Most of the AI you need, you are already paying for

The fastest way to spend money on AI badly is to treat it as a category you buy rather than a capability you already have some of. Most business software has shipped AI features recently, and they tend to arrive switched off, or switched on and never mentioned to anyone.

The order to buy in

There is a sequence here, and it is almost the reverse of how AI gets sold.

How it usually goes

1. 01Start with a tool you saw demonstrated
2. 02Find a process it could apply to
3. 03Discover the process was never documented
4. 04Scope grows to cover the documentation
5. 05Pilot ends, nothing reaches production

How it should go

1. 01Start with a process that runs often and hurts
2. 02Check what your existing stack already does for it
3. 03Trial that for a month against a real number
4. 04Buy a point tool only where the gap is real
5. 05Build only where no product covers the workflow

First, what you already own. Your CRM, your helpdesk, your office suite and your accounting system have all shipped AI features. They are included, they are already inside your security perimeter, and they improve on the vendor's roadmap rather than yours. They are also, frequently, unconfigured, switched off by default or switched on and never explained to anyone.

Second, a point tool where the gap is real. Transcription, document extraction, translation. These are mature, cheap, and genuinely better as dedicated products than as a feature bolted onto something else.

Third, a horizontal assistant. Useful, hard to measure, and worth buying on the assumption that the measurement will be indirect. Do not build a business case that depends on proving it; buy it or do not.

Fourth, and only fourth, something custom. Where the workflow is specific to how your business operates and no product covers it.

Why the order matters more than the choice

Each rung costs more than the one above it and takes longer to show a result. Skipping to rung four is how a company ends up with a bespoke system that does something a setting in their existing CRM would have done, except the setting would have kept working after the developer moved on.

The questions that actually separate tools

Once you are genuinely choosing between products, the feature comparison is the least useful part of the exercise. What differs, and what rarely appears on a pricing page:

• Where the data goes, and on what terms. Enterprise agreements with major model providers commonly exclude your inputs from training. Consumer and low tier plans frequently do not. This is a contract question, not a technical one, and the answer changes by tier within the same product.
• Whether it fails loudly. A tool that tells you it is unsure is worth more than a more capable one that guesses with confidence. In any workflow where a person checks the output, the checking cost is the real cost.
• What happens at your volume. Per-seat pricing and per-use pricing behave very differently once a process is genuinely adopted. Several tools are cheap in a pilot and unaffordable at full rollout, which is a discovery better made before the rollout.
• Whether anyone will use it. A tool that fails outright gets cancelled. A tool nobody opens gets renewed, because cancelling it requires someone to decide it was a mistake.

What to do this week

Pull your software invoices and, for each line, find out what AI features the current tier includes. Most vendors publish this badly, so it is a real exercise rather than a five-minute one. Then pick the single process in your business that runs most often and annoys people most, and check whether anything on that list already addresses it.

That exercise costs nothing, and anything your current tier already covers is a line item you do not have to add.

Where this goes next

The build case gets its own treatment in build an AI agent or buy an AI tool, and the reason most trials never become systems is in why AI pilots do not reach production.

Working out which of this applies to your stack, without a reseller margin pointing at an answer, is AI consulting.