There are six of them, from a chatbot in a browser tab to a model running on your own hardware. This is what each one is for, what it asks of you, and where the larger gain sits once the first one is in place.
Most companies already have AI in the building. In PagerDuty’s survey of office professionals in June 2026, 66% said they used AI tools at work even though they believed their employer’s policy did not allow it1, and other surveys through 2025 and 2026 put the figure as high as 80%. People went looking on their own because the tools help with the paragraph that will not start, the formula nobody remembers, the first pass at a translation, the forty-page document due tomorrow.
That is demand most organisations spend a year trying to manufacture, and it arrived by itself. It is a good thing to build on. Individual use is genuinely fine for a great deal of work, and for some people it is all they will ever need from AI.
Two practical facts shape what comes next. The first is that consumer tiers train on what goes into them by default. ChatGPT and Gemini do unless the setting is changed2, while Claude’s consumer plans have been opt-in since August 20253. Which tier you are on therefore decides which material suits the tool, and enterprise terms turn training off by contract. The second is that individual use tops out at individual productivity. Forty people each saving an hour a week is forty hours a week, without changing the process any of them work in.
Six shapes cover almost everything a company can actually buy. Each is the right answer for some organisation, and most firms end up running two or three at once. The product names below are passing examples and will change; the categories move more slowly. Hapax resells none of them and takes no commission from any vendor.
These cost nothing, or around €20 a month on someone’s own card. ChatGPT, Claude, Gemini and Copilot sit in a browser tab, bought by the person who uses them. They are good at drafting, summarising, explaining and unblocking, and the person reading the answer is the quality control, which is exactly the right arrangement for a first draft. What they do not come with is anything shared. There is no administrator, no common instructions, and no record of what the organisation is doing with them. As a company’s only shape they stop being enough the moment the same task has to be done the same way by more than one person.
The same tools, bought by the company, are sold as ChatGPT Business or Enterprise, Claude for Work and Gemini for Workspace. Content is not used for training4, an administrator can see and control who has access, and work becomes shareable as projects and standing instructions instead of living in one person’s tab. These tiers typically cost $20 to $25 per user per month. For most firms this is the shape to buy first. It disappoints in the ordinary way a licence disappoints, when everyone is given one and nobody is shown what to use it for.
Here the AI sits inside software the company already runs, in Microsoft 365 Copilot, Gemini in Workspace, and the assistants now shipped with Salesforce, ServiceNow and SAP. The appeal is context. The assistant can already see the documents, mail and tickets that the person is entitled to see, so its answers are grounded in your business. This shape is right when the work genuinely lives in that estate. It is often bought because procurement was easy, one line added to an agreement that already existed, and it then sits next to the line-of-business system where the hours actually are.
Microsoft states that prompts, responses and Microsoft Graph data are not used to train foundation models and are processed under its data-processing terms, with the EU Data Boundary covering the EU and EFTA since 1 May 2026; its Flex Routing, on by default for eligible tenants since April 2026, can send inference outside the EU at peak demand, while storage stays inside5.
You pay per use, by the million tokens, and build the thing around the model yourself. Here a model stops being a destination and becomes a component, which is what makes measurement and checking possible, and it is the shape behind the case studies on this site. An API build is the wrong answer when one person needs prose once a week, which is engineering effort spent rebuilding a chat window.
The terms are business terms. Anthropic’s commercial terms, as published in September 2026, state that the API does not train on inputs or outputs and holds inputs for seven days by default under the policy it set in September 2025. OpenAI’s API has not trained on business data without opt-in since March 2023, offers a data-processing agreement, and makes EU data residency available on approval, not by default6.
An agent takes a task, reads files, edits them, runs commands and comes back with a result for a person to review. It does the work itself. Anthropic’s Claude Code, OpenAI’s Codex, GitHub Copilot’s cloud agent and Cursor all read a codebase, plan a change, edit the files, run tests and hand the result back. Cowork, Anthropic’s desktop agent, does the same for people who are not developers, reading files on a computer, browsing the web and carrying out tasks7.
The practical difference from the shapes above is that an agent acts on your files and systems. It can create, overwrite, delete and send things, which makes it useful for multi-step work but dangerous when the access and review are not thought through.
Coding agents are sold as seats, typically $19 to $40 per user per month depending on vendor and plan, with usage beyond the included allowance billed by the token8. Under commercial plans, Anthropic states that it does not train on code or prompts sent through Claude Code; GitHub states that neither Copilot Business nor Enterprise data trains its models; Cursor’s privacy mode, available on every tier including the free one, prevents training by Cursor or its model providers9.
An open-weight model is one whose trained parameters are published for anyone to download. You run it on hardware you control, in your own data centre or cloud tenancy, air-gapped if that is what the work requires, with no per-token fee. The field turns over fast; the names below are current as of October 2026 and some will be gone by the time you read this. The main open families are Qwen (Alibaba), Gemma (Google), Mistral, OpenAI’s gpt-oss, DeepSeek and Meta’s Muse Glimmer. Most of these are released under Apache 2.0 or MIT licences, meaning free for commercial use with no revenue threshold. Qwen’s largest model carries a separate licence for businesses offering it as a service above $50 million in revenue10.
