you sit.

If you're a founder building something new, start with the agent experience. Ship an MCP server before you build a settings page, and write skills before you write tooltips. Then look for a problem that mid-market companies currently pay people to handle and ask yourself whether an agent could do it with the right tools. If the answer is yes, build those tools. You'll still want a clean dashboard where people can check what their agents have been doing, but it doesn't need to carry the whole product. And building custom software has never been easier or cheaper than it is right now.

If you run an existing software company, you still have time, but it's shrinking. Here's a simple test: if a customer never logged into your product again, would they still get value from it? If the answer is no, you have work to do. The good news is that putting an MCP server on top of an existing API is usually a matter of weeks, and there are open-source reference implementations to start from.

If you're a business owner buying software, add one question to every vendor renewal: can my AI agents use your platform, through an MCP server, a command-line tool, or something similar? A vendor that can't answer yes is making it harder for your team to get the most from AI, and you should think twice about signing a long contract with them. Just as important, ask what controls the platform gives you over what an agent is allowed to do, and where a person gets to approve the work. Stripe's confirmation step is a good standard to hold vendors to.

Where people fit in an agent economy

It would be easy to read all of this as a story about software replacing people. I read it the other way. As agents take over more of the clicking, sorting, and copying that used to fill people's days, the human work that remains becomes more important, and a lot of it is new.

Someone has to decide which tools your agents get access to and what they're allowed to do with them. Someone has to define the checkpoints where a person reviews the work before money moves or a customer hears from you. Someone has to notice when an agent's output starts drifting from what the business actually needs, and someone has to teach the rest of the team how to hand work to an agent and trust what comes back. In an agent-first world, the dashboard becomes the place where people supervise what their agents did. That's real responsibility, and it takes skill and training to do well.

That's the work we do at Avolis. We help companies figure out which tools their agents should use, connect them properly, set the guardrails, and train their people to work confidently alongside them. It's also why we're always looking for well-built tools that agents can use. If you're building one for the kinds of businesses we serve, we'd love to hear about it.

Every fifteen years or so, the way people interact with software changes, and a generation of established companies misses it. It happened with the web, and it happened again with mobile. This time, the change is that for a growing share of work, the interface is the agent. The companies that design for that, and the businesses that prepare their people for it, are going to have a remarkable decade.

If you're a business owner trying to figure out what this means for your team, a good place to start is our AI Readiness Evaluation.

Sources

• Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation" (December 9, 2025): https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation

• Stripe Documentation, "Model Context Protocol (MCP)": https://docs.stripe.com/mcp

• Resend Documentation, "Resend CLI for AI Agents": https://resend.com/docs/cli-agents

• Salesforce Ben, "Salesforce Headless 360 and Agentforce Vibes 2.0 Revealed at TDX 2026" (April 2026): https://www.salesforceben.com/salesforce-headless-360-and-agentforce-vibes-2-0-revealed-at-tdx-2026/

• DeepLearning.AI, The Batch, "Investors Panic Over Agentic AI: Claude Cowork plugins trigger a SaaS stock selloff" (2026): https://www.deeplearning.ai/the-batch/claude-cowork-plugins-trigger-a-saas-stock-selloff-but-partnerships-lead-to-slight-rebound

• Fin by Intercom, Pricing: https://fin.ai/pricing