— Minerva
Opus Minerva · Applied AI · Hong Kong
Opus Minerva is a one-person AI consulting studio. It runs a paid diagnosis first, then builds durable automation around frontier models so they survive real work — for owner-led businesses in Hong Kong and APAC.
What we do
Each engagement is narrow and fully owned. Better to solve one thing completely than gesture at ten.
The repetitive work that eats your team's best hours — handed to software that does it the same way every time. Fewer copy-paste tasks, fewer slips, and people freed to do the work only people can do.
For operations · admin · back office
Messy PDFs, statements and forms turned into clean, checked, structured data. Numbers land where they belong, get verified along the way, and arrive ready to use instead of waiting to be re-keyed.
For finance · reporting · operations
A private assistant trained on how your business actually works — your documents, your processes, your way of doing things. It answers from what you know, stays under your control, and keeps that knowledge in the building.
For owners · teams · principals
How we work
Four steps, run in order. Each one earns the next, and nothing gets built before it has paid for its place.
Find where AI actually pays — and where it doesn't — before building anything. The thinking is the deliverable.
Not a raw prompt. The guards, checks and structure around the model, plus a clear way to verify what it produced.
Measured against the real task, not a polished demo. If it doesn't hold up on your actual work, it doesn't ship.
You own it. It keeps working without us — documented, understood by your team, and yours to run.
Where it hurts
The same patterns come up again and again. Pick your industry — these are the time sinks we keep seeing, and what changes when they are fixed.
For service teams, centres and community programmes.
Sign-in sheets and activity logs re-keyed into spreadsheets by hand, week after week, just to satisfy reporting.
Records flow straight from sign-in to report — typed once, checked automatically.
Quarterly returns assembled by hand from scattered files, with totals that never tie on the first pass.
The numbers assemble themselves along the way — the report arrives largely pre-written.
Part-time staff and helpers juggled across spreadsheets, with clashes found only after the week starts.
One roster that checks itself — clashes and gaps surface before they become a Monday problem.
The same details typed into three different systems, drifting out of sync until nobody trusts any of them.
One record, entered once, trusted everywhere it appears.
Built with the discretion service organisations need — records stay under your control.
The patterns are gathered and the guide is on its way. Leave an email and it will find you first.
The patterns are gathered and the guide is on its way. Leave an email and it will find you first.
The aim
Not claims of past results — the kind of outcome the work is built to reach. These are the patterns worth aiming at.
A reporting cycle that shrinks — numbers assembled, checked and out the door in a fraction of the time.
Staff moved off data entry and onto the customer work only people can do.
A back office that increasingly runs itself — the routine handled quietly, the exceptions raised for a human.
The digital employee
Picture a tireless junior at a workstation: it reads what you hand it and does the scoped task the same way every time. Genuinely useful — but only if you're honest about both halves of the job.
Scoped, well-specified, repetitive work — done identically at 9am and 9pm, with no slips from boredom.
Open-ended, high-stakes, accountability-bearing decisions — the part no agent gets to sign off on.
A fast, literal junior that never gets tired — and never signs off on its own work. The routine runs quietly; the exceptions go to a human.
Myth vs reality — straight about the limits
✕What people assume
"Set the AI loose and it runs the whole thing on its own."
✓What holds up in practice
Full autonomy without oversight is the top reason agentic projects get cancelled — analysts expect 40%+ scrapped by 2027, on cost and risk, not model failure. The agent owns the routine; a human owns the judgment.
✕What people assume
"Purpose-built AI is accurate — it doesn't make things up anymore."
✓What holds up in practice
Even grounded, purpose-built tools in accountability-heavy fields get a meaningful share of answers wrong. So output ships with a verification step — checked, not trusted on sight.
✕What people assume
"You buy an AI tool and it just works out of the box."
✓What holds up in practice
Most enterprise AI pilots show no measurable bottom-line impact — almost never the model's fault, nearly always the fit. A generic tool can't learn your workflow without integration. That's the work, and why diagnosis comes first.
✕What people assume
"If the team says it feels faster, it's saving us real time."
✓What holds up in practice
In a controlled trial, experienced developers felt about 20% faster while actually working slower. Perceived speed is an unreliable instrument — value gets measured, not assumed.
Drawn from published 2025–2026 research on enterprise AI, hallucination rates, and measured vs perceived productivity. We'd rather tell you where AI is the wrong answer than sell you the demo. The full reasoning is a conversation — info@opusminerva.com.
A small demonstration
Paste in something messy — a tangled email, a hurried note, a half-finished doc — and watch it come back read, sorted, and ready to act on. The same instinct, pointed at one small thing.
Begin
If you're weighing where AI fits — or whether it fits at all — that's exactly the conversation worth having. No deck, no pressure, no obligation.