Private & Secure AI
How much does a secure private business AI system cost?
The short answer
A secure private AI system for a UK business typically costs £6,000 to £35,000 to build, depending on how many data sources it connects to and whether it faces customers as well as staff. Our AI concierge deployments start at £2,950 for setup on a defined scope, while a full internal knowledge system across SharePoint, Outlook and a document store usually falls between £12,000 and £30,000. Running costs are normally £150 to £900 a month covering hosting, model usage and support.
Indicative price ranges
Focused deployment
£2,950 – £8,000
One clear job done properly: a customer-facing concierge, a document Q&A assistant, or an internal helpdesk over a defined knowledge base.
Internal knowledge system
£12,000 – £30,000
Connected to SharePoint, Outlook, a document store and your CRM. Permission-aware answers, source citations, admin controls and audit logging.
Enterprise / multi-department
£30,000 – £70,000+
Multiple assistants, voice, CRM write-back, workflow triggers, SSO, retention policies and a full governance layer.
These are indicative UK ranges from systems we have built and run. The £2,950 figure is our published AI concierge setup fee for a defined scope. Model and hosting usage is billed at cost or included in a support retainer — we will show you the running-cost model before you commit.
Typical timescales
- 1 week
Data and access audit
What the AI may see, who may ask it, and where the boundaries sit.
- 1–3 weeks
Ingestion and indexing
Connecting sources, cleaning content, building the permission-aware index.
- 2–6 weeks
Build and guardrails
Interface, retrieval, refusal behaviour, escalation to a human, logging.
- 2–4 weeks
Pilot and tuning
A small group uses it in anger and we tune on real questions, not imagined ones.
Who this suits
- Businesses whose staff waste hours hunting for information across SharePoint, inboxes and shared drives.
- Organisations that cannot put client or commercial data into a public AI tool.
- High-volume enquiry operations — bookings, memberships, events, support — where most questions are repeats.
- Regulated or contract-bound sectors that need audit logs and data residency guarantees.
Who it does not suit
- Businesses with no documented knowledge to index. AI cannot retrieve what was never written down.
- Teams under about ten people with low enquiry volume — the payback rarely justifies the build.
- Anyone expecting it to replace headcount outright. It removes repetitive work; it does not remove judgement.
- Organisations unwilling to assign an owner for content accuracy.
What actually drives the cost
Number of data sources
One document library is straightforward. SharePoint plus Outlook plus a CRM plus a legacy file server is a project.
Permission awareness
Making sure a user only gets answers from documents they are allowed to read is the single most technically demanding requirement — and the most important.
Voice
Voice interaction adds telephony, latency engineering and a different testing regime.
Write access
An assistant that reads is one risk profile. One that books, updates or emails on your behalf needs far more guardrail work.
Hosting choice
UK or EU-hosted infrastructure with a zero-retention model agreement costs more than a default consumer API and is usually worth it.
What you are actually paying for
Very little of the cost is the model itself. The money goes on the boring, essential parts: getting your content into a state an AI can retrieve accurately, enforcing who can see what, making the system say "I do not know" instead of inventing an answer, and logging every interaction so you can prove what happened.
That is the difference between a private AI system and a chatbot. A chatbot guesses. A private system retrieves from your documents, cites the source, refuses when it has no basis, and hands off to a human when the question needs one.
Running costs, honestly
- Hosting and infrastructure: £60–£400 per month depending on volume and residency requirements.
- Model usage: typically £50–£500 per month for an internal system; consumption-based, so it scales with use.
- Support and content upkeep: from £250 per month. Someone has to keep the source material current — that is the ongoing job.
What the return looks like
On our AI concierge deployment for York Racecourse, the assistant resolves 92.9% of enquiries without a human, saving around 330 staff hours a month across peak race weeks. That is the shape of the business case: not headcount removed, but a repetitive load lifted off a team that was drowning in it during peaks.
For internal knowledge systems the equivalent measure is search time. If twenty people each lose 25 minutes a day looking for information, that is over 2,000 hours a year. Recovering even half of it changes the maths quickly.
Before you commission anything
Ask any supplier these five questions and insist on written answers:
- Where is the data hosted, and under which jurisdiction?
- Is our content used to train any model? (The answer must be no.)
- How does the system enforce document-level permissions?
- What does it do when it does not know?
- What is logged, for how long, and who can read the logs?
If a supplier is vague on any of these, that is your answer. More detail on the technical side in our guide to using AI without sharing confidential information and UK-hosted AI and GDPR.
Common questions
Is £2,950 really the starting price?
Yes — that is our published setup fee for an AI concierge on a defined scope, plus a monthly running cost. Broader internal systems that connect to multiple data sources start considerably higher, typically £12,000 upwards.
Do we need our own model?
Almost never. Training or fine-tuning your own model is expensive and rarely necessary. Retrieval over your own documents using a commercial model under a zero-retention agreement gives better accuracy for a fraction of the cost.
How long until it is live?
A focused deployment can be live in four to six weeks. A full internal knowledge system across several data sources typically takes eight to twelve.
What happens when our documents change?
The index re-syncs automatically on a schedule, so answers follow your live content. Someone still needs to own document accuracy — the AI faithfully repeats whatever is in the source.