Software and intelligent automation

On-premises AI

We set up AI where off-the-shelf tools cannot reach: on your own server, in your network and server infrastructure, in your systems that have no API. Together we write down what it will read and what it will change.

Fiber optic patch panel connected with neatly arranged turquoise fiber cables

Who it is for

  • Organizations that do not want to send customer, patient or staff data to an AI service abroad
  • Multi-location businesses that run network and server operations with a small team
  • Organizations whose critical work runs on legacy software or a device without an API
  • Hotels, construction sites and production sites where the connection can drop
  • Teams that have tried off-the-shelf tools and hit a data, integration or authority limit

Mature off-the-shelf tools exist for invoice reading, e-mail classification and chat assistants. If that is what you need, we say so plainly during discovery.

Which problem it solves

Problems and solutions
AreaCurrent situationThe system we set up
Network and server operationsAlerts pile up, and fault analysis is done by hand, device by device.The agent correlates alerts, collects status from devices and reports the probable cause with its evidence. The change recommendation comes for your approval.
Configuration auditDevice settings change over time, and no one tracks the difference.It reads configurations regularly, compares them with the approved template and reports risky differences with an explanation.
On-premises AICloud AI services cannot be used because of sensitive data.An open-weight model runs on your own server. Document search, correspondence and analysis stay inside your organization.
System without an APIData in legacy software is moved to another system by hand.It reads and writes through the database, files or a device protocol. Screen automation is the last option.
Field sitesThe connection is intermittent, and a cloud-dependent solution can stop.The model runs on the server at the location; when the connection returns, it syncs the records to the central site.

API: the official interface through which software systems talk to each other. SSH, SNMP, syslog: protocols for management, measurement and logging with network devices and servers. Open-weight model: an AI model that can be run on your own hardware.

How we work

  1. Discovery

    From youA description of the task, sample data, the systems to connect, where the data will stay

    From usWe examine the task, the data flow and the access paths; we report on suitability and risks

    Output: Suitability analysis and success criteria

  2. Pilot

    From youTest data, the people who will give approval, a test environment or read-only access

    From usWe set up the limited-scope system and measure it

    Output: Working pilot

  3. Evaluation

    From youJoint review of the results and the decision on whether to continue

    From usWe compare the results with the criteria; we recommend the installation scope, the authority and the hardware

    Output: Pilot evaluation report

  4. Installation

    From youLive system access and authority; for an on-premises installation, server space

    From usWe set up the model, the server and the integration; we run the acceptance tests and train the users

    Output: Working system and delivery documents

  5. Operations

    From youChoice of the monitoring and maintenance scope

    From usWe monitor accuracy and exceptions, improve the system and report

    Output: Monthly performance report

The pilot is a limited trial run with real or anonymized data. For infrastructure agents, the pilot starts with read-only authority; the agent does not change systems, it only makes recommendations.

Authority stays with you

Before work starts, we write down what the AI will do without approval, where it will stop and wait for your approval, and how every step will be logged. Infrastructure agents start with read-only access.

Automation levels
  1. RecommendationThe automation prepares the work; an employee checks and applies it.
  2. Approved actionThe automation prepares; an employee approves, and the automation applies.
  3. Automatic except for exceptionsWork within the defined limits is applied without approval; cases outside the limits or unclear cases are routed to an employee. The move to this level is made if the pilot results are sufficient, and with your approval.

What is delivered

  • Suitability analysis and success criteria
  • Pilot evaluation report
  • Workflow definition: steps, approval points, action thresholds
  • Authority matrix: which system is accessed with which authority
  • Action log and monitoring screen
  • Data flow document: which data is processed in which model, on which server and in which country
  • For an on-premises installation, server, model and network configuration
  • Source code and configuration (according to the intellectual property model in the contract)
  • Operations guide and user training

Acceptance

Acceptance is done in two stages. In the pilot, the criteria are measured on an evaluation set prepared in advance, and the move to installation is decided on that result. During installation, the acceptance tests are run with real or sample records; the error rate and the share of tasks that go to approval are reported. Work that meets the criteria is accepted.

