LiveAI agent foundation
AI supplies the judgement; LiveBPM guarantees reliable execution. Eight layers of capability add understanding, extraction and judgement to every step of a process, while every change of business state is still made by the deterministic engine under human authorisation.
- LiveAI is our enterprise agent foundation; all eight layers L1–L8 are shipped
- AI never changes business state; it produces evidence-backed recommendations and confirmation items
- Four authorisation levels L0–L3, widened or tightened by measured data
- Three levels of degradation; on-premise or masked-API data modes
AI never changes business state
What it changes is the information a person has when they decide.
The system is cut in two. Anything that changes state, records an amount, writes an approval or moves a document is executed by the process engine. Anything that judges whether something is reasonable, how risky it is, or which policy applies is handled by an agent, which returns a recommendation with its evidence, attached to the process instance as structured data.
Even if an agent hallucinates, is manipulated by prompt injection, or drifts after a model upgrade, the worst it can do is produce a wrong recommendation that a person rejects. It cannot change the amount on an order.
Conclusion: this quotation is 7.8% above the cost pass-through baseline, outside the 5% tolerance. Review suggested.
Evidence: (1) raw material index up 3.1% quarter on quarter, external data, two days old. (2) comparable quotations from other suppliers up 2.4% over the same period. (3) this supplier's twelve-month delivery rating: A.
Applicable policy: Procurement Management Rules, clause 6.3, price anomaly review.
Suggested actions: request a cost breakdown; compare with the second bidder; or keep the price and record the reason.
Confidence 0.86, with the numeric part computed deterministically. Authorisation L2, explicit confirmation.
LiveAI L1–L8
All eight layers are shipped and carry production scenarios. New capabilities follow the same rule: validated in a sandbox first, then admitted to delivery scope.
Model access and routing
OpenAI-compatible protocol, routing by data classification; sensitive data to a dedicated instance, general traffic to a public API.
Data ingestion and classification
Connects internal systems and external sources, tagging sensitivity field by field to drive masking and routing.
Anti-hallucination grounding
Full RAG pipeline. Anything uncertain is marked as unstated rather than invented, and every conclusion traces back to source text.
Information extraction
Multi-format documents into structured data with field-level confidence; low-confidence results routed to human review.
Rule engine
Hard rules implemented as code rather than prompts: no hallucination, reproducible, auditable.
Tool wrapping and agents
System operations wrapped as tools, natural-language querying across databases, multi-step task orchestration.
Human-machine collaboration
Review routing, approval records and write-back, connected directly to LiveBPM work items and approvals.
Governance and observability
Sensitivity tagging, PII detection, lineage auditing, fabrication guards, cost quotas and compliance boundaries.
Continuing above the eight layers: multi-agent orchestration and task decomposition, an industry template library, and appliance-style on-premise packaging. Time-series forecasting, computer vision and classical machine learning sit outside L1–L8 and are scoped separately.
Ten scenarios, without a new line of foundation code
All of them assembled from the eight layers by configuration. Business logic lives in configuration; the foundation stays generic.
Extraction · unstructured to structured
Field-level extraction from contracts, supplier documents, invoices and forms, with confidence scores and source positions.
Question answering · querying and explaining across systems
Natural-language questions across databases, with an explicit refusal when the data is insufficient rather than an invented answer.
Scoring · grading and normalising by rule
Classification, definition alignment and valuation recomputation. Calculation runs deterministically and can be reproduced.
Review · clause-by-clause comparison
Contract alignment, policy comparison and applicability checks, answering what matters about this specific document.
