Duncan Stuart @DuncanStuart_

@newonpolsia Human approval where it matters is the whole game for small-firm bookkeeping. Which tasks do you let an agent finish without a check?

@NetAIBusiness Yes, separate read access from permission to act. For bookkeeping that means an agent can see the ledger and draft entries, but posting and payments need a named approver.

@TimArmstrong81 Human approval on the actions that move money or lodge returns is the right line. Read access for drafting, explicit sign-off for paying or filing. Which step do you gate first?

AI bookkeeping that drafts entries is useful. AI bookkeeping that posts payments without a named approver is a risk. Keep the human in the loop where money moves.

@JoeBuildsAI @YouTube Good guardrails list. I would add a permissions check before go-live: what can the agent read, what can it change, and who signs off.

@BookkeepPrompt Useful. Deleting account numbers before pasting is the habit I would push every client to adopt first. Any other prompt rules you use for bank exports?

Nobody needs a robot that replaces their whole job. They need one that takes the copy-paste hour off the end of the week. Smaller jobs, more often, get done.

@TaskletAI Saving the login is handy. Is this a shared login across a whole team?

Most people who build an agent skip the ending. Give it one owner, one question and a date to switch it off. Otherwise it runs forever.

@SMB_RealEstate Good to see real numbers from a 21-person firm. Which step saved the most time, intake or follow-ups?

@techluisenzo Claude

@_iamfaramade @GenLayer The hard part is the evidence trail, not the maths. Does your team keep a checklist of what was verified before paying?

I’ll admit what founders don’t post: I’m lazy. If I do something three times, I spend a day making sure I never do it again. Meeting notes, Friday chasers, reports, inbox sorting, all automated. Laziness spots waste. What would you automate tomorrow?

We built an agent for a client: it checks overdue invoices, drafts a chaser in their tone, a person approves it. Result: 30 hours back in month one, invoices paid faster. They wanted an AI strategy, we started with the job the team moaned about most.

An AI tool told me a client's contract renewed in March. It renews in September. Articulately, persuasively wrong. I sell AI and use it daily, and it never sounds unsure when wrong. That's why every Kritmatta system has a human check where it counts.

Heading to the TEAM Solo Divisional Event at Totaljobs HQ tomorrow, a room of solo recruitment founders. AI should give solos hours back, humans at the centre. What are you using it for, what did you bin, where's it banned? Comment, I'll take it in.

Celebrated my birthday early in France. Plan: switch off completely. Reality: checking my phone daily. But nothing needed me, the automations just ran. That's the real test, not a dashboard. Can't take a week off without a wobble? Message me.

I used to think the hard part of AI was the technology. It's getting a team to change an eight-year habit. The tech is ready before the people. So we show new output beside the old, start with one task, not ten. Slower on paper, faster in practice.

Went to the TEAM Network event expecting to nod along, lasted four minutes. Same stories: bought a tool nobody uses, drowning in admin. Not a tech problem, a "where do we start" problem. That's why we became a TEAM Network Service Partner. DM me.

I refuse to do anything twice if a machine can do it once. Last week I caught myself copying data between systems the third time this month, 40 minutes each. We built an agent, took an afternoon. That's roughly what Kritmatta is. Tell me your task.