Essay 07

What to buy and what to skip in year one

Your inbox is full of AI products, and every one of them sounds necessary. You do not want to fall behind, and you do not want to waste money finding out.

In the first year, the useful spending is small and boring. The expensive mistakes are usually large and early.

This is not a product list. Products change every few months. These are categories, and the questions worth asking before you pay for any of them.

Usually worth buying

Assistant seats for the people on your first process. A business plan of a common assistant such as ChatGPT or Microsoft Copilot, for people whose jobs are heavy on email, documents, summaries, and drafts. Not for everyone at once. Start with the people attached to the process you picked. Buy the business tier rather than letting people use personal accounts, so company information stays under company terms and you can see who has access.

The AI features in software you already pay for. Accounting, CRM, help desk, and document tools are all adding AI features. Turning on a feature inside a system people already live in is often more useful than adding a new system. Check what it costs, what data it uses, and whether you can turn it on for one team first.

A small amount of outside help, scoped to one process. A few weeks of someone setting up one process properly, with your process owner in the room, is worth more than a large outside project that ends in a roadmap. Ask for something working at the end, not a document.

Training tied to the actual work. Not a general course on AI. An hour with the team on the process they are changing, using their real documents and their real examples.

Usually worth skipping in year one

A custom model. Almost no company this size needs to build or train its own. The common tools are good enough for the first several processes, and the cost of building and maintaining something custom is hard to justify this early.

An all in one AI platform bought before you know what you need. These are sold on breadth. Breadth is the problem in year one. You need one step to work, not forty possible ones.

Company wide licenses as the plan. Buying a seat for everyone feels decisive. Use tends to follow the people who already had a reason for it. Buy for the process, then expand when there is a reason to.

Tools a department signed up for without anyone checking the data terms. Experimenting is fine. Company information sitting in accounts nobody tracks is not. Make it easy to ask before signing up, and quick to get an answer.

Long contracts before a trial. If a vendor will not let you test on a small scope first, that tells you something.

Questions to ask before you pay

1. Which process does this change, and who owns it?
2. What does it connect to, and who sets that up?
3. Where does our data go, who can see it, and is it used to train anything?
4. What does it cost at the size we would actually use it, not the starter tier?
5. Can we start with one team and stop cleanly if it does not work?
6. What do we stop paying for, or stop doing, if this works?

The last question is the one most purchases skip. If nothing gets retired, the spending only adds up.

A sensible order

First, assistant seats for the people on your first process, under business terms. Second, the AI features in systems you already use. Third, outside help on one process if your team needs it. Anything bigger waits until a few processes have actually changed and you can see the pattern.

On Monday

Ask finance for a list of every AI related subscription the company already pays for, including the ones on personal cards that get expensed. It is usually longer than owners expect. Cancel whatever nobody can tie to a process. Then decide what you need for the one process you picked.