AI on-premise vs cloud: start with what your contracts let you send

At a glance

• For confidential sponsor or client work, the first question is what your contracts allow you to send to a third party. Cost and scale come after.
• There are three options, not two: consumer cloud AI, business or enterprise cloud tiers, and on-premise AI, which for a firm without IT staff means a device in the office that a vendor runs.
• A business tier improves the terms (OpenAI and Microsoft both say business data is not used to train their models), but the AI company still receives and processes the material.
• Under the GDPR, a company that processes personal data on your behalf is a processor, and using one requires a contract with set terms and, for transfers outside the EU, a lawful transfer basis.
• Cloud wins on model quality. On-premise wins when nothing may leave the building, and on a managed device the upkeep is the vendor’s job.

On-premise AI runs the model on a device in your own office. Cloud AI runs it on the AI company’s servers, so every prompt and pasted document travels there. For a CRO, biotech or medtech consultancy handling sponsor material, the choice between the two is usually made by the contract before cost or scale come into it. If the contract bars disclosure to third parties, every cloud option is out for that material, however good the vendor’s terms. If it allows approved service providers, the fair comparison is a vendor-run device in your office against a cloud business tier with enterprise terms, and on model quality the cloud tier is often the better choice.

The question that decides it: what your contracts let you send to a third party

Before you compare products, read the confidentiality and subcontracting clauses in your largest sponsor and client agreements, including any quality agreement. You are looking for three things:

1. Whether confidential information may be disclosed to third parties at all, and on what conditions.
2. Whether subcontractors or service providers need the sponsor’s written approval.
3. Whether the data must stay in a particular country or region.

If the contracts allow named service providers with the sponsor’s consent, a cloud business tier can be a legitimate option once the sponsor has agreed. If they bar disclosure to third parties outright, every cloud option is ruled out for that sponsor’s protocols, reports and study files. You may find you have both kinds of client, and need both kinds of tool.

When the material includes personal data of people in the EU, such as patient details in a safety narrative, the GDPR adds its own structure. It defines a processor as a person or body that processes personal data on behalf of the controller. A cloud AI service that receives your staff’s prompts and documents to produce answers for you is doing exactly that. Using it requires a processor contract with set terms and your authorisation before it brings in other processors, and transfers of personal data to a country outside the EU are allowed only if the conditions in Chapter V are met. None of this forbids cloud AI. It does mean a business tier comes with a contract to review, a list of subprocessors to check and possibly a transfer to assess. Our guide to GDPR compliant AI walks through each of those obligations.

Three options, not two

Consumer cloud AI

The free and personal plans of the public chatbots. Staff sign up with their own email, and the firm has no contract with the AI company. OpenAI’s enterprise privacy page says it uses data from versions of ChatGPT for individuals in training its models. An individual user can switch this off in ChatGPT’s data controls (“Improve the model for everyone”), but that is a setting each employee controls on their own account, and it still leaves your firm with no contract. For sponsor material, this option is almost always the one the contract was written to prevent. It is also the one staff use when nothing else is offered.

Business and enterprise cloud tiers

Paid plans with a company contract, admin controls and better data terms. As of October 2026, OpenAI (ChatGPT Business, Enterprise and the API) and Microsoft (Microsoft Copilot, formerly Microsoft 365 Copilot) both say business data is not used to train their models by default, and both offer a data processing agreement. The detail for each plan, including OpenAI’s 30-day deletion terms and the Anthropic models Microsoft keeps outside its EU Data Boundary, is in private AI for business.

These are real improvements over consumer plans. They do not change the basic data path: the prompt and any pasted document travel to the vendor’s data centers and are processed there, and the vendor’s subprocessors may be involved.

On-premise AI, run by a vendor

A device in your office runs an open-weight model. Prompts and documents travel across your office network to that device and stop there, so no AI company receives them. A small firm does not build or run this device itself: a vendor delivers it, connects it to your documents and keeps it patched. That vendor goes on your list of third parties in place of the AI company, so ask what it can see. Our guide to on-premise AI covers who runs the device and what to ask before you sign.

