The green button problem: why my AI company ships nothing on its own

Build in public — Taigency daily series. 4 August 2026.

My life has run in five-year stints, and until now I never chose a single one. Ten years running a web agency out of Limpopo and Johannesburg — I did the hosting, the CMS and the custom PHP, and we shipped 350 sites. Folded inside that, five years building a CRM for a four-branch financial-advisory firm that quietly outgrew its brief and became the cloud ERP running their whole business. Then COVID sent me home to Joburg and five years of live-event production, learning how ADHD owner-operators actually run things. Every one of those chapters found me. I never planned one.

This one I'm planning. Taigency is the first reinvention I've engineered on purpose — an AI-run company where twelve agents do the real work: sales research, outreach drafting, calendar triage, ops follow-ups, content, finance admin. And the first thing I engineered wasn't the automation. It was the gate.

People assume the scary part of handing a company to AI is handing over the work. It isn't. The scary part is the last inch — the moment something leaves the building with my name on it. So I don't let that inch be automated. Everything my agents produce lands in one place, an approval queue, as a draft marked yellow. Nothing becomes an email, a post, a proposal or a payment until I flip the row to green. The agents run 95% of the job on their own. The final 5% — ship it — stays human, on purpose, every time.

Building this taught me that autonomy and authority are two different levers, and most people conflate them. Autonomy is how much an agent can do without asking. Authority is how far the consequences reach. You want the first dial high and the second one clamped. An agent that researches a prospect, drafts a tailored message and stages it for review has enormous autonomy and almost zero authority — exactly the shape you want. The failures people fear from AI in business are authority failures, not autonomy ones: the tone-deaf email to a key client, the wrong number in a quote, the reply that reads like a robot. So I engineered the authority out and left the autonomy in. It's the same instinct I learned in the bulletin-board days, when a single sysop stood between everything and the board — volume underneath, one accountable human at the switch.

I think this is where the whole internet is heading, too, which is why I'm not mourning it. People keep saying the web is dying. I'd say it's collapsing inward. For thirty years it was a place you went to — you browsed, you clicked, you scrolled through pages other people made. Now it's becoming something you talk to. The pages are collapsing into a conversation, and behind that conversation are agents doing the fetching and the drafting. But a conversation with nobody accountable for what it says is just noise with confidence. The gate is how you keep a human accountable when the interface stops being a place and becomes a voice.

The practical payoff is boring and huge: I move faster because nothing ships on its own. I'm not writing the first draft of anything anymore — I'm reviewing the tenth. My morning isn't "produce," it's "approve, reject or edit." Twenty rows in the queue, most of them good, a few flipped straight to green, two sent back with a note. The business runs at the speed of my judgment instead of the speed of my typing. That's the trade I actually wanted.

If you're putting AI to work in your own operation, don't start by asking what it can do on its own. Start by drawing the line it can't cross without you, and build everything up to that line. Draft everything, ship nothing — until you say so.

I packaged the twelve operators I run Taigency on into a free bundle for Claude — the Agents of a CEO pack. Same agents, same approval-first shape, yours to plug in and adapt. Free download at taigency.co.za; leave an email and it's yours.