Bounded intelligence around deterministic control
Use industrial AI without giving it undefined authority over the machine or your engineering evidence.
Axiotech designs on-premise or controlled AI services around explicit data contracts, operating envelopes, confidence, provenance, escalation and fallback. Deterministic PLC and safety responsibilities remain visible.
Good fit
Use this service when
- Bounded decision support, classification, retrieval, vision or anomaly use cases with a named owner and measurable acceptance.
- Edge or on-premise deployments that need PLC, OPC UA, ADS, historian or engineering-document integration.
Scope boundary
Not assumed or silently included
- Placing an unvalidated probabilistic output directly in a safety function or unrestricted control loop.
- Claims of autonomy, accuracy or return before representative data, drift and operating conditions are assessed.
Why act
The operational risk behind the technical symptom
The first work package is shaped around reducing an explicit risk or enabling a named decision.
A model performs well in demonstration data but fails silently as the process changes.
Users act on outputs without provenance, confidence or an understood escalation path.
Private deployment is assumed secure despite ungoverned models, packages, access and update channels.
Purchasable entry points
Choose a bounded engagement before expanding scope
Final inclusions, dependencies, location, schedule and commercial terms are confirmed in a formal quotation.
Package 1
AI use-case & boundary assessment
Decision enabled: A go, reshape or stop decision grounded in data and operational risk.
Package 2
Controlled proof of capability
Decision enabled: Representative evidence with limitations and production gaps exposed.
Package 3
Production integration work package
Decision enabled: A monitored service with fallback, release and operating ownership.
Artefact manifest
What remains after the engineering work
The exact document set scales with risk and the approved work package.
- Intended decision, authority and AI boundary record
- Dataset provenance, quality and representativeness assessment
- Model/service evaluation and limitation statement
- Interface, fallback, monitoring and cybersecurity design
- Release, update, rollback and operational-review plan
Technical and responsibility boundaries
Make ownership reviewable
- Safety and deterministic machine control remain outside the AI authority unless separately engineered and approved.
- Human review is designed where uncertainty or consequence requires it; it is not a slogan.
- Model, data, software and infrastructure changes follow named release and rollback control.
Review and acceptance
Six gates from authority to handover
Gate depth changes with the work; named decisions and evidence remain.
- G0Scope authorityOutcome, boundaries, roles, assumptions and commercial basis agreed.
- G1Baseline acceptedKnown installed/source state, dependencies, constraints and unknowns recorded.
- G2Design approvedRequirements, interfaces, risks and acceptance evidence ready for implementation.
- G3Release candidateBuild reviewed, verified, versioned and accompanied by defect disposition.
- G4Site acceptanceCommissioning evidence, deviations and release decision recorded.
- G5Handover acceptedSource, configuration, records, recovery and residual risks transferred.
Assurance option
Compliance is a declared scope—not a badge
Applicable AI, machinery, data-protection, cybersecurity and regulated-system obligations depend on intended use and authority. Axiotech documents the technical boundaries and evidence in its scope; legal classification and organisational approval remain named customer responsibilities unless separately commissioned.
Review validation servicesRelevant engineering evidence
Private engineering assistant with controlled evidence
A reference architecture for retrieval from approved engineering sources with citations, access boundaries, abstention, review and no direct control authority.
Procurement FAQ
Questions to resolve before quotation
Can the system run without sending data to a public cloud?
Potentially. Feasibility depends on model, hardware, update, support and security requirements; the deployment boundary is an architecture decision.
Can AI write to the PLC?
Only through a deliberately constrained interface with explicit authority, validation, permissives, fallback and risk acceptance. Many use cases should remain advisory.
How do you prove an AI feature works?
Acceptance combines representative evaluation, baseline comparison, operational scenarios, failure behaviour, limitations and a monitoring plan—not one accuracy figure.
Next decision
Describe the outcome, installed baseline and constraints you already know.
Use “Not known” where evidence is missing. The first review will separate facts, assumptions and discovery work.