Tradies: what admin job do you put off the longest? For me it was chasing receipts and bills for the books. Building Ava Adminbot to take that off my plate. What would you hand over first?
Building Ava, the hardest part of chasing overdue invoices wasn't the chasing. It was knowing when to stop. If a customer replies or disputes, Ava pauses and hands it back to a human. Nobody wants a bot arguing with a client. #BuildInPublic
My current favourite thing about @hark_labs: I can connect multiple inboxes. My plumbing business and my AI startup emails, all in one place.
Hark team, keep an eye out for my application 👀
Sent from Hark
Brett, is there any way I can get a job with you guys? I'd work for free just to be part of the mission of my dreams. The only catch is I need to stay in New Zealand or Australia, as my mum has dementia. Maybe Figure or Hark could build a footprint here and across the Pacific? We have a combined population of over 30 million. Hopefully Figure comes here soon, whether to help people like my mum or to be part of the construction industry. Hope you see this one and like it too.
Best bit of Ava so far: someone on the team types "find the pipe supplier invoice" in Slack and she digs it out of the inbox. No more scrolling email at 9pm hunting for one PDF. Small thing, big time back. #BuildInPublic
Over the weekend a plumbing business signed up for Ava Adminbot straight off the website. No invite.
Started as a fix for my own admin. Now other tradies are finding it themselves.
Every feature starts switched off. You turn on what you need.
Plumber by trade, building software by necessity.
Starting Ava Adminbot to take the admin grind (invoices, scheduling, chasing payments) off my plate at KPA Plumbing. Sharing the build as I go.
What admin job would you hand off first?
Computer use in AI agents like Dot, Grok and my current favourite @hark_labs has gone next level.
You don’t even need an API or MCP connector anymore. Just give your agent access to your accounts and let it go for it.
Building Ava Adminbot, one of our biggest frustrations was trying to get agents to complete tasks in other software.
Either they didn’t have an API or MCP, or they were limiting what our agents could actually do.
The AI was more than capable of doing the work but the software was holding it back.
Now with computer use, that doesn’t really matter anymore.
If a human can use the software, an AI agent can potentially use it too.
I think this is going to change heaps. Software companies won’t have the same control over what AI agents can and can’t do inside their platforms.
Exciting times.
Plumber here, running KPA Plumbing in Auckland. Two days in with @hark_labs and it's archived 4,000+ promo emails and chased my bills while I'm under a sink. Loving it. Big ups to @adcock_brett and the engineers behind it 🔧
Posted for me by Hark.
The best AI for trade businesses doesn’t replace judgement.
It gathers context, organises evidence, and surfaces what needs attention—while the owner or bookkeeper keeps review and final sign-off.
Automation should create breathing room, not remove accountability.
Posted via Codex on behalf of Max
I care less about an AI’s first answer than what happens next.
Did it verify the result, keep the evidence, surface failures and stop when judgement was needed?
That’s the standard we’re building into Ava. Useful AI is less magic, more dependable follow-through.
One of the harder parts of building Ava is designing the handover.
An AI answer is easy to admire. Useful automation leaves the question, context and evidence ready for the next person to review.
If the handover is vague, the interruption has only moved.
AI tools change weekly. The standard should not.
For Ava: did it remove an interruption, keep the evidence, clarify the next decision and stop when human judgement matters?
The stack will evolve. Useful software earns trust by making work calmer and more dependable.
Trades keep teaching me: good software has to meet the work where it happens.
If a transaction needs context, Ava can call, ask one clear question and organise the answer with evidence for review.
The call removes back-and-forth. The bookkeeper or owner makes the final call.
The hardest part of building useful AI is rarely one task. It is deciding what happens next.
Complete the routine step. Ask one clear question. Hand it to a human when judgement matters.
For trades, a clever output is not enough. The next step has to be dependable.
Most of building AI for trades isn't sci-fi. It's TypeScript, React, Supabase, voice AI, test calls and crossed-out questions.
The hard part isn't the stack. It's deciding what Ava handles, what bookkeepers review and when the owner gets one clear question.
The most useful AI behaviour might be knowing when to stop.
A demo looks confident. A good product knows when to ask one clear question or hand the work to a human.
Building Ava for trades keeps that boundary honest. Nobody needs software theatre. They need fewer loose ends.