The Stack
Report

Start here, before you buy anything. The single most common finding when we look at a company’s AI spend is that they are paying three vendors to do one job, and that at least one of those jobs is already included in a tool they were paying for anyway.

This happens for a structural reason, not because anyone was careless. Between 2024 and 2026 essentially every incumbent software vendor shipped AI features into products you already licensed. Your note-taking app got a summariser. Your CRM got a writer. Your design tool got generative fill. Meanwhile individual teams, moving faster than procurement, bought point solutions for exactly those jobs.

The audit is one page and takes an afternoon:

1. List every AI tool anyone expenses. Include the $20/month personal subscriptions, in most companies that shadow spend is larger than the sanctioned spend.
2. Next to each, write the job, not the product category. Not “AI writing” but “drafts first-pass replies to inbound support email.”
3. Group by job. Any job with more than one tool against it is a candidate for consolidation.
4. For each remaining tool, check whether something you already pay for now does that job. This is the step that finds the money.
5. Only then look at what to buy.

Two numbers worth having before you start: 413 of the 716 tools in our index have a free tier, and the gap between the cheapest and dearest frontier model is roughly 385×. Both mean the same thing: the price you pay is much more a function of what you chose than what the task required.

There are three ways to add AI to a workflow, and most bad decisions come from not knowing which one you are looking at.

Buy

A finished product. Someone else owns the model, the prompt, the interface and the maintenance. Costs the most per seat and the least in time. Right for anything that is not your differentiator, which is nearly everything.

Wrap

You call a model API and build a thin layer of your own: your prompts, your data, your interface. Cheap per unit, and the layer is genuinely yours. Right when the workflow is specific to how your business works and no product fits it. The trap is underestimating maintenance: a wrapper is not finished when it works, it is finished when it still works after the model changes underneath it.

Build

Training or fine-tuning your own model. Right in a genuinely small number of cases: you have proprietary data nobody else has, the task is narrow and repetitive enough that a small model beats a general one, or per-unit API economics genuinely break at your volume. If you cannot state which of those three applies, you are not in this category.

The eighteen-month test. Price all three over eighteen months including the engineering time, not over one month including only the licence. Buying usually wins on a one-month view and loses on a three-year one; building loses on both more often than its advocates expect. Wrapping wins more often than people expect, because the expensive part, the model, is rented rather than owned.

Every figure here is drawn live from our index, which means this document is accurate on the day you read it rather than the day it was written. It also means it will disagree with itself over time. That is the intention.

If something in it is wrong, and in a market moving this fast, something will be, tell us and we will fix it. If you want a second opinion on a stack you are putting together, reply to the email this came with. We read everything.