AI agents move closer to the tools you already own

Four items worth your attention today share a single spine: the question of who owns the layer of software that now does the reasoning, and what it costs to keep that layer running. Two vendors moved their agents closer to the tools a business already owns, and two essays looked at the economics underneath, where a customer with a modest subscription can now do work that used to require a paid seat inside someone's product.

An agent that works on top of your existing helpdesk

Start with the idea, because it matters more than any single product. For most of the last decade, adding a capable new feature to your support stack meant one of two things: wait for your current vendor to build it, or move your data to a vendor who already had it. Migration is expensive in ways that do not show up on the invoice. You retrain a team, you rebuild routing rules, and you risk losing the history that makes past conversations searchable. The cost of switching is often the real reason a business stays with a tool it has outgrown — a tension we looked at in when a spreadsheet stops being enough.

So the more interesting design pattern is an agent that does not ask you to move. Intercom said today that its Fin agent can run as a Service Agent on top of HubSpot and Freshworks [1]. The claim is narrow and worth reading literally: you keep your helpdesk, and the agent operates over it. For a business choosing tools, that reframes the decision. You are no longer picking one platform and accepting everything bundled with it. You are choosing a system of record and, separately, choosing which agent reasons across it. Those two decisions used to be welded together. Pulling them apart is what makes the announcement matter.

The trade-off is worth naming. An agent that sits on top of another vendor's product depends on that product's interface staying stable, and on the two companies keeping their integration current. We have written before about how integrations break quietly, and a layered arrangement has more joints that can fail. That is not a reason to avoid it. It is a reason to ask, before you commit, who is responsible when the layer beneath shifts.

What it costs to run AI, and why that sets your price

The second idea is about money, and it explains a great deal of what you will see vendors do this year. Running a capable model costs something every time it answers. If you build AI into your product and pay per call to a model provider, each customer interaction has a marginal cost that never reaches zero. If your customer instead pays a flat subscription for the same class of model, they can do a large amount of work for a fixed monthly fee.

SaaStr put this plainly today, describing a conversation with the CEO of a B2B company who was proud that their AI cost had fallen to pennies per customer [4]. The essay's point is not that low cost is poor engineering. It is that pennies-per-customer can also mean the cheaper model was chosen, and that a customer paying between twenty and two hundred dollars a month for a frontier model directly may simply be able to do more than your product does. For anyone choosing tools, the lesson runs both ways. When you evaluate an AI feature, ask what model sits behind it and whether the vendor's unit economics push them toward the cheaper option over time. A feature priced to protect a margin can quietly get less capable. We covered the deciding-what-you-pay-for question in what a subscription plan is really selling.

This is also why the honest answer to "should we build AI into this?" is often "only where it earns its cost." Not every task deserves a model call, and the discipline of deciding which ones do is the same discipline we described in how to tell whether a task should be automated.

Turning one-off prompts into repeatable work

The third item is about a habit, not a headline. A prompt typed once into a terminal produces a result once. It is not a process. It cannot be reviewed by a colleague, run again next week, or trusted to behave the same way twice. The gap between a clever one-off and a dependable process is the gap between a demonstration and a tool your team can rely on.

GitHub described custom agents for its Copilot command-line tool today, framing them as a way to turn one-off terminal prompts into repeatable, reviewable processes that understand a team's stack and workflows [2]. Read that as a general principle rather than a single product feature. The value is in the words repeatable and reviewable. An automation you cannot inspect is a liability, because when it does the wrong thing you have no way to see why. An automation you cannot repeat is a party trick. The moment a prompt becomes a named, versioned thing that a second person can read, it crosses into being infrastructure. That is the line worth watching whenever a vendor adds "agents" to a tool you use.

When sales is the caboose, not the engine

The last item is a shift in how software companies grow, and it affects the tools you will be sold. SaaStr wrote today about an argument making the rounds that sales is no longer as central as it once was [3]. The author is careful — they do not agree with it, but they concede the logic. The logic is that when a product is good enough and cheap enough to try, adoption can lead and sales can follow, rather than the other way round.

For a business on the buying side, this is quietly useful. If a vendor's growth comes from people using the product first and buying later, then the product has to be genuinely usable before anyone talks to you. That is a healthy incentive. It also means you can learn more from a free trial than from a demo, because the trial is where the company has put its effort. When you evaluate a tool, weight the hands-on hour over the sales call. We made the broader case for judging software by use rather than pitch in choose software worth using.

The thread, pulled together

Put the four together and a picture forms. Agents are becoming portable, so the tool that holds your data and the tool that reasons over it need not be the same. The cost of running those agents is real and uneven, so price and model quality are now linked in ways worth interrogating. Repeatability is what separates a demonstration from a process. And the way software reaches you is shifting toward try-first. None of this requires you to act today. It does change the questions you ask before you sign. For the same flow of decisions from a different week, our previous briefing covers similar ground.