Log in
Hivemind Labs · Australian frontier AI lab

We prove, with evidence, when AI can act on its own, and build whatever it takes to get there.

Follow one agent from its first refund to its five hundredth
The problem

Meet a customer service agent. A customer wants a refund.

It reads the message, finds the order, issues the refund and replies.

It's done in 13 seconds, for 12 cents.

Nobody on your team had to touch it.

So you give it sixty more.

Thirteen minutes of work. Seven dollars. Every reply sounds right.

Ten of them were wrong.

Each one sounded exactly as confident as the other fifty.

So now someone checks every single one.

Two hours of checking to supervise thirteen minutes of work. Supervision swallows the value.

Nobody can show when AI is safe to act alone.

So every business putting AI to work is stuck choosing between these two.

Customer
I ordered a desk lamp three weeks ago and it never showed up. I'd like a refund please.
AI agent
No worries, I've refunded $42.00 to your card. It should land in 3 to 5 days. Let us know if there's anything else you need.
13.0s
agent time
$0.120
cost
0s
your time
0
refunds
0m
agent time
$0.00
cost
Refunded on day 45. Your policy allows 30.
Refunded someone else's order.
Refunded $189.00 on an $18.90 order.
Agent doing the work13 min
Your team checking it2 h

60 refunds × 2 minutes of review each. Slower than doing it yourself.

Let it run

Fast and risky

  • 13 seconds a refund
  • No staff time
  • 1 in 6 wrong refunds reach customers
Check everything

Safe and slow

  • Every refund reviewed
  • 2 hours of staff time per 60
  • The agent saves you nothing

Earned Autonomy

Autonomy grows as uncertainty shrinks. Supervision falls only as evidence clears the bar, task by task.

Earned Autonomy

First, agree the error bar.

Would automating refunds still pay off even if 1 in 50 went wrong? If yes, that's the most risk you'll accept. It isn't what you expect to happen.

Supervised: every action reviewed.

Already faster than doing it by hand, but a person is still the bottleneck.

Every correction teaches it.

Fewer get sent back, and the track record builds.

Spot-checked: the multiple climbs.

The evidence clears the bar, so hidden spot checks replace full review. It's judged on its worst plausible error rate, not its average.

Independent. Then it levels off.

Spot checks and critical actions keep a named person in the loop, by design. If errors rise, the task steps back down.

Standing still isn't free.

Check everything forever and the multiple stays at 2×. Let it run unchecked and 1 in 6 refunds go wrong. Earned Autonomy is how the multiple grows safely.

The error bar

Would automating refunds still pay off even if 1 in 50 went wrong?

If yes, 1 in 50 is your error bar. The agent has to prove it stays inside it before it works alone.

1.6×
The multiple: work done per hour of your team's time, compared with doing it by hand
Supervised
Refunds handled →Placeholder curve
27 hrs
of extra checking every month if you check everything forever
1 in 6
refunds wrong if you let it run unchecked instead

Assumes 1,000 refunds a month at 4 minutes each by hand.

When nothing fits, we build it

No API? We trained a model for that.

This store's portal has no API. Scout, the browser model we built when nothing on the market fit, works through it like a person would, with every step checked.

0s<1¢ per taskEarns autonomy like any task

And you see everything it does.

Every action comes with an explanation, and every decision and review lands in the audit record. Critical or irreversible actions always go to a named person.

admin.yourstore.example/orders
Scout · ready
Orders
Search orders
Search
OrderItemTotalStatus
#4468Ceramic mug$18.90Delivered
#4471Desk lamp$42.00In transit
#4474Headphones$129.00Delivered
#4479Notebook set$24.50Packing
38
conversations today
112
actions on its own
0
waiting on you
Whatever it takes

And everything else you'd expect.

Model-neutralLeading models, mainly Anthropic Claude. Move models without a rebuild.
IntegrationsAPIs, webhooks, email, SMS, Teams, Outlook, Slack, CRM and finance.
Named ownersEvery agent has an owner. Nobody approves their own agent's promotion.
Audit recordEvery decision and review, tamper-evident.
Single sign-onYour existing login.
Sovereign by defaultAWS Sydney. Your data stays in Australia.
Case study

In production today.

Regulated customer service: an enquiry agent working
around the clock for a five-person team.

Autonomy

70%

of interactions automated under Earned Autonomy

Volume

10k+

live customer interactions handled

Time back

7.6hrs

of staff time every week, 0.2 FTE across a team of five

Who we are

Lab-grade research.
Enterprise delivery.

Hivemind Labs is an Australian frontier AI lab. We prove, with evidence, when AI can act on its own, and build whatever it takes to get there.

  • Australian-owned
  • All R&D in Australia
  • Data stays in Sydney
  • Model-neutral
  1. 01

    We measure it, task by task

    Earned Autonomy lets supervision fall only as evidence clears the bar. Results are published, failures included.

    83–94%fewer serious mistakes than unchecked agents, in pre-registered experiments
  2. 02

    We build whatever it takes

    Off the shelf where it works, from scratch where it doesn't. When no browser model fit, we trained Scout.

    <1¢per task with Scout, on our own hardware in Sydney
  3. 03

    We partner on one real task

    Pick the task, agree the error bar with your risk team, run it on real work. You own what we build together.

    3–6 wksto a first experiment, live on your work

There's a lot we didn't show you.

How the bar is set, what happens when an agent changes, how Scout was trained. That's what a conversation is for.

Get in touch

Pick one task, agree the error bar, and have the first experiment live in three to six weeks. Partners and investors, start here too.

Emailhello@hivemind.au
Mobile · Jake0401 241 019