Picture a kitchen with the best equipment money can buy. Sharp knives, a walk-in fridge, ovens that hold their heat perfectly. The cooks are skilled and the ingredients are fresh. And yet plates are barely leaving the pass, the dining room is half-empty, and orders are piling up on the rail.
You walk in expecting to find a broken oven. You don't. The equipment is flawless. The problem is that every single dish, before it can go out, has to be tasted and signed off by one head chef — and that head chef can only taste so fast. The kitchen isn't slow because it's badly equipped. It's slow because everything funnels through one person.
That kitchen was us.
We went looking for a missing tool
We build intelligent systems for a ministry called HOFMI — its education arm, its missions work, the Bible school, the church. By the middle of 2026 the setup underneath all of it was genuinely good. It could draft, research, and get real work ready to go.
So we asked for an honest review of how it was actually performing, and we braced for the usual answer: some missing piece, some tool that still needed building. That is almost always what these reviews turn up.
That is not what this one found.
The review was built to disagree with us
We ran it deliberately as a check, not a celebration. One group inspected how things really ran, day to day. Another went off and researched independently. A third group had one job only — to argue, to fact-check every claim the first two made and try to knock holes in it. A final pair sat above all of them as judges. Nobody was allowed to change anything; the whole point was to see the system as it truly was, not as our own notes claimed it was.
The verdict came back in a single sentence: the machine is excellent, and it is pointed at the wrong target.
The number nobody thinks to look at
When something feels slow, the natural instinct is to ask how much got done and how fast. Fair enough — but that only tells you about the work that actually moved. It says nothing about the work stuck in the queue, waiting.
The number that gave the game away was a different one: of all the decisions that needed a yes, how many were signed off by a person, and how many the system was trusted to decide on its own. Over two months, nearly every decision of any weight came back to one desk — mine — and almost none were left to the system. It could do the work. It just wasn't allowed to say yes to anything. Everything queued behind one person's attention.
Two other signs pointed the same way. New work was coming in far faster than finished work was going out. And the actual ministry output the whole thing existed to serve had dried up — a key pipeline had been dead since March, and the spaces set up for new trainees sat empty.
A brilliant setup with one person approving everything isn't an automated system. It's an expensive way to turn that one person into the bottleneck.
Why a normal catch-up meeting would have missed it
A standard review asks what slowed down the work in front of you. It looks at the jobs that moved. So it never weighs them against the much bigger pile that never even got a decision. You can hold a perfectly tidy catch-up every fortnight, conclude the team is doing well, and completely miss the real jam — the queue waiting on one approver — because that queue never comes up in the conversation.
Counting approvals against finished work drags the problem into the open in a single figure. When all the sign-offs cluster on one name, and new work keeps outpacing finished work, you are not looking at a tools problem. You are looking at a delegation problem wearing a tools costume.
The bit worth taking away
If you have spent money on clever software and it still feels slow, resist the urge to buy more of it. Check two things first. Who signs off on the work — and how much gets out the door without them. If one person owns nearly every yes, the cheapest improvement available is not another tool. It is moving more of those decisions off that one desk.
That is the harder fix, which is exactly why most people reach for the software instead.
For the technical reader
The platform under HOFMI's operations was, by mid-2026, genuinely capable: agents that could draft, research, provision and deploy.
The audit was deliberately read-only and adversarial, structured as a multi-agent workflow. Several agents ran live probes against production; others ran independent research tracks; a separate set acted as sceptics whose only job was to fact-check every load-bearing claim the first two groups made; and a final pair sat as judges over the lot. Nothing wrote to anything — the goal was to observe the system as it actually ran, not as the change log claimed.
The decisive metric was the ratio of human approvals to agent approvals. Across eight weeks, almost every decision of consequence routed back through one person, with a vanishingly small share decided by the agents themselves. The platform could do the work; it could not authorise it.
Two supporting figures corroborated the diagnosis: intake (new work entering the system) ran well ahead of the ship rate, and ministry throughput — the output the platform existed to serve — was close to zero, with the sermon pipeline dead since March and the trainee workspaces empty.
The transferable instrument is the approval-to-ship ratio: who signs off, and how much the system ships without them. A clean sprint retro is silent on the work that never got a decision, so it systematically hides a single-approver bottleneck. When approvals cluster on one name and ship rate trails intake, the constraint is delegation, not tooling — and that is the harder, software-resistant fix most teams avoid.