Side by side

	Consumer cloud AI	Business or enterprise cloud tier	On-premise AI, run by a vendor
Where the prompt goes	AI company’s servers	AI company’s servers, under a company contract	A device in your office
Who processes the material	The AI company, under its consumer terms	The AI company and its subprocessors, under your contract	The device in your office, plus the vendor that runs it
Retention and training terms	Set by the vendor. OpenAI trains on individual plans unless each user opts out	Contract terms, business data not used for training by default at OpenAI and Microsoft	Set in your contract with the vendor
Model quality ceiling	The vendor’s strongest consumer models	The vendor’s strongest models	Open-weight models that fit one office device, a step below the largest cloud models on hard reasoning
Cost shape	Free or a small monthly fee per person	Monthly or yearly fee per person	A one-off charge for the device and a yearly fee for the service
Who maintains it	The AI company	The AI company	The vendor that runs the device

“Who processes the material” decides whether you can use an option at all for a given sponsor. “Model quality ceiling” decides how good the answers are once you can.

When the cloud is the right answer

• The material is public or your own. Summarizing a published FDA guidance, drafting a job advert or rewriting your own website copy involves nothing a sponsor contract protects.
• Your contracts allow it and the sponsor has agreed. Then a business tier with a signed data processing agreement gives you the strongest models with nothing to maintain.
• You need the strongest model available. The largest cloud models are ahead of anything that fits on one office device for long, hard reasoning tasks and broad general knowledge.
• Use is light and occasional. A few seats on a business tier cost less than a device sized for the whole firm.

When on-premise is the right answer

• Your contracts bar third parties. Then the cloud options are out for that material, and the choice is between on-premise AI and no AI.
• Sponsors ask how you use AI and you need a short, checkable answer. “It runs on a device in our office and nothing goes to an AI company” is easier to defend in a vendor qualification questionnaire than a list of subprocessors and retention periods.
• Most of the work is reading and drafting from your own documents. Summarizing monitoring reports, finding the relevant SOP, comparing protocol versions and drafting from templates depend more on what the model can see than on how large it is.

On-premise is the weaker choice for raw model capability. On a managed device the upkeep is the vendor’s job, so for your firm it is a contract term rather than work. The capability gap is a real limit to accept.

What each costs a year

A business tier is priced per person. For a 25-person firm the yearly bill is the seat price times 25 times twelve, and it rises with every hire. Check the vendor’s current price list, because these prices change often. A managed device is a one-off charge plus one yearly fee for the whole firm, so the cost per person falls as the firm grows. Compare the totals, then what each buys: the seat price buys the strongest models, and the device buys keeping sponsor material in the building.

What a firm without IT staff should do

1. Sort your work into material that may go to a cloud vendor and material that may not, based on the contracts.
2. For the first group, pick one business tier, sign its data processing terms, check its subprocessor list and turn off anything you do not need. Ban consumer accounts for work.
3. For the second group, decide whether you need AI on it. If you do, it has to run inside your building.
4. Have a vendor deliver and run the device, and get in writing what the vendor can see and how fast it applies security patches.
5. Write down what you did. A sponsor audit or questionnaire will ask.

Six questions to put to any AI vendor, cloud or on-premise, are in private AI for business. If you also hold US patient data, see is there a HIPAA compliant AI.

Where IntraMuros fits

IntraMuros is the on-premise column for firms that have no one to run a server. Mayet, a British company that builds software for pharma, biotech, medical devices, clinical research and other regulated industries, ships a device to your office in a couple of days, configures it remotely in a few hours after you plug it in, and connects it to your documents and email, and can connect it to software such as an eQMS, eTMF, CTMS or ERP, on your premises or in the cloud. It then patches, monitors, backs up and, if it fails, replaces the device. It is running within a week of signing.

The model runs on the device, so prompts, documents and answers do not go to an AI company. For the training row, your documents and conversations are never used to train any AI model, by Mayet or by anyone else. Staff open the agent, MURO, in a browser and sign in with their usual Microsoft or Google work account, and MURO answers only from what that person can already open, in shared folders and in each connected piece of software. Every answer shows the document it came from. For the “who processes the material” row, the answer is the device in your office plus Mayet as its administrator: Mayet connects to the device to update, monitor and back it up. As the one outside party that administers it, Mayet goes on your vendor list like any other.

Everyone in the firm can use it, and maintenance is £1,000 a month or £10,000 a year for one device. The device is a one-off charge chosen by company size, with set-up included: Base, £4,500, or Pro, £7,500, for a firm of about ten, and Enterprise, £12,000, for a firm of 25. For 25 people, maintenance at £10,000 a year is £400 a person a year, about £33 a month. When it is busy, questions wait their turn. See how it works and pricing.

Some firms will use both: a cloud business tier for public and permitted material, and IntraMuros for the sponsor work that may not leave the building. If none of your contracts restrict third parties and you want the strongest models, a cloud business tier alone is the simpler choice.

Sources

• Regulation (EU) 2016/679, General Data Protection Regulation — Articles 4(8), 28 and 44