Skip to main contentSkip to footer
Talk to us

APPLIED AI / LOCAL AI

Local AI, on infrastructure you control

We run open-weight models and AI agents on your own servers or private cloud, so sensitive data never goes to a third-party AI service.

Talk to us

Local AI means the model, the data and every prompt stay on infrastructure you own or control: your data center, your private cloud account or a single server in your office.

Sensitive work stays in. The rest can go out.

Most firms do not need every request to run locally. A routing gateway keeps privileged and regulated work on your servers and lets the rest use the strongest model available, under rules you set.

  1. Step 1

    Classify

    The gateway checks each request against rules your compliance team sets: the matter, the client, or the kind of data attached.

  2. Step 2

    Route

    Privileged and regulated work goes to the model on your own servers. Public or low-risk work can use a frontier model in your own cloud account.

  3. Step 3

    Log

    Every request is recorded with who asked, the data class and the model that answered, on your systems.

Where local AI pays for itself

Each of these involves data that clients, regulators or contracts keep in-house, and work that eats senior hours today. The model drafts and checks; a person decides.

  • Law firms

    Privileged document review

    “We have tens of thousands of documents for an investigation, and the client will not let them go to an outside AI service.”

    What runs locally
    First-pass relevance ranking against the issues you define, possible-privilege flags with the passage and the reason, and short summaries by custodian.
    Why it stays local
    Privileged and confidential client material, often under outside-counsel guidelines that restrict third-party AI tools.
    Who decides
    Lawyers make every relevance and privilege call. The model proposes and explains; it never decides.

    Reviewers start from a ranked, pre-tagged set instead of a raw pile, and quality checks go where model and reviewer disagree.

  • Law firms

    Contract review against your playbook

    “Every NDA and services agreement is checked by hand against positions we agreed years ago.”

    What runs locally
    Clause extraction, comparison with your playbook positions, and suggested redlines that cite the playbook rule behind each change.
    Why it stays local
    Client contracts and your fallback positions are confidential, and the playbook itself is know-how you do not want leaving the firm.
    Who decides
    The lawyer accepts, edits or rejects each suggestion. Nothing goes to the other side without them.

    Routine agreements turned around faster and more consistently, with senior time kept for the clauses that genuinely differ.

    AI for law firms

  • Banks and lenders

    AML alert triage and case write-ups

    “Most alerts turn out to be nothing, but each one still needs an investigator to pull data from five screens and write it up.”

    What runs locally
    An agent gathers the transactions, KYC file and earlier alerts into one case summary, drafts the disposition narrative with every figure linked to its source, and flags what is missing.
    Why it stays local
    Customer and transaction data under banking secrecy and residency rules, and suspicious-activity work that is legally confidential in most jurisdictions.
    Who decides
    The investigator closes or escalates, and decides whether anything is reported. The agent never files.

    Investigators spend their time on judgment instead of copying data between systems, and every write-up follows the same structure.

    AI for financial services

  • Banks and lenders

    Credit memo drafting

    “Our analysts spend days turning financial statements and borrower files into a credit memo in our house format.”

    What runs locally
    Figures read from the statements, ratios calculated by code rather than by the model, and memo sections drafted in your template with each statement cited and policy exceptions highlighted.
    Why it stays local
    Non-public borrower financials and your credit policy. Lenders rarely want either in a third-party AI service.
    Who decides
    The analyst owns the memo and the credit committee owns the decision. The draft is a starting point.

    A consistent first draft early in the process, so analysts spend their time on the risk rather than the formatting.

  • SaaS and engineering

    AI features your enterprise customers can approve

    “Our biggest prospects want the AI features, but their security review rules out sending their data to a third-party AI API.”

    What runs locally
    The same AI features, served by an open-weight model inside the customer’s own cloud account or your single-tenant deployment, behind the API your product already calls.
    Why it stays local
    The customer’s data never leaves their environment, which is often the only answer their security team will accept.
    Who decides
    The customer controls where it runs, what it can see and how long anything is kept.

    Deals that were stuck in security review can move, without building and maintaining a second product.

    AI for SaaS companies

  • SaaS and engineering

    Coding assistant on code that cannot leave

    “Our code is our IP, some of it belongs to clients, and our contracts say it stays in-house.”

    What runs locally
    Code completion, review and agent-style coding against a model on your own servers, connected to your repositories, with usage logged.
    Why it stays local
    Source code, client code under contract, and the secrets that turn up in history all stay on your network.
    Who decides
    Engineers review and merge. AI-written code goes through the same review and tests as any other change.

    Teams get AI-assisted engineering without an exception to their client contracts or security policy.

    AI-assisted engineering

One privileged document review, run locally

An illustrative internal investigation at a law firm, with the review run on the firm’s own server.

StepHow it usually goes todayWith local AI
Collect

TodayDocuments loaded into the review platform, with its AI features switched off because the client has not approved an outside AI service.

With local AIDocuments processed on the firm’s own server. The model has no route to the internet, so the client’s terms are met by design.

First pass

TodayJunior reviewers read every document, usually in date order.

With local AIThe model ranks every document against the issues the team defines and tags each with its reasons, so reviewers start with the most relevant.

Privilege

TodayReviewers look for lawyers’ names and legal advice as they read.

With local AIPossible privilege flagged with the passage and the reason, for a lawyer to confirm or reject.

Quality check

TodayA senior associate re-reviews a random sample.

With local AIChecks start with the documents where the model and the reviewer disagreed.

Record

TodayDecisions sit in the platform; how each call was reached is in reviewers’ heads.

With local AIEvery model suggestion and every lawyer decision logged on the firm’s systems, ready if the review is ever challenged.

