Putting AI where the work actually happens

The day's announcements circle a single question, and it is no longer whether to use AI. It is where to put it and who stays in charge of it once it is there. A platform vendor framing AI as an enterprise opportunity, a startup turning recorded sales calls into agent playbooks, and practical guidance on moving past ad hoc prompting all point the same way: the tools worth choosing this year are the ones that fold AI into real work, not the ones that leave it sitting in a tab you remember to open.

The bigger story is not AI replacing software

The loudest narrative about business software right now is that AI is dismantling it — replacing seats, collapsing pricing, eating the workflow. It is worth reading the argument that this framing, while not wrong, is not the main event. One industry commentary published this week opens by naming exactly that reflex: "Everyone wants the story to be about AI." [1] The point it goes on to make is that many of the problems attributed to AI were already present in how software companies priced, retained, and served customers.

For a business choosing tools, the practical lesson is to separate two things that get merged in the noise. One is a genuine change in what software can do. The other is a vendor's own commercial trouble dressed up as a technology story. When you evaluate a product, ask what it does for your work today, not what category-level disruption its marketing invokes. A tool earns its place by the job it does, and that test does not change because the surrounding conversation is loud.

Turning what your best people do into something repeatable

Much of the value a small business loses is tacit. Your most effective salesperson knows which objection to answer first and when to stop talking, and that knowledge usually leaves when they do. A startup called Encore AI raised $30M this week to attack exactly that gap. It "analyzes calls, messages, and CRM data to identify effective sales techniques and turn them into playbooks for AI agents." [2]

Set the specific product aside and look at the pattern, because it is the one to understand. The idea is to capture what already works inside your own conversations and turn it into something the next person — or the next automated step — can follow. That is a reasonable ambition, and it rests on one precondition most companies underrate: the underlying record has to be worth learning from. If your calls, messages, and customer history are scattered across tools that do not talk to each other, there is no clean corpus for anything, human or machine, to learn from. Before you buy a system that learns from your data, it is worth being honest about whether your data is in one place and current. We have written before about what a customer record should contain, and the same discipline decides whether this kind of tool has anything to work with.

Moving past the open browser tab

There is a wide gap between using AI and building it into how work gets done. The common state in most companies is a chat window someone keeps open and pastes into when they remember. Guidance published this week names that state plainly: "If your company's idea of 'using AI' is keeping ChatGPT open in a browser tab—congrats, you're doing the bare minimum." [4]

The distinction matters because ad hoc prompting and workflow automation solve different problems. Prompting is good for one-off requests. Automation is for the repeated step that happens whether or not anyone remembers to trigger it — the follow-up that should go out, the record that should update, the summary that should land before a meeting. The move from one to the other is not about a cleverer prompt. It is about deciding which repeated tasks are stable enough to hand over, and then wiring them so they run on their own. That decision deserves care rather than enthusiasm; we set out a way to make it in how to tell whether a task should be automated. The short version is that a task is a candidate when its inputs and its rules are predictable, and a poor one when either still needs judgement every time.

When a platform vendor calls AI an enterprise opportunity

On its second-quarter earnings call, Meta signalled how broadly it intends to sell into business. Its chief executive said the company sees a "large enterprise opportunity" spanning AI agents, APIs, compute, and internal software. [3] The breadth of that list is the thing to notice. It is not a single product; it is agents, the interfaces to reach them, the compute beneath them, and the software a company runs internally.

For a buyer, a large vendor describing an opportunity that wide is a signal to read carefully rather than a reason to act. When one supplier offers the model, the interface, the infrastructure, and the applications together, the convenience is real and so is the concentration. The more layers you take from a single vendor, the harder it is to move later and the more a change on their side ripples through yours. None of that makes such an offer wrong. It means the question to ask is what you could still do if you needed to leave — whether you could export your data and carry your process elsewhere. We made that case in the data you should be able to export on any Tuesday, and it applies with particular force when the vendor is large enough to want every layer.

Keeping automation from becoming noise

Not every automation story this week was about AI, and one of the most useful was about the opposite problem: automation that works so eagerly it drowns the people it serves. GitHub published guidance on taming Dependabot, the tool that keeps software dependencies current. It states the trade-off directly: "Dependabot keeps your dependencies current, but its defaults can flood your repository with pull requests." [5]

The remedy described — grouping related updates, slowing the routine cadence, while keeping security fixes fast — is a good template for any automation you run, not only this one. The principle generalises. An automated process that fires on every small change trains people to ignore it, and an ignored alert is worse than none because it hides the one that matters. The fix is almost never to switch the automation off. It is to tune what it surfaces and how often, so that the urgent stays fast and the routine batches into something a person can actually read. Speed and signal are different settings, and good automation lets you choose each separately.

What faster actually changes

Most automation is sold on time saved, and that is the least interesting part of it. A Zapier account of a lawyer who made his practice several times faster reframes the question: "This one is about what happens after you save it." [6] The saved hour is only the input. What you do with it — take on clients you previously could not afford to serve, lower a price, or simply go home — is the actual decision, and it is a business decision, not a technical one.

That is the thread worth carrying out of the day. The tools are converging on the same promise, which is to do more of your repeated work for you. The value does not come from the promise; it comes from what you decide to do with the capacity it frees, and from keeping a person in charge of the judgements that should never be handed over. We have drawn that line before in what AI should and should not do in your business. The announcements will keep coming. The discipline of deciding where AI belongs, and where it does not, is the part that stays yours.