Agent workflow definitionIllustrative

Network fault first-analysis workflow

Target level: approved action · 6 steps

  1. TriggerThe monitoring system raises a connection loss alert at a branch.
  2. CollectionThe agent collects interface, routing and log information from the relevant routers and switches with read-only access.
  3. AnalysisIt correlates alerts from the same minute and lists the likely cause with evidence.
  4. RecommendationIt prepares the configuration diff and the rollback steps for the fix.
  5. ApprovalThe change goes to the network owner for approval; it is not applied without approval.
  6. LogEvery read and action is logged; the incident report goes to the team.
Authority: read-only + approved changeNo authority to delete or shut down

Out of scope and limits

Work that is not part of this service unless written separately in the proposal, and the rules that apply to every project.

Out of scope

  • Licensing, installation and vendor support for the connected systems (ERP, CRM, network devices)
  • Developing or training a model from scratch; existing models are configured to work with your data
  • End-user computer and application support

Working rules

  • AI outputs can contain errors. Error limits are measured and reported in the pilot; no specific accuracy or savings are guaranteed.
  • Infrastructure agents start with read-only authority. Authority to make changes depends on the pilot results and your written approval.
  • The response quality and speed of an on-premises model depend on the hardware; the hardware recommendation is based on the pilot measurement.
  • Irreversible actions such as payment, deletion or a configuration that could interrupt service require human approval unless you request otherwise in writing.
  • In projects that process personal data, a data processing agreement is signed. Personal data is not sent to a model abroad without a documented legal basis.
  • When the connected systems change, the integration is updated; who is responsible for maintenance is determined at delivery.

What determines the price

The price is set in a written proposal once the need is clear. The service fee and the hardware, license and usage costs appear on separate lines.

  • The systems to connect and the access path (API, database, device protocol)
  • Model choice: cloud service or on-premises model
  • For an on-premises installation, server, GPU and network requirements
  • Agent authority, approval points and automation level
  • Processing volume and exception rate
  • Scope of monitoring, maintenance and improvement
Proposal line items
ItemContentIn the proposal
Service feeDiscovery, pilot, installation, testing, documentation, trainingDiscovery and pilot at a fixed fee; installation at a project fee
Model and hardwareCloud model usage (API) or on-premises server and GPUSeparate line. API usage is paid from your provider account, and the hardware is your property
Licenses and subscriptionsConnected software, cloud servicesSeparate line, in your name
Monthly operationsMonitoring, maintenance, improvementOptional, under a separate contract

Frequently asked questions

Does our data go to AI providers?

In an on-premises installation, it does not; the model runs on your own server. If a cloud model is to be used, which data goes to which service and which country is documented in advance and submitted for your approval.

Does an on-premises model give results as good as cloud models?

It depends on the task. Open-weight models can give sufficient results in tasks such as classification, summarization and document search; there can be a difference in complex reasoning. The decision is made in the pilot, by measuring with your data.

What happens if the agent makes a mistake on our network or servers?

The agent starts with read-only authority. When authority to make changes is granted, a configuration backup is taken before every change, a summary of the differences is submitted for approval and a rollback plan is prepared. Every action is logged; the agent can be stopped at any moment.

What happens if our system has no API?

The connection path is determined during discovery: database, file transfer or device protocol (SSH, SNMP). Screen automation is the last option; the maintenance need is stated clearly in the proposal.

How long does the pilot take?

It depends on the scope. The duration and the success criteria are written down before the pilot starts.

Who owns the source code?

Ownership is defined in the proposal and the contract. The code of the workflow and software developed specifically for you is kept in a repository you have access to.

Let’s talk about the task that off-the-shelf tools cannot solve.

In the form, select “Intelligent automation and AI”; describe the task, the systems to connect and where the data must stay in a few sentences.

Discuss Your Project