Before any lawyer relies on it, accuracy is measured on a sample of documents your own team has already reviewed. Names and details in this example are illustrative.

What we build

Is local AI right for you?

Yes: the data cannot go to a third party

Privileged legal files, customer data your contracts keep in-house, or client terms that rule out outside AI services. Keeping the model inside your perimeter settles the question by design instead of by vendor promise.

Yes: you need to run disconnected

Air-gapped sites and networks with no internet route. Local models are the only option there, and the major platforms now support it, for example Foundry Local on Azure Local for Microsoft estates.

Yes: steady, high volume

Fixed hardware cost against per-token pricing only wins when the load is large and constant, such as overnight batch extraction across millions of pages. Published break-even estimates vary widely, so we model it on your own volumes.

Probably not: you need the strongest reasoning

The best closed models still lead on the hardest reasoning and agentic coding tasks. If accuracy on complex judgment is what matters, a frontier model in your own cloud account may be the better trade.

Probably not: low volume and no one to run it

GPUs need patching, monitoring and capacity planning. For a handful of users, a managed model with zero data retention in your cloud region usually costs less in total.

Where it can run

A single GPU server in your office or data center for one team or one process.

  • One 80–96 GB GPU runs 27–31B models and gpt-oss-120b
  • Fastest route to a working pilot on real data
  • No cloud dependency at all

Open-weight models we evaluate (November 2026)

  • Gemma 4 (Google DeepMind)

    Apache 2.0. Sizes from phone-class models to a 26B mixture-of-experts and a 31B dense model with up to 256K context. The 31B fits unquantized on one 80 GB GPU.

  • Qwen3.8-27B (Alibaba)

    Apache 2.0. A dense 27B model with image input and 262K context, strong for its size on coding and office tasks. Larger Qwen releases use their own license, so we check each one.

  • gpt-oss-120b and 20b (OpenAI)

    Apache 2.0. Reasoning models built for tool use. The 120b runs on a single 80 GB GPU; the 20b on much smaller hardware.

  • Mistral Small 4 (Mistral AI)

    Apache 2.0. 119B parameters with about 6.5B active per token, 256K context, reasoning and image input in one model.

  • DeepSeek V4 Flash (DeepSeek)

    MIT license. 284B parameters with 13B active and a 1M-token context, aimed at agent work.

  • How we choose

    We check license and publisher against your procurement policy, shortlist two or three models and test them on your own documents against a cloud baseline, starting from any AI stack you already run.

Local is not automatically secure

Network isolation

The model endpoint is not reachable from the internet, outbound traffic is blocked, and only named applications can call it.

Access that follows your existing permissions

Search and agents only retrieve documents the person asking can already open, using your directory groups and matter or account permissions.

Model supply chain

Weights come from the publisher’s official repository, are checked against published hashes, and use safe formats such as safetensors. Each version is recorded so you know exactly what ran.

Logging and retention

Prompts, sources and outputs are logged to the extent your policy allows, kept on your systems and covered by your existing retention rules.

Prompt injection

Running locally does not stop a malicious document from instructing an agent. Agents get only the tools they need, and actions that change records wait for a person.

How a local AI project runs

  1. Classify

    We map which data and tasks must stay local and which can use a cloud model, and agree the rules with your compliance team.

  2. Bake-off

    Two or three open-weight models tested on your own documents and questions, against a cloud baseline, with an accuracy target you set.

  3. Size and cost

    Hardware or private-cloud sizing for your real volumes, and a three-year cost compared with the cloud option.

  4. Build and harden

    Serving, search, agents and the controls above, integrated with your identity, logging and monitoring.

  5. Run

    We operate it, including model updates and re-testing, or hand over to your team with a run guide.

What you get

  • A data classification showing what runs locally, what may use a cloud model and why.
  • Bake-off results on your own data, including where the local model falls short.
  • A sizing and three-year cost comparison you can take to finance.
  • A working local service with no outbound data path, plus security controls and audit logs your risk team can review.
  • A run guide covering patching, model updates and re-evaluation.

Related

Further reading

Data that cannot leave the building?

Tell us what it is and what you want AI to do with it.

Talk to usStart with a Blueprint

Common questions

What is local AI?

Running an AI model on hardware you own or control, such as your own servers or a private cloud account, instead of sending data to a third-party AI service. The model weights, your data and every prompt stay inside your environment.

Is a local LLM as good as ChatGPT or Claude?

For most document, search and drafting work, current open-weight models are close enough. On the hardest reasoning and agentic coding tasks the best closed models still lead. We test both on your own data before you decide.

Can Claude run locally?

No. Claude is not released as open weights. You can use it through Amazon Bedrock, Google Cloud Vertex AI or Microsoft Foundry under your own cloud account and region, and route only the most sensitive work to a local model.

What hardware do we need?

It depends on the model and the number of users. One server with an 80–96 GB GPU runs 27–31B models for a team. Larger mixture-of-experts models need a multi-GPU server. Small desktop AI boxes are good for building and testing, but slow for many users at once.

Is local AI cheaper than an API?

Only at steady, high volume. Below that, per-token pricing usually wins once you count hardware, power and the people who run it. Most clients choose local AI for data control, not cost, and we model both on your volumes.

Does India’s DPDP law require local AI?

Not in general. The DPDP Rules, notified in November 2025, allow cross-border transfers unless the government restricts them, and most obligations apply from May 2027. Your sector regulator, your contracts or your clients may still require data to stay in-house. This is not legal advice.

Do you train models on our data?

No, unless you ask us to in writing for a model dedicated to you. Most local AI work needs no training: the model answers from your documents at the time of the question.