When the model is the easy part of buying software

The theme running through today is that the hardest part of adopting AI is no longer the model — it is the plumbing, the cost accounting, and the trust that sits around it. Several of today's announcements are less about what software can do and more about what it costs to run, who it is allowed to act as, and how it fits with the tools a team already uses.

The value is moving from the model to the implementation

The most consequential news for anyone buying business software is a shift in where the money and the effort are expected to go. TechCrunch reports that a new company, Ode, has launched on the premise that the next large AI business is not another model but the work of putting models to use inside real organisations [1]. The framing is that embedding engineers directly inside enterprises — people whose job is to make the technology actually run against a company's own data and processes — is what unlocks adoption.

For a business choosing tools, this is a useful correction to a common assumption. The instinct is to compare products by the intelligence of the underlying model. The more honest question is how much work it will take to get a tool doing something useful in your particular setup, with your data, your people, and your existing systems. A capable model that never connects to your actual workflow delivers nothing. A modest one that is wired into the way you work every day earns its keep. When you evaluate software, weigh the implementation cost as heavily as the feature list, and ask the vendor concretely who does that work and how long it takes. This is the same discipline we cover in choosing software worth using: the demo is the easy part, and the integration is where the value is won or lost.

An hour of an AI agent, priced in dollars

SaaStr published a short, pointed piece today about the running cost of a capable AI agent. The headline figure is that an hour with their top agent came to $13.42, and the argument is that no human can be hired for that rate [2]. The account describes their AI agent as having wrapped a focused work session — a discrete block of real output measured against a real bill.

Two things matter here for a buyer. First, AI work now has a unit cost you can read off a statement, the same way you read electricity or a contractor's invoice. That is a healthier way to think about it than "unlimited AI included", because a per-hour or per-task figure lets you compare the machine against the alternative honestly. Second, a low hourly number is only meaningful if the work produced is work you would have paid for anyway. Cheap output that nobody needed is not a saving.

This is why we argue that the decision to automate a task should be made deliberately, not by default. The question is not whether an agent is cheaper per hour than a person — it usually is — but whether the task should be handed to a machine at all. We set out that reasoning in what AI should and should not do in your business: price is one input, judgement about the task is the other.

Who is your AI agent allowed to be

As software starts to act on your behalf, a quieter problem surfaces: identity. When a person logs in, you know who they are and what they may touch. When an AI agent does work, it needs credentials too — and those credentials can proliferate quietly, with no clear owner. TechCrunch reports that Oak has emerged from stealth with $60 million in seed funding to work on identity management, in a market it links to the rise of AI agents [3].

You do not need to buy a specialised product to take the point seriously. Every automation you set up acts as somebody or something. It holds a key. If you cannot say which systems a given automation can reach, what it is allowed to change, and how to revoke it in one step, you have handed out access you cannot account for. The number of these non-human actors inside a business tends to grow faster than anyone tracks, because each one is created to solve a small problem and then forgotten.

The defensive move is the same one that applies to human access. Keep one clear record of who — and what — can reach each system, grant the narrowest access that does the job, and make revocation a single action rather than an archaeology project. We wrote about the human side of this in one login and why it matters; the arrival of AI agents raises the stakes, because an agent can act far faster than a person can, and it never sleeps.

Working with people outside your walls

Not every announcement today is about AI. Google Workspace shipped a practical change: group conversations in Google Chat can now include people from outside your organisation [4]. The company's reasoning is plain — as it puts it, "For many teams, it's essential to be able to work in real-time with partners from outside your organization."

This is a small feature with a large implication for tool choice. A great deal of real work happens with people who are not on your payroll: contractors, agency partners, suppliers, clients. If your collaboration tools stop at the edge of your company, that work leaks into email threads and personal messaging apps where it cannot be searched, governed, or handed over when someone leaves. Bringing external partners into the same managed conversation keeps the record in one place.

The trade-off to weigh is governance. Every external participant you admit is someone who can now see part of your conversation, so the same tool that makes cross-company work easier also widens the circle of who has access. The right posture is to use the feature deliberately — invite the people who need to be there, and keep track of who is in each conversation — rather than treating an external group as no different from an internal one.

Is software dead? No — but winning is harder

Finally, a wider piece of perspective. SaaStr sat down with investor Rory O'Driscoll, who has been backing software companies for more than thirty years, and the conclusion in the title is worth sitting with: software is not dead, it "Just Got a Lot Harder to Win" [5].

For a buyer, that difficulty on the seller's side is quietly good news and a caution at once. Good news, because more competition and more AI-assisted building means more capable tools reaching the market faster. A caution, because harder to win also means more products chasing the same problem, more churn among young companies, and more noise to see through when you are choosing. The tool that demos best this quarter is not guaranteed to be the tool that is still supported and improving in two years.

The practical response has not changed. Judge a tool by whether it fits how you actually work, whether you can get your data out of it on any ordinary day, and whether the people behind it will still be answering the phone when you need them. The market getting harder for sellers does not change what makes a tool worth buying — it just makes the choosing matter more.