Insight · From production
What production teaches: running Digital Workers on remediant.ai
The practice runs its own clinical-intelligence platform in Australian aged care — the environment where Digital Worker patterns ship before any client engagement uses them. These are the lessons that survive contact with production.
The environment
remediant.ai is the practice's own self-funded platform: a clinical-intelligence system for Australian aged care, running on serverless, event-driven architecture with FHIR ingestion and CI/CD quality gates. Two Digital Workers from the practice's roster operate here in production — a Clinical Document Analyst that reads clinical documents and HL7/FHIR feeds and routes structured findings for clinician review, and a Clinical Intelligence Analyst that watches the clinical record for deterioration signals and compliance-relevant events, assembling the evidence trail behind every flag.
Both run behind hard guardrails: human review on every clinical output, flags that never become decisions, every signal traceable to source records. That conservatism is deliberate — aged care is a domain where the cost of a wrong answer is measured in people, and where software sits close to a medical-device boundary regulated by the TGA.
Building slideware teaches you what sounds right. Operating a platform teaches you what breaks. Five lessons follow — each one now baked into how the practice designs Digital Workers for clients.
Lesson 1 — Decision capture is the product
The naïve model of an AI system is that the output is the product and the logging is overhead. Production inverts that. When a clinician, an auditor or a quality manager asks "why did this flag fire?", the answer — this rule, this evidence, this calculation, captured at the moment it happened — is what makes the output usable at all. An unexplainable flag in a clinical setting isn't a partial success; it's noise a busy nurse learns to ignore.
On remediant.ai, decision capture runs as a first-class subsystem: what the worker read, what it concluded, what it escalated, immutably recorded. The practical consequence for any engagement: design the decision record first, and let the worker's architecture fall out of what the record must contain. Retrofitting explainability onto a system that never captured its reasoning is somewhere between expensive and impossible.
Lesson 2 — Golden datasets age
Evaluation evidence is produced before a worker ships: golden datasets encoding what "done well" measures as, red-team suites, regression gates in CI/CD. Production's uncomfortable lesson is that this evidence has a shelf life. Clinical documentation styles drift, referral sources change, resident populations shift — and a worker that scored well against last year's data can degrade quietly against this year's reality.
Two disciplines follow. First, drift monitoring in operation — including demographic drift, because a worker that performs unevenly across population groups is failing even when its aggregate numbers hold. Second, periodic re-evaluation as a scheduled obligation, not a response to incidents: the golden dataset gets refreshed from production reality, and the worker re-earns its score. Evaluation is not a gate you pass once; it is a cadence.
Lesson 3 — Deletion is a feature
Privacy obligations are usually framed as constraints. Production reframes them as engineering requirements with acceptance criteria. Aged-care records carry long retention obligations — remediant.ai maintains a seven-year audit trail — while privacy law simultaneously requires that personal data be deletable, provably, on request.
Holding both at once is a design problem: retention-governed audit trails on one side, and signed deletion certificates on the other — cryptographic attestation that specific data was destroyed, which is itself a record that survives the deletion. None of this is buildable as an afterthought; data minimisation, deletion paths and retention boundaries have to be in the data architecture from the first schema. The lesson for clients: ask your AI vendor to show you a deletion certificate. The pause tells you what you need to know.
Lesson 4 — Escalation design determines trust
The most consequential design decisions in both production workers turned out to be not model choice but escalation calibration: what the worker handles silently, what it flags, and how often. Escalate too little and the safety case collapses. Escalate too much and clinicians drown in flags — and a review queue that people rubber-stamp is a human-in-the-loop control in name only.
The workers hold trust because the boundary is legible: they flag, they never decide, and every flag arrives with its evidence trail attached so the human review is fast and genuine. Escalation thresholds are tuned like a product feature, against real reviewer behaviour, not set once in a design document. A worker earns autonomy the way a junior staff member does — gradually, against a record.
Lesson 5 — Treat the worker like a privileged hire
A Digital Worker reading clinical feeds holds a level of access few humans in the organisation hold. Production makes the implication concrete: the worker is a non-human identity, and it gets the full treatment a privileged human hire would — least-privilege access scoped to its position description, policy-based authorisation, Australian data residency, every data access audited, credentials revocable in one motion.
This is the discipline that turns an uncomfortable executive question — "what can this thing actually reach?" — into a one-page answer. It is also the part most pilots skip, because shared service accounts are easier. They are easier the way unlabelled keys are easier.
Why a consulting practice runs its own platform
Every pattern above — decision capture, evaluation cadence, deletion architecture, escalation tuning, non-human identity — shipped on remediant.ai first, at the practice's own cost and risk, before appearing in any client design. That ordering is the point. Advice about production should come from production; the alternative is a slide deck with confident fonts.
It also keeps the practice honest in a way clients can verify: the core principles on this site aren't aspirations, they are descriptions of a running system. Where an engagement could raise a conflict with the platform, it is disclosed before contracting — that discipline is part of the same posture.
What this means for an engagement
Digital Workers built for clients — through Digital Worker Development, on the platform and governance patterns above — deploy inside your cloud boundary, in your tenant, with these production lessons already designed in. You own and operate the workforce when the engagement ends; the scar tissue comes included.