SAIE for the energy sector is a sovereign AI environment for critical infrastructure that runs in the company’s infrastructure, with an emergency procedure and a plan for rolling back agent actions. Sovereign AI for critical infrastructure - with an emergency procedure and a rollback plan
Why now
Operating and emergency procedures are scattered, and in a critical situation what counts is how fast you reach the right one. Data from critical infrastructure must not leave the company, NIS2 and the KSC Act require evidence of oversight over AI, and staff spread across branches do not adopt new tools without local leaders. Operators must implement the security measures under the amended KSC Act (Poland’s NIS2 implementation) by 3 April 2027, and from 28 October 2026 KRiBSI (Commission for the Development and Security of Artificial Intelligence) can inspect the use of AI. AI in critical infrastructure management is a high-risk system - AI Act obligations will cover it from 2 December 2027. Cybersecurity in the energy sector and NIS2 requirements now set the pace of AI deployments: according to an EY survey (June 2026), 71% of energy companies already have an AI strategy, and data security is their number one barrier (46%)
Three people who need to say yes
Operations / grid director - the right procedure in seconds, controlled agents
CISO / KSC officer - data in the company’s server room, evidence for NIS2
HR / transformation director - adoption across dispersed branches
Business scenarios
Assistant for operating and emergency procedures
Knowledge and documents
TWIN:DESKNEW · SAIE PANTERAProblem: Operating and emergency procedures are scattered, and in a critical situation what counts is how fast you reach the right one
How it works: The central RAG knowledge base points to the right procedure, citing the relevant passage of the source document. Agentic RAG asks again in another round when something is missing, and TWIN:INDEX searches long documents by chapter
What you gain: The dispatcher gets the right procedure with a citation in seconds - in a situation where searching through folders is not an option
Controlled rollback of AI agent actions
Process automation
TWIN:DESKNEW · SAIE PANTERAProblem: An agent has carried out a sequence of operations that must be undone immediately
How it works: A DAG-based rollback of agent actions - the user sees the rollback plan and confirms the operation. In the agentic loop the agent proposes a plan for approval, and the run history shows every step
What you gain: A wrong sequence of agent actions is reversed in a controlled way, with a visible plan - no manual unwinding after the fact
Searching grid diagrams and technical documentation
Knowledge and documents
TWIN:DESKProblem: Grid diagrams and technical documentation are not text-searchable
How it works: Advanced OCR that describes drawings and charts in a form suited to the model
What you gain: A grid diagram can be found with a question, not by browsing an archive of drawings
AI adoption in a dispersed organisation
Adoption and skills
CDFProblem: Staff spread across branches do not adopt new tools
How it works: AI Champions in every branch and the CDF Academy with internal certification (CDF-F3)
What you gain: Every branch has its own change leader and trained users - the tool takes hold in the field, not only at headquarters
Sovereign AI for critical infrastructure
Security and sovereignty
PROXY:AI + CDFNEW · SAIE PANTERAProblem: Data from critical infrastructure must not leave the company
How it works: Sovereignty model (CDF-F1) - Level L4, isolation from the public network; routing only to models in the company’s server room; Energy sector profile. Offline Model Registry matches the model to the hardware and starts it from a Harbor package in a network without the internet
What you gain: AI runs entirely in the company’s server room - and you can document this for auditors and the regulator
A secure AI environment instead of public tools
Security and sovereignty
PROXY:AI + CDFProblem: Staff use public AI tools outside the organisation’s control
How it works: A central gateway as the single point for model calls, with a usage register; a scan for unauthorised AI tools and a path to legalise them
What you gain: The company knows who uses AI and for what; staff get a legitimate alternative instead of a ban
Evidence of AI oversight for NIS2 and the KSC Act
Compliance and oversight
PROXY:AI + CDFProblem: An NIS2 and KSC audit requires evidence of oversight over AI
How it works: A register of AI agents with an owner, autonomy level and emergency shutdown procedure; an immutable event log with export
What you gain: Evidence of AI oversight is produced by the system as you work, not assembled before the audit
Questions decision-makers ask
What legal framework ensures a safe deployment?
The AI Act, the amended KSC Act (NIS2) and an internal AI policy linked to the information security management system. The CDF methodology guides you through the Initial Assessment, the classification of use cases and documents you can show an auditor
On-premises or cloud AI?
For critical infrastructure data - on-premises. At LW Bogdanka, TWIN:DESK runs entirely in the client’s infrastructure, with no data going outside, and the deployment took two months
How do you organize historical data and access to it?
Start with the documents people search for most often, e.g. operating and emergency procedures, and assign permissions by department. At LW Bogdanka, about 80% of the knowledge base consisted of scans, which TWIN:DESK processed for search with citations
Can an energy company covered by NIS2 use ChatGPT?
Yes, if the company has approved it as a tool: with a contract, control over data location and a usage register, because the amended KSC Act requires securing the supply chain of ICT services. Data from critical infrastructure should not, however, go to public chatbots. The best option is to give staff access to models through PROXY:AI, which masks data and records every query, or to run AI in the company’s server room
Can AI in nuclear energy run without the internet?
Yes. SAIE runs in a network completely cut off from the internet, including the isolated operational networks of nuclear facilities: Offline Model Registry matches the model to the hardware and starts it from an installation package, and the knowledge base and event log stay in the operator’s server room. It is the same L4 sovereignty level we use in critical infrastructure
All questionsOther paths: Critical infrastructure
AI for an essential service operator - with an evidence package and maintenance a small IT team can afford
A single AI gateway for the whole organisation, with audit, budget and a contingency plan
Documents read themselves, complaints come with a draft reply, and pilots scale across branches