When agents, not people, choose your software
Across today's announcements, the thread is that the thing choosing and running software is more and more an agent rather than a person. That changes what it means for a tool to earn its place: it now has to be reachable, cheap enough to call, and trusted enough to act, not merely pleasant for a human to open.
When an agent decides what stays, not a person
For years the way a tool left a business was simple. A person grew frustrated, or found something cheaper, or a budget review forced a cut. Someone made a decision. The report from SaaStr this week describes a quieter path. Two long-standing Canva users cancelled, and the reason was not anger. "We weren't angry at it, and nobody sat down and decided to replace it." [1] The work had moved to agents, and the agents never reached for that tool.
This matters for anyone choosing software now. The question is no longer only whether your team will like a tool. It is also whether the agents you run will actually call it. A tool that a human enjoys but an agent never invokes can fall out of use without a single complaint. That is a different kind of churn, and it is harder to see coming, because nobody files a reason for it.
The practical lesson is about reachability. A tool that exposes a clean interface an agent can drive stays in the flow of work. One that only a person can operate is at the mercy of whether a person still opens it. If you are consolidating your stack — a theme we covered in replace five business tools with one — it is worth asking which of your current tools an agent could operate on your behalf, and which depend entirely on a human remembering to.
Spreading AI across a whole practice, not one desk
Xero's announcement of Xforce speaks to a problem many businesses already have. AI inside a company is often scattered. Xero puts it plainly: the trouble is "it's fragmented: one team's workflow here, one person's prompt there, and nothing that holds up across the whole practice." [2] One person has a prompt that works. A team has a workflow nobody else knows about. None of it survives a staff change.
Understand the idea before any tool. Fragmented AI carries three costs. The first is fragility: when the person who wrote the prompt leaves, the capability leaves with them. The second is inconsistency: two teams solve the same task two different ways, and you cannot tell which is right. The third is risk: you cannot govern what you cannot see. If prompts live in private chats, you have no way to review them for accuracy or for what data they touch.
The remedy any business should look for is shared, visible workflows rather than private ones. Whether or not you adopt a specific product, the test is the same: can the whole practice rely on a workflow, or does it depend on one person remembering a trick? A capability that only lives in someone's head is not a capability your business owns yet.
An agent that tells leadership what it would rather not hear
Information tends to get cleaner as it travels up an organisation. Problems get rounded off, and by the time news reaches the top it has lost its edges. TechCrunch reports on a Databricks agent aimed squarely at that gap. "With Databricks at 12,000 employees, AI Ghodsi is the first to admit that he can't know everyone and that they are unlikely to confess the things that go wrong to him." [3] The idea is an agent that surfaces what people might not say directly.
For a business choosing tools, the lesson is not to buy this one agent. It is to notice what the role implies. An agent that reads across your data and reports plainly is only as good as the data it can reach and the honesty you are willing to act on. If leadership asks for an unvarnished view and then punishes it, the tool changes nothing. The trade-off here is cultural as much as technical: a candid agent is useful only in a place that wants candour.
Software that improves its own models after it ships
Most AI tools are fixed once they ship. They are trained, deployed, and left until a vendor pushes an update. Salesforce described something different today. "SFR-AutoR&D, a new self-improving system from Salesforce, trains and improves the embedding models in deployment-grade enterprise AI." [4] Embedding models are the part of a system that turns text into the numbers used for search and matching, so improving them quietly improves how well the system finds the right thing.
The concept worth taking away is self-improvement after deployment. It is appealing because a tool can get better without a human doing the tuning. It also carries a trade-off. A system that changes itself is harder to pin down. If results shift, you need a way to see what changed and why. For a buyer, the question is not whether self-improvement sounds good. It is whether the vendor gives you visibility into the changes a system makes on its own.
Reading an automation bill before it surprises you
Automation platforms often price by consumption, and consumption is hard to predict before you run anything. Zapier's write-up on Power Automate opens on exactly this discomfort. "The first time someone at a Microsoft shop notices their Power Automate bill is higher than expected, they usually look for a pricing page." [5] The surprise is the point. You commit to building on a platform, then learn the real cost once the work is already running.
The lesson generalises beyond any one vendor. When pricing scales with usage, the number that matters is not the headline plan. It is the marginal cost of one more run, one more step, one more agent action — and whether you can see that cost before you are committed. This becomes sharper as agents, not people, trigger the runs, because an agent can call a tool far more often than a person would. We looked at related tool-cost questions in yesterday's briefing.
The thread
The common thread is control. Agents now choose tools, improve themselves, and speak up the chain — and each shift moves a decision away from a person and towards a system. The job of anyone buying software this week is to keep that system visible: know what your agents call, what your workflows contain, and what your bill will be before it arrives. At 360REV we build towards that kind of visibility, so the tools you depend on stay something you can see and account for.
Sources
- [1] Our Agents Never Picked Canva. 10 Learnings on Agent Pricing, Getting Picked, and Security From This Week's "The Agents" With Ulrik Lehrskov-Schmidt — SaaStr
- [2] Meet Xforce by Xero: AI workflows built alongside your firm — Xero
- [3] Databricks' new AI agent tells CEO Ali Ghodsi the unvarnished truth — TechCrunch
- [4] Salesforce Releases R&D Agents That Improve Its AI Models — Salesforce
- [5] Power Automate pricing and plans for 2026 — Zapier