Who actually chose the AI in your stack

The AI you use at work probably arrived without a decision. A new button appeared in a tool you already pay for. Your suite subscription moved to a tier with an assistant in it. The contractor who set up your laptops installed the assistant he holds a certificate in. None of those events involved you comparing anything, and none of them tell you the tool is right for your work. They tell you the vendor solved its distribution problem.

This guide is for a solo operator, freelancer or small team paying for somewhere between four and fifteen software seats, who now has AI in three or four of those products and is not sure what any of it costs or what sits underneath it. It is not for organizations with a procurement function, and it will not tell you which model is best; that answer changes every few months and depends on your work. What it will do is give you a way to price the AI you are already buying, find out whose model is running it, and read a recommendation for what it is.

The AI in your stack arrived through a channel, not a comparison

Model vendors have a distribution problem. Buyers who will pay for AI at scale get reached through the suppliers they already have and the advisers who already have their number, not through model cards. So the vendors buy the channel instead of the customer, and the larger deals are public enough to read.

The clearest recent example ran at enterprise scale. On 13 August 2026, IBM announced a partnership with OpenAI and said “IBM will also launch a dedicated OpenAI Practice, with thousands of consultants and engineers obtaining expert-level certifications under the OpenAI Partner Network”, folding GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage for clients in financial services, government, telecommunications and retail [1]. Reporting on the same announcement put the training figure higher, at tens of thousands of consultants, primarily retraining existing employees rather than hiring, and noted the deal came less than a year after IBM announced a similar alliance with Anthropic [2]. OpenAI’s own partner page describes the arrangement plainly, inviting organizations to “Build, co-sell, and deliver AI solutions with OpenAI”, and lists Accenture, BCG, McKinsey, PwC, KPMG, IBM, Cognizant, Infosys and Capgemini among its partners [3].

Two details in that outlast the news cycle. First, the same channel carried two different model vendors inside a year, so allegiance is rented, not held. Second, nothing in it is dishonest. A consultant certified on one vendor’s stack really is faster and safer with that stack, and recommending it is often the correct call. The problem is only that the recommendation and the certification produce the same words, and you cannot tell them apart from the outside.

You are not buying consulting hours, but you sit at the small end of exactly the same machine. Your accounting software’s AI assistant, the automation platform’s built-in model, the agency that “recommends” one writing tool: each of those is a channel that a model vendor has already paid to reach, in seats or credits or training rather than in a signed alliance.

Price the AI you are already buying before you buy any more

The bundled assistant feels free because it has no separate invoice. It is not free; it is priced into the tier you were moved onto, and the number is usually visible if you look at the plan grid rather than your card statement.

Microsoft’s business comparison page lists Microsoft 365 Business Basic at “$7.00 user/month, paid yearly”, Business Standard with Copilot at “$23.50 user/month, paid yearly”, and Business Premium with Copilot at “$32.00 user/month, paid yearly” [4]. That gap buys more than AI, since the higher tiers carry desktop apps and management features too, which is exactly why the AI portion is hard to see. Notion runs the same shape. Its pricing page shows Plus at $10 per member per month and Business at $20 on yearly billing, and Business is the tier where the Notion Agent, which “completes complex, multi-step tasks using context from Notion, your connected apps, and the web”, actually lives [6]. For contrast, buying the assistant directly and on its own is a published number as well. Claude Pro is listed at $17 per month with the annual discount, billed as $200 up front, or $20 billed monthly, and a Claude Team standard seat is $20 per seat per month billed annually or $25 monthly [8].

Put your own two numbers side by side: what a seat cost before the AI tier and what it costs now. Multiply the gap by seats and by twelve. That annual figure is the budget you are already spending on AI without having chosen it, and it is the right number to compare against a standalone subscription for the one or two people who would actually use one.

Buying the tool does not pin the model underneath it

People pick a bundled assistant believing they have chosen a model. They have not. They have chosen an interface and a data agreement, and the vendor keeps the right to change what runs behind it.

