Scope drift
The symptom. The agent starts doing work it was not designed for.
How we handle it. Scope is documented in the Workshop proposal, enforced in the prompt and in code. The agent refuses out-of-scope inputs and logs the refusal.
Use case · lead
Half your team's day is form-to-form work.
Engineers re-type calculation outputs into report templates. Receptionists copy intake forms into the practice management system, and bookkeepers reconcile supplier invoices against PO numbers by hand. The team is competent and senior; the work is not.
The pain
30 to 60 percent of admin and back-office time in Australian SMEs goes to moving the same information from one form, system, or document to another. The owner cannot justify hiring more people for this work and cannot afford the team to keep doing it.
The mistake most businesses make
Reaching for general-purpose AI (Copilot, ChatGPT) and asking it to "help with the admin", which produces drafts that still need full review. The actual answer is a scoped agent that does one workflow end-to-end with explicit guardrails, an evaluation harness, and an escalation path for the cases it cannot handle. That is a build, not a chat session.
What Core Nova does differently
We have done this work. In an engineering consultancy. In a medical practice. Calculation-to-report and intake-to-note workflows where senior staff were spending hours on transcription that should not have been theirs to do. We will not name the clients without their consent and we will not publish numbers on the internet, but the work exists and the operators who built it are the same operators who will build yours.
We name the failure modes most AI consultancies do not. Every Build is built to handle all three.
Evaluation rot
The symptom. The agent's outputs degrade as the underlying data shifts and no one notices.
How we handle it. An evaluation set is built with the agent and re-run on a documented cadence with a named owner. Drift is detected, not discovered.
Escalation gap
The symptom. The agent silently handles cases that should have gone to a human.
How we handle it. Explicit escalation gates for any case outside the documented scope, with a named human owner and a defined SLA.
A$1,500 + GST
Workshop: two to four hours, prepared in advance. Paid in full at booking. The proposal is yours whether or not we build it.
Book a free discovery hourWhat you leave the Workshop with
- Process map of the function in scope.
- A written proposal: what the first version will do, how long it will take, and a fixed price for the Build.
- A proposal you keep whether or not we build it.
What happens after
The proposal is yours whether or not we build it. If you go ahead, the Build is delivered at the fixed price in the proposal, and Evolve is there afterwards if you want it monthly.
Frequently asked
Administrative work that moves the same information between forms, systems, and documents. Engineers re-typing calculation outputs into report templates. Receptionists copying intake forms into the practice management system. Bookkeepers reconciling supplier invoices against PO numbers by hand.
Asking ChatGPT or Copilot to 'help with the admin' produces drafts that still need full review. The actual answer is a scoped agent that handles one workflow end to end with explicit guardrails, an evaluation harness, and an escalation path. That is a build, not a chat session.
A written proposal with a process map and a fixed price, from a Workshop of two to four hours at A$1,500 + GST. It is yours whether or not we build it.
The workflows we run inside our own business are listed on the Still Running board at corenova.au/still-running, each with the month it went live and whether it is alive today, including the ones that stopped. Client builds are under way and go on that same board, named, when the client agrees to it. We do not publish unattributed client numbers.