Spending this week going deeper on the wildlife AI system instead of citing the topline number again. The part I keep coming back to is the confidence threshold: every image gets a classification and a score, and anything below the threshold gets routed to a person instead of logged as fact. Setting that threshold is the part nobody talks about. Get it wrong in one direction and you're back to reviewing almost everything by hand, which defeats the point of building the system at all. Nobody hands you a formula for the right number. It comes down to a judgment call: how much risk of a missed or wrong classification is acceptable before a human needs to look. I don't think there's a universally correct number here. What matters is picking one that's honest about how wrong the model still gets things. Curious how other people building anything with a human-review step think about where to set that line.

I went back through everything I wrote in September and tried to find the thread connecting it, because at the time each piece felt like its own separate rabbit hole. Turns out the thread was simple: check the claim before repeating it, then go find what's already running for real. Week one was about email drafting. I kept seeing the same pitch everywhere, that letting AI draft your replies is basically free productivity. So I went looking for the research behind it instead of the marketing version. One clinical documentation study found unreviewed AI drafts were wrong 42% of the time. That number changed how I build now. The rule I run on is plain: AI drafts, a human sends. No exceptions, no matter how good the model gets. Week two moved off claims and onto client work I can point to directly. Workflow automation projects have been clearing 8 to 20 hours of manual admin a week for the people running them, worth somewhere between £20,000 and £57,000 a year in recovered value, a range that keeps repeating across different engagements rather than one standout client. The number that surprised me most wasn't the money, it was how consistently that range showed up across businesses doing nothing alike. Separate from that, a wildlife conservation project cut manual camera-trap image review by more than 90%, putting over 1,200 specialist hours a year back into people who'd rather be doing conservation work than sorting photos. And on the outreach side, an AI business development pipeline produced 140+ personalised emails across two niches in under two weeks, the kind of volume that used to mean hiring someone just for that. Week three tried to put a shape around all of it, because "AI helps businesses" is too vague to check against anything. I landed on a five-stage maturity model: Stage 0, manual everything. Stage 1, AI bolted on with no integration. Stage 2, AI inside real workflows with a human still reviewing. Stage 3, agents that each own a recurring job with human approval at the decisions that matter. Stage 4, a self-managing operating layer. Most small businesses I talk to sit at Stage 0 or Stage 1, which tells you where the real starting line is before anyone promises Stage 4 results.