Probative Co / Services / Private AI deployment

the AI practice · your perimeter, your precedents, your evidence

Private AI deployment

Design, pilot and deploy legal AI that runs inside your own perimeter: models you control, retrieval over your own documents with the permissions you already enforce, and an evaluation record built from your matters rather than a vendor's demo.

Start with the free 3-file check → Start the onboarding form See the options

$3,000 – $4,000

fixed · 10 working days

AI readiness assessment

~10 working days

Estimated range $3,000 – $4,000 · how quotes work

$12,000 – $18,000

from · 4–6 weeks

Private AI pilot

4–6 weeks

Estimated range $12,000 – $18,000 · how quotes work

$23,000 – $45,000

from · 6–10 weeks

Private AI deployment

6–10 weeks

Book the free check →

No charge, no obligation · findings yours either way

$1,900 – $3,500 / month

per month

Managed private model

Continuous

Estimated range $1,900 – $3,500 / month · how quotes work

Costs and quotes — fixed fees, agreed in writing →
Fees exclude VAT where it applies. Encrypted transfer, written DPA, conflict check before we start. Not a law firm; not legal advice.

the thinking

Why architecture beats assurance

A vendor's data-processing agreement limits what a supplier may do. It does not change the fact that client content leaves your systems. For firms with confidential, high-profile or regulated clients, that transfer is the problem — and the only complete answer is a model that runs where your documents already live.

Open-weight models are now genuinely useful for focused legal tasks with retrieval
Inheriting DMS permissions solves confidentiality structurally, not contractually
An evaluation record built on your own matters is defensible; a vendor benchmark is not

what you receive

Stated in advance, delivered as described.

Readiness and feasibility

What is in use today, what a private layer would replace, which hardware or private tenancy fits your size, and what it costs to run per month — in writing, before you commit to anything.

Architecture and build

Hosting, networking, key management, logging and matter walls, designed for your environment and documented for your IT lead and your insurer.

Model selection and licensing

Which open-weight family, which size, which quantisation — and a licence review so you know what you may do with it commercially, indefinitely.

DMS-scoped retrieval

Retrieval over precedents and matters that respects ethical walls and permissions rather than bypassing them. This is where most private AI projects succeed or fail.

Evaluation harness

Fifty to 150 tasks drawn from your own matters, scored by your fee-earners, with the results and failure modes documented.

Governance pack

DPIA, register entry, policy updates, vendor diligence notes and an incident runbook.

Training and handover

Fee-earner sessions, administrator runbooks and a named internal owner, so the capability outlives the project.

how it runs

Four steps, no surprises.

STEP 1

Assessment

Ten working days to establish what is in use, what is feasible and what it costs. Fixed credited against a pilot if you proceed.

STEP 2

Pilot

Four to six weeks on a scoped document set: model chosen, retrieval wired, evaluation run with your fee-earners, security and DPIA documented.

STEP 3

Deployment

Six to ten weeks to production: full retrieval scope, access controls, monitoring, training and the governance pack.

STEP 4

Managed

Optional retainer a month: model updates with regression evaluation, quarterly re-testing, register upkeep and incident support.

options and fees

Fixed numbers. The invoice matches the proposal.

Engagement	Cost	Turnaround	What it covers
AI readiness assessment	$3,000 – $4,000fixed · 10 working days	~10 working days	The honest starting point: what AI is actually in use, what data it touches, what it would take to run any of it privately, and what to fix first. Shadow-AI discovery across fee-earner workflows AI use register with data-flow notes per tool Private-deployment feasibility read: models, hardware, DMS retrieval, cost bands Prioritised 90-day plan with owners and effort estimates Board-ready one-pager	Contact us → Estimated range $3,000 – $4,000 · how quotes work
Private AI pilot	$12,000 – $18,000from · 4–6 weeks	4–6 weeks	A controlled pilot of a private model on your own matters, so the decision to deploy is made on evidence from your documents rather than a vendor demonstration. Model selection and hosting design (in-firm, private cloud or dedicated tenancy) Retrieval set up over a scoped document subset with permissions inherited from the DMS Evaluation harness: 50–150 tasks from your own matters, scored by your fee-earners Security, DPIA and AI-policy documentation for the pilot Pilot report with go / no-go recommendation and full cost model	Contact us → Estimated range $12,000 – $18,000 · how quotes work
Private AI deployment	$23,000 – $45,000from · 6–10 weeks	6–10 weeks	Design, build and hand over a private AI capability your firm owns: models inside your perimeter, retrieval scoped to your own precedents, and an evaluation record that satisfies a regulator or an insurer. Architecture: hosting, networking, key management, logging Model choice and licence review, including open-weight options DMS-scoped retrieval with permission inheritance and matter walls Evaluation and red-teaming against legal tasks you specify Fee-earner training, runbooks and an internal owner programme Governance pack: DPIA, AI register entry, policy updates, vendor diligence	Book the free check → No charge, no obligation · findings yours either way
Managed private model	$1,900 – $3,500 / monthper month	Continuous	We keep the deployment healthy and the governance current: model updates, evaluation re-runs, prompt and policy revisions, and a quarterly board note. Model and dependency updates with regression evaluation Quarterly re-evaluation against your task set Register and policy upkeep as AI use changes Incident support and vendor escalation Quarterly governance note your board can minute	Contact us → Estimated range $1,900 – $3,500 / month · how quotes work

Payment. 50% on booking, 50% on delivery unless the option says otherwise. Card payment through Stripe, or invoice with bank transfer. Card details never touch our servers. Larger deployments are billed to milestones agreed in writing before work starts.

who this is for

Firms with confidential or high-profile client bases where public tools are a hard no
Firms whose AI use is currently blocked and needs to become possible
Firms paying per-seat licensing for capability they could own
Firms asked by a client or insurer where their data actually goes

questions we get asked

QWhat does 'private' actually mean?+

Inference happens on infrastructure you control — in-firm hardware, a private cloud tenancy, or a dedicated managed instance with no shared capacity — with retrieval over documents held in your own store. Client content does not reach a third party's model for processing.

QIs a self-hosted model as good as the frontier models?+

For focused legal tasks with retrieval and a well-built prompt, open-weight models are close enough to be genuinely useful. For open-ended reasoning they are behind. A pilot on your own matters is the only honest way to judge it — which is exactly what we run.

QWhat does it cost to run?+

A capable single-node setup for a small firm sits in the low thousands of pounds of hardware, or a few hundred a month in a private tenancy. Heavy usage across a larger firm needs more. You get a written cost model with the pilot, not a range in a proposal.

QDo we have to abandon the tools we already use?+

Usually not. Most firms run a private layer for confidential work and governed commercial tools for the rest, with a policy that says which is which. We write that policy as part of the deployment.

QWill this pass our insurer's questions?+

The governance pack is written for that audience: where the data rests, who can access it, what is logged, how incidents are handled, and who at the firm is accountable. Several clients have handed ours straight to their broker.

QDo you resell hardware or models?+

No. We take no vendor payments or reseller margin, which means our recommendation is free to be that you need less than you were quoted.