You’re not buying AI.
You’re buying 8 things.

In plain English. The humans and habits around the technology: who decides, who checks, who is trained, and how work changes. It is the layer most diagrams leave out.

Where the power sits. With whoever owns the decision at the end. Tools are available to everyone, so the difference comes from judgment, empathy and creativity.

Ask these. Who is accountable for each AI-assisted decision? What do we review by hand, and why? How are people trained, not just given access? What work do we want to keep human on purpose? How will we notice if quality slips?

Cost shape. Mostly time: training, change management, review. Easy to underfund because it never appears on a vendor quote.

How it goes wrong. Staff trust the output too much, or avoid the tool entirely. Nobody owns the result when it is wrong.

The last 10%. Everything on this layer. AI does the heavy lifting; you own the call.

In plain English. The rules, controls and records that keep AI use safe, legal and explainable.

Where the power sits. With whoever can say no, or can show an auditor or regulator what happened and why.

Ask these. What data may go into which tool? Who approves a new AI use? Can we show what the system did and who reviewed it? What happens when it errs or leaks? Which rules apply in each place we operate?

Cost shape. Mostly fixed and front-loaded: policy, review, tooling. Far cheaper than fixing a problem afterwards.

How it goes wrong. Staff use unapproved tools because the approved path is slow. Policies exist on paper only.

The last 10%. Deciding what risk the organization accepts. That is a leadership call, not a technical one.

In plain English. The products people actually use: assistants, copilots, industry software with AI built in, or something you build yourself.

Where the power sits. With whoever owns the workflow and the relationship with your users. Changing an app means retraining people.

Ask these. What specific task gets faster or better, and how will we measure it? Could our existing software add this next quarter? Who pays when usage grows? How will we know it is being used?

Cost shape. Per-seat or per-use fees that rise with success. Check the price at ten times today’s usage.

How it goes wrong. Buying a demo, not a workflow. Paying for many overlapping tools nobody has compared.

The last 10%. Choosing which problems are worth solving at all.

In plain English. The plumbing and rules that connect a model to your tools, data and steps, so it can do a task from start to finish. An agent is a model allowed to take actions, not just answer questions.

Where the power sits. With whoever defines the workflow and the permissions. This is where a helpful answer becomes an action with consequences.

Ask these. What can the agent do without asking a person? Where is the approval step? Can we see every action it took? What is the worst thing it can do with the access we gave it?

Cost shape. Variable and easy to underestimate, because one task can trigger many model calls.

How it goes wrong. Too much access, too little oversight. No one watching what the agent does day to day.

The last 10%. Deciding which actions need a human sign-off, and holding to it when speed tempts you.

In plain English. The trained “brain” that reads, writes, reasons and generates. Some are closed, reached through a provider. Some are open-weight, meaning you can download and run them yourself under a license.

Where the power sits. With whoever has the right model for your task, which changes often. Different tasks suit different models.

Ask these. Which tasks need the strongest model, and which only need a good-enough, cheaper one? Where does our data go? What do the license and indemnity terms say? Who tested it on our real work, not a benchmark?

Cost shape. Closed: pay as you use. Open-weight: lower per-use price, but you carry hosting, security and talent costs.

How it goes wrong. Treating the choice as a belief system. Assuming “open” means free, or “closed” means safe.

The last 10%. Judging whether an output is good enough for your customers.

In plain English. The information a model learns from or draws on: your documents, records, customer history, and the examples humans use to teach and test systems.

Where the power sits. With whoever has data nobody else has, in a usable shape. Models are widely available; your clean, permitted, well-organized data is not.

Ask these. What do we have that competitors do not? Are we allowed to use it this way? Is it accurate, current and findable? Who can see what?

Cost shape. Often the biggest hidden cost: cleaning, organizing, permissions, keeping it current.

How it goes wrong. Messy or out-of-date data produces confident wrong answers. Sensitive data reaches a tool it should not.

The last 10%. Deciding what data you should not use, even if you could.

In plain English. The rented or owned computers, storage and networks where AI runs, plus the tooling around them.

Where the power sits. With the people who run the platform your data and systems already live on. It decides your speed, reliability and much of your bill.

Ask these. What does this cost at today’s usage and at ten times today’s? Where is our data stored, in which country? What are the security and uptime commitments? Who watches spend so a pilot does not become a surprise bill?

Cost shape. Committed spend plus usage. AI workloads can grow quickly, so track cost per task from day one.

How it goes wrong. Sizing commitments before usage is known. No one owns the monthly bill.

The last 10%. Deciding what is worth running at all, and setting a budget for it.

In plain English. The specialized processors and data centers that train and run AI. Think of it as the power plant under everything else.

Where the power sits. With whoever can make and secure scarce capacity. Supply, energy and geopolitics all shape it.

Ask these. Is capacity guaranteed for us, or first come, first served? How do price changes pass through to us? What does a supply shock do to our plan? Is there an efficiency path, such as a smaller model, that cuts our need?

Cost shape. Mostly invisible to a business buyer, but it flows into every price on the layers above.

How it goes wrong. Plans that assume prices and availability stay as they are today.

The last 10%. Deciding how much capability you really need, instead of buying the biggest option by default.