Notion states this in its own documentation: “Notion currently utilizes various large language models (LLMs) hosted by Notion as well as by organizations such as Anthropic and OpenAI” [7]. Microsoft published the same fact as a feature. In a post dated 24 September 2025 it wrote that “Copilot will continue to be powered by OpenAI’s latest models, and now our customers will have the flexibility to use Anthropic models too”, named the versions current at that date, Claude Sonnet 4 and Claude Opus 4.1, as options in Copilot Studio, and said an organization’s admin enables the access in the Microsoft 365 admin center [5]. That those particular version numbers now read as dated is the whole point.

Two consequences follow, and they last longer than any model name. Do not choose a tool because of the model it currently uses, because that is the part most likely to change without your involvement. Do choose on the things written into the contract: what happens to your text. Notion’s page is specific there, saying that by default Notion and its AI subprocessors do not use customer data to train models, that for all non-Enterprise workspaces LLM providers retain customer data for 30 days or fewer before deletion, and that zero data retention applies by default on Enterprise workspaces [7]. Those terms differ by plan tier, which means the answer you found when you signed up may not be the answer that applies to the tier you were later upgraded onto.

Ask every adviser what they are certified in, and treat vagueness as data

Anyone who configures your systems, sets up your automations or writes your AI policy has a stack they know. Ask which vendor they hold certifications or a partner relationship with, and whether that relationship involves resale margin or referral fees. Much of it is published anyway. IBM put its certification plans in a press release [1], and OpenAI names its partners and invites them to “Co-sell with OpenAI” on its own site [3]. If an adviser will not answer a question their vendor already answers in public, the reluctance is itself the answer.

What you are listening for is not a disqualifier. A partner-trained adviser may still be the cheapest competent person you can hire, and their familiarity is worth real money in setup time. You are listening for whether the alternatives were considered at all. An adviser who can tell you why the other option loses for your specific case has done the comparison. An adviser who cannot name what the other option is has given you the sales sheet.

The same question applies inside your own head. If you have three years of habit in one assistant, you are that consultant. Track how much of your preference is fit and how much is fluency, because the two feel identical from the inside and only one of them survives a change of vendor.

Run five real jobs before the renewal date, not a comparison table

Vendor comparison tables are written by the vendor. The test that settles it costs a fortnight and no money you were not already spending.

Pick five tasks you actually do, in your actual files: the weekly client update, the quote that needs numbers pulled from a spreadsheet, the tidy-up of a messy call transcript, the first draft of a proposal section, the search across your own notes for something you half remember. Run all five through the assistant that came in the bundle, and the same five through one standalone tool on a monthly plan. Keep the outputs. Two weeks later, count how many you shipped without a rewrite, and how many times you gave up and did it by hand.

Then set the decision against a date rather than a feeling. Note when each subscription renews, and put a reminder two weeks before it. Annual billing is where drift becomes expensive; the Claude annual plan, for instance, is billed as $200 up front rather than monthly, which is cheaper per month and a longer commitment [8]. Cancel the loser before the renewal, not after, and downgrade any tier you climbed only for an assistant nobody uses.

What still goes wrong

The bundled option often wins on the merits, and this guide can be read as cynicism about that. It is not. If the assistant in the suite you already pay for handles your five jobs, the correct decision is to use it and cancel everything else, and the distribution deal that put it in front of you has done you a favor. The point is only that you should reach that conclusion after the test rather than before it.

You also cannot see most of what you would need to see. Notion’s own description is that it uses “various” large language models, hosted by itself and by others [7], which is accurate and tells you nothing about which one answered your last request or whether that changed last week. A drop in quality you feel in March may have a cause you can never confirm. Terms move too. The retention and training commitments quoted here are the ones published today for the tiers named [5][7], and they are worth re-reading at renewal rather than trusting from memory.

The five-job test is noisy. Two weeks of one person’s work is not an evaluation, and a tool can lose your test and still be the better tool for a task you did not think to try. Treat it as a cheap tiebreaker between options that are close, not as evidence about models in general. And it does nothing about switching costs: the longer an assistant holds your notes, prompts and automations, the more the next decision is about migration effort rather than quality, which is precisely the position every distribution deal is designed to put you in.