Aqera Advisory
Agentic AI Transformation
The process, not the pilot. We decide with you which work an AI agent should do, at what autonomy, under what oversight and at what cost. Then we build it so it can be audited, and redesign the human roles around it so the change lasts.
The problem
Every organisation now has AI agents in it, sanctioned or not. What goes wrong is rarely the model. Nobody mapped where the work was, nobody said which decisions an agent may take, nobody set the number that would stop it, and nobody redesigned the jobs around it. The result is pilot purgatory, and in a NSW agency it can also mean an AI system that should have been assessed under the AI Assessment Framework but never was.
Our method — four stages, nothing advances on assertion
| Stage | What happens | Exit criterion |
|---|---|---|
| 1 · Map | Map three to five value streams at operation level. Build a work inventory of steps, hand-offs, volumes, cycle times and rework. Catalogue every decision point and the evidence it relies on. Take a census of AI tools already in use, and overlay the legal, privacy and framework obligations on each step. | The sponsor signs the map as a fair description of the operation. |
| 2 · Evaluate | Score each decision point against six criteria: decision structure, evidence availability, reversibility and error cost, data readiness, obligation exposure, and volume and value. Classify each as not suitable, experiment (with a mandatory expiry date) or investment. | An approved agent portfolio. Every investment has a scale-or-kill threshold, kill criteria, a business owner outside IT and a date. |
| 3 · Build | Write a design record for each agent before any code. It sets the autonomy level and the human-in-the-loop control, the tools and data the agent may and may never touch, a held-out evaluation set and threshold, cost per unit at three adoption levels, and the supply chain. | The agent passes its held-out evaluation, runs under recorded standing conditions, and has its first production evidence registered. |
| 4 · Adopt | Redesign the roles at each decision point and set out oversight and escalation protocols. Plan how senior expertise will still be developed as agents absorb routine work. Track adoption through usage, override and escalation measures. Each benefit has a named executive owner, and a post-implementation review is scheduled at baseline. | An owner outside delivery reports realisation to the executive as a standing measure. |
Every agent is re-evaluated on schedule and whenever its underlying model changes. Experiments expire unless renewed by decision, and the agentic register is reviewed quarterly.
Engagement options
| Offer | Duration | Outputs |
|---|---|---|
| AI Readiness Diagnostic | 2–3 weeks | Facilitated assessment against the published Aqera Maturity rubric, heat map of constraining dimensions, gate-readiness findings. |
| Agentic Opportunity Scan | 3–4 weeks | Signed value stream maps for two or three streams, suitability assessments, a prioritised agent portfolio with thresholds, kill criteria and owners, and a recommendation on what to build first and what to stop. |
| Agent Pilot | 8–12 weeks | Design records, agentic register entries, held-out evaluation reports, deployment under standing conditions and first production evidence for one or two agents. Scale-or-kill threshold set in week one and judged at the end. |
| Scale and Adopt | Quarterly | Human-agent teaming design, role change, oversight protocol, benefit register with named owners, realisation reporting and quarterly register review. |
| Fractional Chief AI Officer | 2–8 days a month | The whole method run inside your executive, with a 90-day outcome plan and an explicit handover. |
| NSW AI Assessment Framework services | Scoped | Health check, compliance program and assurance retainer for agencies with AIAF obligations. |
Deliverables
Value stream map · Work inventory register · Decision-point catalogue · AI tool census · Obligation overlay · Agent-suitability assessment · Candidate register · Buy-partner-build decision record · Scale-or-kill threshold record · Agent design record · Agentic register entry · Held-out evaluation report · Standing-conditions record · Human-agent teaming design · Role change record · Oversight and escalation protocol · Judgment-continuity plan · Adoption measures · Benefit register entries · Post-implementation review schedule.
The method produces records, not slide decks. Each deliverable is either a record in AqeraOne or maps directly onto one, so the transformation is auditable by a Secretary, an Auditor-General or a framework reviewer without anyone assembling a pack.
Outcomes buyers can measure
- AI use in the organisation inventoried and registered, including tools nobody approved.
- Agents in production with recorded autonomy, oversight and evaluation thresholds.
- Experiments that ended on their expiry date rather than becoming unfunded production systems.
- Cost per unit of work, compared before and after.
- Override and escalation rates, showing whether human oversight is actually working.
Standards and alignment
Designed to align with the Australian Government's policy for the responsible use of AI in government, the NSW AI Assessment Framework, the Privacy Act (including automated decision-making transparency obligations), ASD guidance on secure AI, and ISO/IEC 42001 AI management system principles. Alignment means designed with reference to these, not certified against them.
Who it is for
The Deputy Secretary, COO, Chief Information and Digital Officer or Registrar who owns an operation, rather than the AI council that governs it. Governance answers "may we"; this answers "how, how fast, and does it become an investment".
Talk to us about your program.
Tell us what you are trying to change and we will say how we would help.