Automation and AI
Automate the known. Apply AI to the ambiguous. Keep people in control.
A workflow that resolves clear cases with rules, uses a self-hosted model only where judgement is genuinely needed, checks that recommendation against policy and confidence, and puts uncertain cases in front of a person before anything happens.
- AI only where needed
- Rules firstAI only where needed
- before any auto-action
- Policy checkedbefore any auto-action
- where confidence is low
- Human reviewedwhere confidence is low
- decision and reason
- Fully recordeddecision and reason
The operational problem
Either everything is manually reviewed, or a model is trusted more than it should be.
Most organisations sit at one of two unsatisfactory extremes: every case gets a manual review regardless of how clear-cut it is, wasting the reviewer's attention, or AI output is wired directly into an action with no policy gate, which nobody wants to defend after the first mistake.
The audit gap shows up when a model made a call and nobody can say afterwards what the confidence was, whether policy was checked, or who, if anyone, looked at it before the action was taken.
What a working solution has to do
Separate the deterministic, the ambiguous and the accountable.
Resolve the clear cases without a model
Deterministic rules should decide anything that can be decided deterministically. A model is a cost and a review burden that should only be spent on genuine ambiguity.
Treat model output as a proposal
A recommendation, classification or summary from a model has to be checked against policy and a confidence threshold before anything happens with it.
Put a person in front of anything uncertain
Cases that fail the confidence or policy check need to reach a reviewer with full context, not be auto-approved because a threshold was close.
Capture the decision as structured evidence
What was recommended, what was decided, and why, has to be recorded in a form that supports both audit and deliberate improvement later.
The learning loop
Event, rules, recommendation, check, review, action, capture, improve.
This is the chain Nexus is built around for any process that mixes deterministic logic with model judgement and human accountability.
- 01
Event
A case enters the workflow: an invoice, a support ticket, a document, a request.
- 02
Deterministic rules
Rules resolve everything they can. Only genuinely ambiguous cases proceed.
- 03
AI recommendation
A self-hosted model proposes a classification, extraction or summary for the remaining cases.
- 04
Policy and confidence check
The recommendation is checked against policy and a confidence threshold before it can proceed unreviewed.
- 05
Human review where required
Cases that fail the check, or that policy always routes to review, reach a person with full context.
- 06
Approve, reject or correct
The reviewer's decision, not the raw recommendation, is what the workflow acts on.
- 07
Operational action
The workflow performs the write-back or next step authorised by the decision.
- 08
Capture and improve
Decision and outcome are recorded, and used deliberately to improve prompts, retrieval, policies or evaluation.
The loop repeats. Improvements to prompts, retrieval, evaluation, policy and thresholds are authored, tested and published by people who own the process.
Worked example
An expense claim that needs judgement, not just a limit check.
- 01
Trigger
An expense claim is submitted through the finance system, above the value where policy previously required manual review of every line.
- 02
Rules + Code
Deterministic checks apply per diem limits, category rules and duplicate detection, clearing straightforward claims immediately.
- 03
Sovereign AI
For claims with unusual categories or missing receipts, a self-hosted model classifies the likely category and drafts a justification summary from the attached notes.
- 04
Policy Check
The classification is checked against a confidence threshold and a policy list of categories that always require review, such as client entertainment.
- 05
Human Approval
Low-confidence or policy-flagged claims reach a finance reviewer with the claim, the rule results and the model's summary together.
- 06
System Action
The approved, corrected or rejected outcome is written back to the finance system and the employee is notified.
- 07
Audit
The claim, rule outcomes, model recommendation, confidence score, reviewer decision and reason are retained as one record.
Evidence and output
What the record shows for every case.
Rule outcomes
Which deterministic checks passed, failed or were not applicable to this case.
Model recommendation
The classification, extraction or summary produced, with its confidence score.
Policy result
Whether the case cleared automatically or was routed to review, and which rule triggered that routing.
Reviewer decision
The person, the decision, the reason, and the data they had in front of them at the time.
Products and platform capabilities
Built from platform capabilities, not a separate AI product.
Platform: Sovereign AI
Provides the self-hosted model endpoints and knowledge bases the recommendation step calls.
Platform: Human-in-the-Loop
Provides the approval node, pause and resume execution, and reviewer context described on the platform pages.
Platform: Governance and Audit
Provides the identity, role-based access and execution record that make every decision attributable.
Who owns it
A process owner and a model owner, working together.
Process owner
Defines the policy thresholds, which categories always require review, and who the reviewers are.
AI or platform team
Maintains the model endpoint, monitors confidence distributions, and proposes changes to prompts or retrieval based on recorded decisions.
Reviewers
The people whose approve, reject and correct decisions become the evidence the process improves from.
Operation
Governed improvement.
Changes are authored, tested and approved
Teams use recorded decisions to improve prompts, retrieval sources, evaluation and policies. Model recommendations pass through deterministic checks and human review. Each process change is authored, tested and published by an accountable owner.
Architecture session
Tell us where AI judgement and human accountability need to meet.
We will help you separate what rules can already decide from what genuinely needs a model and a reviewer, and design the policy and confidence checks between them.