Putting a price on AI agents

The announcements worth reading today share one theme. The industry has stopped asking whether an AI agent can do a job and started arguing about what that work costs and who should own it. For a business choosing tools, that shift matters more than any single feature, because it changes how you read a price and how far you trust a vendor to still be standing next year.

Most of what follows is about judgement, not procurement. The point of a daily read like this is to help you tell a durable decision from a reaction to a headline. Where 360REV handles one of these ideas we will say so briefly, but the ideas come first.

What AI work actually costs

The most useful piece of the day is also the least flashy. GitHub wrote about making AI coding cheaper without lowering the quality of the result, and the framing is worth borrowing well beyond code: "Why shorter outputs can cost more, and how GitHub Copilot reduces wasted work across the complete coding task."[1]

The instinct most buyers carry is that a shorter answer is a cheaper answer. That is only true per token. The cost that matters is the cost of finishing the task, and a terse reply that is wrong, or that leaves you to do three more rounds of prompting, can cost more in total than a longer reply that lands the first time. When you evaluate any AI-priced tool, ask what unit it charges for and whether that unit lines up with work you actually value. A price per message rewards a vendor for making you ask twice. A price tied to a completed outcome does not.

This is the same discipline we have written about before when reading a plan. A price is a claim about what you are buying, and the only way to test it is to trace it back to a job you recognise. If you have not read it, what a subscription plan is really selling covers the habit in full.

What it takes to ship an agent worth buying

SaaStr published a conversation with the chief product officers of Harvey, Glean and Rubrik on shipping an agent that wins its category. The piece opens by marking how much the job has changed: "For about ten years, chief product officer was the best job in B2B."[2]

The reason that era ended is the reason this is relevant to a buyer. Shipping a feature and shipping an agent are not the same discipline. A feature does one defined thing on demand. An agent takes an open-ended goal, makes a chain of decisions, and is judged on whether the end state is right. That is far harder to build and far harder to trust, which is exactly why the people who build these things are now talking publicly about what separates a real product from a demo.

You can use that distinction as a buyer. When a vendor shows you an agent, watch what happens at the edges: what it does when the input is messy, where it stops and asks a human, how it explains a choice you did not expect. A polished happy path tells you almost nothing. The handling of the awkward cases tells you whether anyone shipped a product or a slide.

The market is consolidating around automation

TechCrunch reported that Palo Alto Networks paid 500 million dollars for Console, and noted what the deal leaves behind: "The acquisition also leaves Sequoia-backed Serval as the de facto startup leader in AI IT service automation, industry watchers believe."[3]

For a business choosing tools, an acquisition is not gossip. It is information about the life expectancy of a product you might depend on. When a large platform buys a promising startup, the product you liked often changes shape: it gets folded into a suite, repriced, or pointed at a different buyer than the one it was built for. The startups still standing after a round of buying become the independent options, and the field narrows.

None of this is a reason to avoid a young vendor. It is a reason to ask, before you commit, how hard it would be to leave. Can you get your data out in a usable form on any ordinary day, without a support ticket and without a specialist? If the answer is yes, a change of ownership is an inconvenience. If the answer is no, it is a risk you are carrying whether or not you ever thought about it. The tools that talk to each other on open terms are the ones you can walk away from, which is the quieter argument in why your tools do not talk to each other.

An agent for every account, not just the big ones

The clearest example of an agent aimed at a real business process came from SaaStr, describing a renewal agent that builds a tailored renewal deck for every account rather than only the largest. The piece names the problem directly: "Every company cuts the same corner on renewals."[4]

The corner is familiar. Your biggest customers get a custom renewal built on their actual results. Everyone else gets a template, because a person only has so many hours. This is the kind of task that suits automation well, because the bottleneck was never judgement. It was capacity. The work of pulling a customer's own numbers into a tailored summary is repetitive and rule-bound, and spreading it across every account rather than the top few is a straightforward win.

It is also a good place to be clear about the line between help and harm. An agent that assembles the facts of a renewal is doing honest work. An agent that decides which customers to keep, or sets a price without a human seeing it, has crossed into territory that belongs to a person. We have written about where that line sits in decisions automation should never make; the short version is that automation should prepare a decision, not make the ones that carry real consequence for a relationship.

Turning documents into something people will actually watch

Google added a feature to Vids that turns written material into a short video. In its own words, "Google Vids now allows you to transform static Google Docs, PDFs, and Word files into engaging video summaries."[5]

The appeal is obvious to anyone who has written a careful document that nobody read. A video summary meets people where their attention already is. The trade-off is just as plain, and it is the trade-off attached to every generative feature in a suite: the tool writes a script and narration for you, which means the tool can also get it wrong in your name. Treat the output as a draft that a person signs off, not as finished work that ships on its own. That habit — generate freely, review before it leaves the building — is the one worth carrying into every AI feature your suite adds this year, and it is the heart of what AI should and should not do in your business.

Taken together, the day reads as a market maturing. The question has moved from capability to cost, durability and judgement, and those are the questions a careful buyer was always going to ask.