How Locally-Hosted Enterprise AI Works

How Lumen, the private AI platform from Cognetryx, turns your own documents into answers you can check, all inside your network.

Compare it with building your own →

Here, Lumen summarizes a sample bank lending policy and opens the cited page on the right. Select 1, 2 or 3 to see how it works.

  1. Every claim cites its source

    Numbered citations sit next to each point in the answer.

  2. The source opens beside it

    Click a citation and the original file opens at the cited page.

  3. The exact line is highlighted

    A reviewer can check the claim against the policy itself.

The Lumen Platform

Taking Your AI & Data Private

The capabilities your teams use, all inside your network.

Sample data. Select a feature to play its demo.

The Process

How Your Data Stays Private

Every step of the pipeline runs inside your network. Nothing leaves.

Inside your network

  1. 01

    Connect your documents

    Lumen connects to SharePoint, Confluence, NetSuite, Salesforce, Google Drive, and more through their APIs, and to the files on your own servers.

    1. Upload

      Internal files brought in securely

    2. Indexing

      Prepared for search

  2. 02

    Ask in plain language

    Query internal documents using conversational language, no SQL required.

    1. Search

      By meaning, keyword, and link

    2. Rank

      Top passages selected by relevance

    3. Answer

      Written from the top passages

  3. 03

    Check the source

    Every answer links to the passage it used, highlighted in the original file.

    1. Cited Answer

      Each claim tied to its passage

    2. Audit Log

      Question, sources, and answer recorded

Every step above runs on your own servers, from indexing your files to writing the answer. Security matters more than ever. In IBM’s 2026 breach study, 43% of organizations had an incident involving unapproved AI tools, up from 20% a year before. The same controls keep your data and your AI in your hands. You choose the model and when it changes. The AI sees only what each person can open, needs sign‑off to file, send, or approve, and logs every step.

What private AI is, and why it matters →

Reasoning and Governance

It Reasons Before It Concludes

Enterprise workflows are diverse and range in complexity and utility. Take “Does our retention policy require seven years for this record?” versus “Is this transaction suspicious enough to report?” as an example.

One is clear. The other is based on a series of defined situational metrics and ironclad rules. Every industry has its own version of both. Many AI models answer both in the same confident voice, because they are trained to return the most probable answer. But is it the right one?

When a question has a direct answer, Lumen returns it immediately with citation. When a question calls for judgment, whether it arrives through document analysis or a direct query, the model lays out defensible readings and the considerations behind each. From there you decide: review each one, set policy, or automate the pattern once you know how that call should go.

What to notice in the live answer:

  1. Cites what’s settled. The 95% cap comes straight from the policy, with its page.

  2. Names the kind of question. A 97% request is an exception, so it needs committee approval.

  3. Shows where judgment comes in. It weighs how far over the cap the request is and says precedent will matter.

  4. Leaves the call to people. It tells the underwriter to build the case and let the committee decide.

GuideHow AI should handle judgment calls →

A longtime customer wants a home equity line on her owner-occupied home at 97% LTV. Our lending policy caps LTV at 95%. Can we approve it as an exception?

Based on your lending policy, the maximum LTV for a home equity line is 95% [1]. A 97% LTV request exceeds that cap, so it cannot be approved at the underwriter level—it would need an LRC exception.

Here’s the path forward:

  • It’s possible, but not automatic. Your policy establishes that loans exceeding standard LTV thresholds require LRC approval [2]. Rate and other policy exceptions similarly flow to the LRC [3].
  • You’d submit an exception request to the LRC with supporting documentation explaining the rationale—tenure as a customer, credit profile, cash reserves, equity cushion justification, etc. The LRC will weigh the overall risk of the transaction against the relationship value [4].
  • The 2-point overrun (97% vs. 95%) is material enough to warrant review but not so extreme that it’s automatically off the table—LRC precedent on similar cases will matter.

If the LRC denies it, the customer has the right to submit a written appeal [5].

Bottom line: you can’t approve this yourself. Package it as an LRC exception request with the strongest underwriting story you can, and let the committee decide.

5 sources
  1. Lending Policy Generic Bank MI.pdf · p. 34Maximum LTV: 95%. Appraisals: See Appraisal Policy. Lending Area: Contiguous United States, subject to the bank’s state licensing and applicable restrictions.
  2. Lending Policy Generic Bank MI.pdf · p. 44Loans exceeding 80% LTV may be approved by the LRC.
  3. Lending Policy Generic Bank MI.pdf · p. 10The bank’s cost of funds Rate exceptions are handled as follows: Rate reductions exceeding 2% require LRC approval.
  4. Lending Policy Generic Bank MI.pdf · p. 10The proposed rate relative to the overall risk of the transaction Any customer who disagrees with a credit decision may submit a written appeal to the LRC.
  5. Lending Policy Generic Bank MI.pdf · p. 10Any customer who disagrees with a credit decision may submit a written appeal to the LRC within a reasonable time following the
A live example from Lumen, answering from a sample bank lending policy. Select a citation to open its source.

