Private AI systems

What is in scope

• Assessment of what your staff actually need AI for, and what data it would have to touch
• Model and hardware selection sized to the work, not to a benchmark
• Private knowledge ingestion with scoped retrieval, so teams only see what they should
• Access control mapped to your existing identity groups
• Audit logging and retention rules you define
• Staged rollout, with the first group trained properly before expanding
• Ongoing maintenance, model upgrades and knowledge updates

How the engagement runs

1. Week one — assessment. Interviews with the people who would use it, a look at your document estate, and a review of your network and identity setup. Output is a written recommendation including the option of not doing it.
2. Weeks two to three — build. Infrastructure stood up, models deployed, knowledge indexed, permissions wired.
3. Week four — pilot. One real team, real work, daily feedback. This is where the retrieval scoping gets fixed, because it is always slightly wrong at first.
4. Then — expand and maintain. Further groups added as the pilot settles. Maintenance is a standing arrangement or handed to your team.

What determines the price

Three variables: where it runs (owned hardware is capital spend with near-zero marginal cost; a private cloud tenant is the opposite), how much knowledge it holds, and how deep the integration goes into systems that write back.

A bounded version ships at a fixed price inside the AI Business Starter Pack, which is the cheapest way to test whether this belongs in your business.

When this is the wrong service

Stated plainly, because an engagement that should not have happened costs us more than it costs you.

• You want the strongest possible open-ended reasoning and your data has no confidentiality constraint. Use a frontier model; you will get better results for less.
• Nobody internally can own it. A private system needs someone to care about it even when we maintain it.
• The real problem is that your documents are a mess. Fix that first — AI retrieval over disorganised content produces confidently wrong answers faster than a human would.