Build it yourself or buy it, and what you pay either way

Two threads run through today's productivity announcements. The first is the old buy-or-build question, asked again now that anyone can point an AI at a problem and get something working. The second sits underneath it: once AI is woven into the tools you already run, the thing you pay for stops being a seat and starts being usage you can measure by the token.

Neither thread is abstract. If you run a small business, both change how you should read a pricing page and how you should scope your next tool. The sections below take the day's items in order of how much they should weigh on that reading.

The case against building your own CRM

The most useful piece of the day is a plain rebuttal to a claim doing the rounds. Someone on X argued they had built AI agents that replaced a full CRM without being technically inclined, and SaaStr takes that claim seriously before pulling it apart. As they put it, the claim "has truth in, but also radically overstates things" [1]. Both halves matter. The truth is that you genuinely can assemble something that captures leads, sends follow-ups, and stores notes, faster and cheaper than ever. The overstatement is in what happens next.

A CRM is not the demo you build in an afternoon. It is the years of edge cases a mature product has already absorbed: deduplication, permissions, audit history, integrations that keep working when a vendor changes an API, and the quiet reliability that lets you stop thinking about the tool. Building the first ten per cent is now trivial. Owning the last ninety — maintaining it, securing it, and trusting it with the record of every customer relationship — is the part that does not get cheaper because the first part did.

The practical rule is to separate what is core to your business from what is merely necessary to it. Build where the logic is genuinely yours and no vendor can know it. Buy where the problem is common and someone has already spent a decade solving it. We take a firm position on the same line — the decisions automation should never make are the ones where being wrong costs a relationship, and those stay with a person regardless of what you build or buy.

What you are actually paying for: tokens

If you buy AI-powered tools, the second item today explains the meter you are now on. Zapier's primer on tokens makes a point that has crept up on a lot of buyers: tokens "have gone from a background technical detail to the primary usage limit and billing unit for top AI models" [2]. A token is a fragment of text — roughly a short word or part of one — and both what you send a model and what it sends back are counted in them.

This matters for procurement, not just for engineers. When a tool prices AI by usage, two customers on the same plan can pay very different amounts, because one asks the model to read and write far more text than the other. A long document summarised every morning costs more than a one-line classification, even though both are "one AI action" in the marketing. Before you commit to an AI feature, ask the vendor what a typical task consumes and where the ceiling sits. A plan that looks generous can turn tight the first time your team leans on it.

The honest version of this is a bill you can read: which action cost what, and why. If you cannot get a straight answer to that question, treat the pricing as unknown rather than cheap. We wrote separately about what a subscription plan is really selling, and usage-metered AI is the clearest case yet of a plan selling you a number you have to understand before you sign.

When a spreadsheet is no longer the tool

The third item is a survey of database-powered app builders, and its opening is the most honest sentence in it. Zapier grants that a spreadsheet is genuinely capable: "You can put together an accounting system, a task manager, or an inventory tracker with columns, rows, and formulas—all without slamming into a wall of code at any point" [3]. That is true, and it is why so many businesses run for years on a shared sheet.

The question is when the sheet stops paying its way. The wall is usually not features — it is people. A spreadsheet has no real notion of who may see which row, no record of who changed what, and no way to stop two people overwriting each other. When those become daily problems, a database-backed tool earns its cost. The builders in this category sit between a spreadsheet and custom software: structured data, permissions, and views, without a full development project. We covered the same transition in when a spreadsheet stops being enough, and today's list is a reasonable map of where to look once you have decided to move.

A permission that shares access without sharing detail

Google Workspace added a calendar sharing level worth noting because it fixes a specific, common bind. The new option, "Make changes (see private events as free/busy)", grants edit access while hiding the substance of private entries. In Google's words, it "allows you to grant someone edit access to your calendar while keeping the details of your private events entirely hidden" [4].

The reason this belongs in a briefing about choosing tools is that it is a small, correct model of permissions in general. Good access control is not one switch between all and nothing. It lets you say yes to the task — an assistant who can book and move meetings — without saying yes to everything else that role could have seen. When you evaluate any tool that holds sensitive data, this is the shape to look for: can you grant the narrow thing someone needs, and only that. A tool that forces you to over-share to get work done is a liability that grows with your team.

AI moving into tools you already open

The last thread is quieter and, over a year, probably the most consequential. Zapier's piece on ambient AI names the friction that today's assistants still carry: chatbots "keep getting smarter, but they're creating a new kind of busywork" [5]. Every answer still costs you the trip — open the app, start a chat, prompt your way to the goal. Ambient AI is the argument that the assistant should meet you inside the work instead of asking you to come to it.

You can see the same direction in Google extending Fill with Gemini in Sheets to eleven more languages. The feature was introduced, in Google's description, as "a new AI-powered feature designed to make data preparation and manual entry even easier" [6]; widening its language coverage is how a capability stops being a demo and becomes something a global team actually uses. The pattern to watch is not the launch of a new AI product but AI appearing inside the sheet, the calendar, and the inbox you already had open.

For a business choosing tools, that changes the question. It is less "which AI app should we adopt" and more "which tools we already trust are adding AI in a way we can see, price, and switch off if it misbehaves." The buy-or-build decision from the top of this briefing and the token meter in the middle both fold back into that. Prefer AI that arrives where the work is, that bills in units you can read, and that you can turn off without losing the tool underneath it.