Model Choice

Works With the AI Model You Choose

Lumen is our own proprietary platform, and it works with the AI model your firm chooses. Pick the model that fits the work, and decide when it changes.

Lumen connects to the systems you already use and to the files on your own servers. Your people get one place to ask questions, run agents, and capture meetings, with every answer tied to your own documents.

  • Stay in Control

    Lumen runs inside your network, and updates arrive as signed packages you install on your schedule.

  • Choose Your Model

    Run the model that fits the job, and switch when a better one arrives.

  • Connect Your Data

    Bring in the systems you use through their APIs, plus the files on your own servers.

GuideWhich jobs smaller AI models handle well →
The Lumen Platform

Inside your network

ChatAgentsMeetingsExports

  1. Security & Access

    Single sign-on, role-based access, and a full audit log in the Compliance Portal.

  2. Agents & Workflows

    Agents that run on a schedule or start on an event, with approvals built in.

  3. Search & Citations

    Finds the right passages in your files and shows the exact line behind every answer.

  4. Your AI Model

    Runs the model your firm chooses. Switch models as better ones arrive.

  5. Connections

    Links to your systems through their APIs, and to files on your own servers.

  6. Your Data

    Stays on your servers, and each person reaches only what they can already open.

Runs on your own servers

Deployment Options

Flexible Infrastructure Choices

Choose the deployment model that fits your organization's requirements and risk profile.

Locally-Hosted Infrastructure

Deploy and manage at your data center or colocation facility. Complete control over all aspects of deployment while keeping data within your organizational boundaries.

  • Processing closer to data for lower latency
  • Complete control over infrastructure
  • Full compliance and audit capabilities
  • Air-gapped deployment available

Private Cloud / Hybrid

For organizations with existing private cloud infrastructure, Lumen can be deployed within your secure environment with cloud-like agility.

  • Cloud-like agility and scalability
  • Maintains full data sovereignty
  • Integrates with existing cloud investments
  • AWS/Azure isolated VPC options available

Cost over time

One fixed cost as your firm grows

Cloud AI tools charge for each person, every month. Business and enterprise seats list at $20 to $30 per person on yearly plans.1 Fees for agents and heavy models come on top, and they grow as people use AI more.2

  • For a firm of 1,000, seat fees alone come to $1.2 million to $1.8 million over five years. At 2,500 people, they reach $3 million to $4.5 million.
  • Metered use adds to that. Microsoft sells Copilot Studio agent credits at $200 a month for each pack of 25,000,2 and GitHub moved Copilot to billing by the token in June 2026.3
  • Lumen has one fixed price, set by the scope of your deployment, and runs on servers you own. Firms spread a server’s cost over about five years.4 Your cost holds steady as use grows, so agents can run on every file, every night. It rises in steps only when you expand.

For a small team, cloud seats can cost less at first. The more people use AI, and the more they use it, the sooner your own servers pay for themselves.

  1. List prices, September 2026: Microsoft 365 Copilot $30; Gemini Enterprise from $30; Claude Enterprise $20 plus usage; ChatGPT Business $20.
  2. Microsoft, Copilot Studio pricing, September 2026.
  3. GitHub, GitHub Copilot is moving to usage-based billing, April 27, 2026.
  4. IRS, Publication 946: computers are five-year property.
Illustration: cloud AI cost rises as staff and use grow, while a private deployment stays flat and rises only in steps when servers are added, so the gap widens after the lines cross ILLUSTRATIVE Staff using AI, and how much they use it → Cost per year → Lines cross Cloud AI: seats plus usage fees Private: one fixed cost, added in steps The gap grows
Cloud fees climb with each hire and each task. Your own servers hold one fixed cost that steps up when you expand.

Featured Reading

Line drawing of an answer's citation on a monitor linked to a highlighted line under a magnifying glass, and a review sheet

Citations & Review

How to check an AI answer before it reaches an examiner

A source link beside an AI answer is a start. How to check that the passage backs the claim, the document is current, and the record holds up in an exam.

 Line drawing of a server rack being filled from a workbench of parts, under a project timeline, with a checklist

IT & Infrastructure

The private AI stack you’d build yourself: what it takes and how long

Running AI on your own servers: the parts to build, where it can run, what sets the timeline, and what to ask a vendor.

 Line drawing of an employee at a computer whose search reaches only the two open file drawers; the rest stay locked

Access & Agents

Permission-aware AI: answers and agents that reach only what the user can open

Make AI answers and agents reach only what each user can open: permission checks inside search, SSO roles, agent identity, audit logs, and tests.

See it answer from real documents

A 30-minute live demo. We reply within 24 hours.

See it in action