When your agent tells you to stop paying for software

A run of reports from one software publisher this week share an uncomfortable theme: the tools a business runs on are increasingly chosen, priced, or replaced by an agent rather than a person. In two of them, a team's own AI told it to stop paying for something and build or rent a cheaper version instead — and the team concluded the agent was right.

This matters because the decisions involved are not new. Build or buy, per-seat or per-use, headline growth or committed revenue: these are the questions any business has always asked when picking software. What has changed is who is now sitting at the table when the question gets answered, and how fast the economics underneath each answer are moving.

When an agent tells you to build instead of buy

Build versus buy is one of the oldest decisions in software, and the standard advice for a small team has always been the same: buy. A scheduling link, a billing engine, a help desk — these are solved problems. A three-person company has better uses for its time than maintaining its own version of a tool it could rent for the price of a coffee. That logic has not gone away. What moved is the cost of the "build" side.

A team at SaaStr described exactly this. They had used a scheduling product for years, were happy with it, and had no real reason to replace it — then built their own anyway, because an agent produced a working version on a hosted development platform in about twenty minutes [1]. They were careful to say the bought product was not the problem. In their words, "Calendly is cheap, proven, and works well." [1] The tool was fine. The point was that building a narrow, single-purpose replacement had stopped being a project and become an afternoon.

That should change how you weigh the decision, but not which way you lean by default. When building was slow and expensive, buying won for almost anything outside your core. When building a small tool costs twenty minutes of an agent's time, the real question becomes which side you want to own and maintain for the next few years. Buying still wins when the thing you would build has to keep working without your attention, handle edge cases you have not imagined, and stay secure as the world changes around it. A scheduling link that quietly breaks costs you meetings you never find out you missed. The discipline of asking what a tool is actually for, and whether you want to be responsible for it, is the same discipline covered in choosing software worth using. The agent makes building cheap; it does not make maintenance free.

The real cost of letting agents do the work

The second report makes the trade-off concrete from the other direction. The same team runs with three humans and more than twenty agents [4], and one of those agents touches enough systems that, as they put it, "Across all the apps it touches, it makes 35,000 to 40,000 API calls a day." [4] When they priced formal access for that volume of agent traffic, the estimate came back at $240,000. Their agent proposed a small database instance costing a few dollars a month instead, and they took the suggestion [4].

The lesson is not that cheap always beats expensive. It is that pricing models built for human users behave strangely when the user is an agent. A per-seat tool is almost free to an agent, because one seat can drive unlimited work. A per-call or per-request tool can become very expensive very quickly, because an agent does in a single day what a person would spread across a month. Before you let automation loose on a tool you pay for by usage, you need to know how that tool charges, and what happens to the bill when the volume goes up by two orders of magnitude.

This is also a warning about delegation. An agent that can spend money on your behalf — by making API calls, provisioning infrastructure, or signing up for services — needs the same guardrails you would put on any junior employee with a company card. The habit of being deliberate about what automation is allowed to decide, and what must come back to a person, is the subject of decisions automation should never make. A $240,000 surprise is cheaper to prevent than to explain.

Pricing is being rebuilt around AI

Sit on the vendor side of that same shift and you get the third report. Chargebee, a billing company the publisher notes "has been around 14+ years and runs billing for 6,500+ businesses" [3], was highlighted specifically for how central it has become to how AI companies charge. The telling observation was about its customers: "Every AI company we talk to has changed its pricing at least twice" [3], usually because the old model stopped matching how the product was actually used.

For a business choosing tools, this is a signal to read the pricing page carefully and expect it to move. Seat-based pricing assumes a human at a desk. Usage-based pricing assumes you can predict your usage. AI features break both assumptions, because a single person armed with an agent can generate the workload of a whole team. Vendors are rebuilding their pricing around consumption — tokens, actions, runs — precisely because that is where their own costs now sit. When you commit to a tool today, ask not only what it costs this month but how it charges for the thing you will do more of next year. The moment a tool stops fitting the way you work is the moment to look again, which is the thread running through when a spreadsheet stops being enough.

At 360REV we price AI work by the work itself, not by the machine that runs it, which is our attempt to keep the bill legible as usage climbs. That is one answer among several; the point for any buyer is to understand the model before the volume arrives.

A re-rating worth understanding

The fourth report is about the market rather than a single tool, but it carries a lesson for buyers too. Okta's shares roughly tripled on what the publisher described as around 11% growth — a large move for a modest headline number. Only months earlier, the mood was the opposite: "In April, Okta was one of the most hated names in B2B." [2] The report points to the forward indicators behind the turn, including 14% growth in committed remaining performance and about 30% of new bookings coming from products the company had launched more recently [2].

Why should a business picking software care about a share price? Because the same numbers tell you something about the vendor you are about to depend on. Headline growth is a rear-view mirror. Committed future revenue tells you how much customers have agreed to keep paying. The share of sales coming from newer products tells you whether a vendor can still ship things people want, or is living off what it built years ago. A tool from a company whose customers keep committing, and keep adopting its newer work, is a safer long-term bet than one coasting on a single ageing product — regardless of what this quarter's growth rate happens to be. The habit of looking past the headline figure to the number that actually changes a decision is one we have written about in the numbers that change a decision, and it applies as much to choosing a supplier as to running your own business.

The thread

Four reports, one direction of travel. Agents are now cheap enough to build the tools you used to buy, hungry enough to blow up usage-based bills, and central enough that vendors are redrawing their pricing to match. The decisions themselves — build or buy, how to charge, which numbers to trust — are the same ones businesses have always faced. The answers are moving because a new participant has joined the table. The teams that come out ahead will be the ones who let the agent propose and keep a person deciding. For yesterday's items, see the briefing for 2 October.