Four levels of authorisation
Graded by consequence and confidence, each level mapping to a different process engine configuration. Most agents sit at L1 or L2; L0 is used only where no money is involved.
| Level | Name | Relationship | Typical use | Engine configuration |
|---|---|---|---|---|
| L0 | Automatic | AI acts; reviewable afterwards | Due-date reminders, document generation, OCR pre-fill, scheduled exception lists | System step, logged only |
| L1 | Silent recommendation | Accepted by default, reversible at any time | Ranking, suggestions, field mapping, first-pass classification | Automatic step plus record, one-click revert |
| L2 | Explicit confirmation | A person must confirm the judgement | Exception interception, variance alerts, tolerance breaches, clause deviations | Creates a work item; overriding requires a reason |
| L3 | Human-led | AI supplies material only | Supplier removal, major contract award, budget exceptions | Multi-level countersign; AI output is an attachment |
Autonomy is earned
The system records acceptance rate, revision rate, override rate with reasons, and post-hoc accuracy for every class of judgement. Three consecutive months above the threshold with no serious error can move a judgement from L2 to L1; a rising override rate escalates it automatically and triggers a knowledge review.
Every interaction is training data
When an agent proposes and a person overrides with a reason, that record is at once a negative example, a piece of tacit organisational knowledge and a candidate rule. Keeping it structured turns judgement into an asset that survives staff turnover.
We answer the AI-specific risks up front
| Risk | What it looks like | How it is controlled |
|---|---|---|
| Data sovereignty | Whether business data, policies and customer information leave the company | Fully on-premise deployment supported; in API mode a model gateway enforces masking under a data processing agreement. The customer chooses. |
| Hallucination | Conclusions that read well but are wrong | Evidence chain and sources mandatory; numeric conclusions computed deterministically; critical judgements cross-checked by two models; L2 and above require human confirmation. |
| Prompt injection | Malicious instructions embedded in external text | External input filtered; prompt templates strictly separated from user data; agents hold no write access to master data, so manipulation causes no material loss. |
| Behaviour drift | Judgement changes after a model upgrade | Model versions under configuration management, with a golden test set of real historical cases regressed automatically on every upgrade. |
| Runaway cost | Call volume spikes | Quotas and budget alerts per agent and per organisation; batch work off-peak; simple tasks on small models, complex judgement routed to larger ones. |
| Explainability | Being unable to explain a decision to an auditor | Every call recorded: input snapshot, retrieved knowledge, model version, conclusion, confidence and the human outcome. Fully replayable. |
LiveAI questions
What is LiveAI?
LiveAI is our enterprise agent foundation. It has eight layers: model access and routing, data ingestion and classification, anti-hallucination grounding, information extraction, a rule engine, tool wrapping and agents, human-machine collaboration, and governance and observability. It adds extraction, review, scoring and question-answering to each step of a process.
Can the AI change business data or approve on its own?
No. LiveAI never changes business state. It can read process context, write a recommendation with its evidence, and create a work item for a person to confirm. In L0 authorisation it may trigger reminder-type helper flows. State changes, amounts and approval records are always executed by the LiveBPM engine under human authorisation.
What happens when the AI is wrong or hallucinates?
Every judgement carries an evidence chain and its data sources. Numeric conclusions are computed deterministically rather than generated by a model. Critical judgements can be cross-checked by two models. Anything at L2 or above requires explicit human confirmation. In our ten-scenario validation, confidence-based routing and the rule engine together intercepted 75% of the error cases before they reached a user.
Does the business stop if the AI service is unavailable?
No. The intelligence layer and the business layer are physically decoupled, with three levels of graceful degradation: a single failing agent simply hides its recommendation; if the model service is down the system falls back to pure rules; if an external data source is down the most recent cache is used and clearly marked as stale. Process integrity is guaranteed by the engine, not by the AI.
Does our data leave the company or train public models?
Fully on-premise deployment is supported, in which case no data leaves your environment. In API mode traffic passes through a model gateway that enforces masking, under a data processing agreement. The model layer speaks the OpenAI-compatible protocol, so sensitive data can be routed to a dedicated instance. Customer data is not used to train public models.
What is the four-level authorisation model?
AI actions are graded by consequence and confidence: L0 automatic execution, L1 silent recommendation, L2 explicit confirmation, L3 human-led. Most agents sit at L1 or L2. The level is widened or tightened based on measured acceptance rate, override rate and post-hoc accuracy, not on opinion.
Want to know what AI can and cannot do in your processes?
Send us your policy documents and a few sample documents. We will tell you which judgements suit an agent and which must stay with a person.