Private & Secure AI
Can a company use AI without sharing confidential information?
The short answer
Yes. A business can use AI on confidential information without that information training a public model or leaving its control, provided it uses enterprise or API access under a zero-data-retention agreement rather than consumer chat tools, keeps its documents in its own permission-controlled store, and sends only the minimum relevant extract to the model at query time. Setting this up properly typically costs £6,000 to £30,000 depending on how many systems it touches.
Indicative price ranges
Policy and safe tooling
£2,000 – £6,000
Enterprise licences configured correctly, a written AI use policy, staff training and DLP rules. Stops the immediate leakage risk.
Private retrieval system
£8,000 – £25,000
Your documents in your store, permission-aware retrieval, zero-retention model access, citations and full audit logging.
Fully self-hosted
£25,000 – £60,000+
Open-weight models running on UK infrastructure you control. Full data residency within your boundary. Higher cost, lower answer quality than frontier models — choose it only when contracts demand it.
These are indicative UK ranges based on projects we have actually delivered. Scope drives the number, not the sales conversation — we quote a fixed price after a scoping session, and we will tell you if an off-the-shelf tool is the cheaper answer.
Typical timescales
- 1–2 weeks
Risk and data classification
What is genuinely confidential, what is merely internal, and what may never be sent anywhere.
- 1–2 weeks
Policy and tooling
The fastest risk reduction available — usually done before any build starts.
- 4–10 weeks
Private system build
Retrieval, permissions, guardrails, logging.
- 1–2 weeks
Assurance
Penetration testing where required, plus documentation for your clients' due-diligence questionnaires.
Who this suits
- Professional services handling client-confidential material under contractual restrictions.
- Healthcare, legal, financial and public-sector suppliers with data residency obligations.
- Manufacturers and product businesses with commercially sensitive designs or pricing.
- Any business where staff are already pasting sensitive content into free AI tools — which is most of them.
Who it does not suit
- Organisations that want a policy document and nothing else. Policy without tooling does not change behaviour.
- Businesses with genuinely no confidential data — you can use standard tools and save the money.
- Anyone who needs absolute certainty that no third party ever processes a byte, unless they are prepared to fund full self-hosting.
What actually drives the cost
Level of assurance required
Contractual data-residency clauses and client audits push you towards self-hosting, which is materially more expensive.
Number of systems in scope
Each connected source needs its own permission mapping and sync.
Existing identity setup
A tidy Microsoft 365 tenant with proper groups makes permission-aware AI far cheaper to build.
Audit and retention requirements
Long retention with searchable logs adds storage and engineering.
Where the risk actually is
The risk is almost never the technology. It is an employee pasting a client contract into a free consumer chat tool at 4pm on a Friday. Consumer tiers may use submitted content to improve models; enterprise and API tiers, under the right agreement, do not. Most businesses are exposed today not because they adopted AI, but because they did not.
So the first move costs very little: give people a sanctioned tool, write a one-page policy that says what may and may not be entered, and block the consumer endpoints. That alone removes the majority of practical exposure within a fortnight.
The three safe architectures
1. Enterprise tooling with zero retention. You use a commercial model through an enterprise agreement or API where the provider contractually does not retain or train on your inputs. Cheapest, fastest, appropriate for most UK SMEs.
2. Private retrieval (the usual answer). Your documents stay in your own store. When someone asks a question, the system finds the handful of relevant passages they are permitted to read and sends only those to the model, which writes an answer with citations. Your corpus remains under your governance; only a minimal, permission-filtered extract is ever transmitted, and it is not retained.
3. Fully self-hosted open-weight models. Nothing leaves your infrastructure at all. Necessary when a contract or regulator demands it. The trade-off is real: higher cost, more maintenance and answers that are good but not frontier-quality.
Non-negotiables in any build
- Permission-aware retrieval. If a user cannot open the document, the AI must not quote it. This is the requirement most cheap implementations skip.
- Citations. Every answer shows its sources so a human can verify it.
- Refusal behaviour. No source, no answer. Confident invention is worse than silence.
- Full audit logging. Who asked what, what was returned, and from where.
- Retention control. You decide how long prompts and logs live.
Proving it to your clients
Increasingly, the reason to do this properly is commercial rather than legal: your clients' procurement teams are now asking how you handle AI. Being able to hand over an architecture diagram, a zero-retention agreement and an audit-log policy is becoming a differentiator in tenders. We build that documentation as part of every private AI engagement.
Common questions
Does ChatGPT train on our data?
Consumer tiers may use conversations to improve models unless you opt out. Enterprise, Team and API access under the standard business terms do not train on your inputs. The distinction matters enormously — check which tier your staff are actually using.
Is private retrieval genuinely safe for client-confidential material?
For most UK businesses, yes, provided it runs under a zero-retention agreement with permission-aware retrieval and audit logging. Where a client contract explicitly forbids any third-party processing, self-hosting is the correct answer.
What is the fastest way to reduce risk this month?
Provide a sanctioned enterprise tool, publish a one-page acceptable-use policy, and block consumer AI endpoints on the corporate network. That is typically a two-week piece of work.
Can we restrict which staff see which answers?
Yes. Permission-aware retrieval maps your existing Microsoft 365 or Google Workspace groups, so answers respect the access rights each person already has.