What fits on a given machine depends on memory. A laptop with 16 GB of RAM can run a model in the 20-billion-parameter range; OpenAI’s gpt-oss-20b downloads at about 14 GB. A workstation with a 32 GB graphics card runs the 27 to 35-billion-parameter models, among them Qwen 27B and 35B and Gemma 4 31B, which download at 18 to 24 GB. A single 80 GB server GPU handles models up to about 120 billion parameters. NVIDIA’s DGX Spark, a desktop unit with 128 GB of unified memory, is rated for up to 200 billion11. Beyond that, the largest open models need multiple GPUs or a cluster.
The best open-weight models trail the frontier by four to six months on Epoch AI’s capability index12. Those are trillion-parameter models running in data centres. Artificial Analysis found the leading open-weight models priced at half to one-sixth of the leading closed models, with a wide gap on the hardest reasoning and agentic coding evaluations13. For the workstation-class models that a company would actually self-host, independent evaluation is sparser but not absent. EuroEval, an independent benchmark suite, publishes a Finnish leaderboard covering both open and closed models. Gemma 4 31B and Qwen 27B share the fifth tier, with eight models above14.
Sovereignty is usually the reason a local model gets raised. Mistral is domiciled in Paris and hosts in the EU by default, with on-premises and private-cloud options. Anthropic has no native EU data zone, although EU-region routes exist through AWS Bedrock and Google Vertex. Under the GDPR the processor terms carry more weight than the map pin, and EU residency strengthens a position without replacing a data-processing agreement. Most companies that ask us about a local model want an enterprise contract and a data-handling policy instead.
One duty applies whichever shape is chosen. Article 4 of the EU AI Act requires every organisation whose staff use AI at work to take measures to support the development of AI literacy among their staff, taking into account their technical knowledge, experience, and the context the systems are used in. Following the Digital Omnibus of July 2026, the amended text expressly clarifies that no specific level of literacy must be guaranteed. No certificate and no curriculum are prescribed, though the Commission’s guidance notes that in many cases simply relying on the instructions for use may not be effective. The duty has applied since February 2025, with national enforcement powers from August 2026, and it draws no distinction between a browser tab and a model on your own hardware15. The training page sets out what it means in practice.
Four questions separate the six in practice.
| Who buys it | How it is paid for | What you control | What it takes to run |
|---|---|---|---|
| iConsumer tools | |||
| The person who uses them | Nothing, or around €20 a month on someone’s own card | No administrator, no common instructions, and no record of what the organisation is doing with them | A browser tab |
| iiEnterprise subscriptions | |||
| The company | $20 to $25 per user per month | An administrator can see and control who has access | Someone to manage it and show people what to do with it |
| iiiEmbedded copilots | |||
| The company, through a vendor it already has | One line added to an agreement that already existed | Access follows the permissions already in the software | Already running the platform it sits inside |
| ivAPI buildsThe shape behind the case studies | |||
| The company | Per use, by the million tokens | The whole workflow, and what goes in and comes out | You build the thing around the model yourself |
| vAgents | |||
| The person or the company, sold as seats | Seats, typically $19 to $40 per user per month, with usage beyond the included allowance billed by the token | What the agent can see, change and send | Access rules and someone reviewing the output |
| viSelf-hosted open-weight | |||
| The company | Hardware and staff, no per-token fee | Everything, from the hardware to the data, air-gapped if needed | Hardware with enough memory, and people to maintain it |
Which one fits depends on where your hours go, what your material is, who will maintain the thing afterwards and what your policies already say.
A person asking a chatbot gets a better paragraph; a workflow with AI inside it gets a better process. That is the whole argument of this page, and it is where the compounding happens, because a process improved once stays improved for everyone who uses it afterwards.
Integrated means three things in practice. The model runs inside the step where the hours go, and not in a separate window someone has to remember to open. Its output is checked by design, by a rule that has to pass, a second pass over the same material, or a named person who signs. And the result is measured against what the step used to cost in time, in rework, or in errors that reached a customer.
None of that requires the most expensive shape. An enterprise subscription with a few well-made standing instructions is an integrated workflow if the work runs through it, and an ordinary API key covers most of the rest. Choosing the shape is the smaller decision. Choosing which process to put it inside is the larger one.
The case studies on this site are all this kind of work, run end to end on data built with the answer known in advance, so every result could be marked against that answer. The claim this section makes has since been measured directly, in the tender case, where the same sixteen documents were answered both ways.
Sammalkoski Oy does not exist. It is a 340-person engineering-services firm in Turku where about seventy people had each found their own way to a browser chatbot and were getting real use out of it. The firm bought an enterprise workspace for everyone, with terms that keep its material out of training and an administrator who can see what is licensed. It then picked the two places the hours actually went, tender responses and site-report write-ups, and rebuilt both as workflows with the model inside the step and a check on the output before it moved on. Training ran on the firm’s own tenders.
Six months later a tender response takes 9 hours of a senior engineer’s time against 26 before, and 210 of 340 staff use the tools weekly, up from the seventy who had found them alone. The win rate on tenders sits where it was, at 31%.
Exemplum · hypothetical
We have since measured the step this exemplum describes, from private habits to one workflow, on sixteen engineered tenders answered both ways in the tender case.
If you are working out which shape fits, tell us which tools your teams already use, officially or not, and where their hours go. That is most of what anyone needs in order to give you a useful answer.
Sources
Reviewed · October 2026