ContextLab · Core technology
Context first. Agents follow.
ContextLab is Axiobit's technology for capturing how an organisation works — its language, its processes, its rules and its expert judgment — and turning it into structured context that AI must work within. Every Axiobit agent is built on it.
01 / The bottleneck
Why context is the bottleneck.
Large models brought general intelligence to every company at once. What they did not bring is knowledge of any particular company.
In regulated and process-heavy sectors, an answer that is plausible but outside the process is worse than no answer. ContextLab closes that gap from the side of the enterprise, not the model.
02 / The four layers
Four layers of enterprise context.
Layer 01
Ontology
Your entities, terms and relationships, defined once and versioned.
- Role of process (BPM)
- Defines and versions the vocabulary
- Role of AI / ML
- Reasons within it; never redefines it
- Maturity
- Discuss availability
Layer 02
Process
Your workflows as explicit models: steps, owners, exceptions. The AI may act only where the process allows.
- Role of process (BPM)
- Source of truth for valid states and next steps
- Role of AI / ML
- Predicts likely next steps, flags anomalies, routes exceptions
- Maturity
- Discuss availability
Layer 03
Formal Knowledge
Policies, rules, contracts and case history, retrievable and applied at the right step.
- Role of process (BPM)
- Decides which knowledge applies at which step
- Role of AI / ML
- Retrieves and applies grounded knowledge; does not invent it
- Maturity
- Discuss availability
Layer 04
Tacit Capture
Expert judgment captured from real decisions and proposed as new rules, always confirmed by a human.
- Role of process (BPM)
- Flags where the process lacks a rule
- Role of AI / ML
- Proposes candidate rules from human overrides; humans confirm before promotion
- Maturity
- Research track
| Layer | What it holds | Role of process (BPM) | Role of AI / ML | Maturity |
|---|---|---|---|---|
| Ontology | Entities, terms, relationships of the business | Defines and versions the vocabulary | Reasons within it; never redefines it | Discuss availability |
| Process | Workflows as state models: steps, owners, transitions, exceptions | Source of truth for valid states and next steps | Predicts likely next steps, flags anomalies, routes exceptions | Discuss availability |
| Formal Knowledge | Policies, rules, thresholds, contracts, case history, regulation | Decides which knowledge applies at which step | Retrieves and applies grounded knowledge; does not invent it | Discuss availability |
| Tacit Capture | Expert judgment not yet written down | Flags where the process lacks a rule | Proposes candidate rules from human overrides; humans confirm before promotion | Research track |
03 / Design principle
Process as a boundary.
- 1Request
- 2Current process state
- 3Allowed actions and fields
- 4Model
- 5Validated output
- 6Next state
- The process model stays a separate, versioned artefact that a compliance reviewer can read without opening any model.
- At run time it restricts what the AI may output: valid next actions, valid data fields, valid escalation paths.
- Process changes do not require retraining. They are edited, versioned and deployed like any business rule.
04 / Small data
Built for small data.
Specialised ML where the data is structured.
Classification, propensity, forecasting and anomaly models trained on the company's own records, with explainable drivers.
Language models where the data is language.
Used for understanding requests, documents and conversations, always grounded in the Formal Knowledge layer.
Most companies do not have big data. They have good data, in modest volumes, spread across systems. ContextLab is designed for that reality.
05 / Deployment
Runs on your infrastructure.
Designed for deployment on company infrastructure, on-premise or in a private European cloud. Deployment requirements are agreed for each engagement.
06 / Learning
A system that learns, under control.
- Execution traces (state → decision → outcome) are collected from live processes and used for periodic improvement of the specialised models.
- Nothing becomes a rule without human confirmation.
07 / Comparison
How ContextLab compares.
| Approach | What it does well | What it lacks for enterprise use |
|---|---|---|
| General AI assistants | Broad language ability | No knowledge of your processes or rules |
| Document chatbots (RAG) | Answers from your documents | No process awareness; cannot act safely |
| Process mining | Shows how processes really run | Observes; does not constrain or act |
| Agentic workflow tools | Chains model calls across steps | The model is not bounded by the process |
| ContextLab | Structured context + process as boundary | Requires upfront modelling of the process (we do it with you) |
08 / Inside ContextLab
Context layers and AI Technologies.
ContextLab has two parts that work together. The context layers (Ontology, Process, Formal Knowledge, Tacit Capture) make agents reliable: they define what an agent knows and where it may act. AI Technologies make agents capable: the models chosen for each task — gradient boosting, language models, forecasting, speech and document understanding — plugged in only where the process allows.
Want the technical detail?
We share the architecture brief with technical teams and research partners on request.