Plenty of agencies will theorise about AI agents from the outside. We build and run a growing portfolio of our own AI-agent businesses — real products, with real visitors, real conversion targets, and real money on the line. That changes what you learn.
When it's your own funnel, you can't hide behind a deck. You watch live conversations, you see exactly where people drop, and you feel the difference between a flow that converts and one that quietly loses everyone at step three. Most of what follows we didn't predict — we discovered it the moment real traffic hit something we'd built and were sure was fine.
We bring those scars to client work. Everything we'd deploy for an Australian business, we've usually already shipped, broken, and fixed for ourselves first. The lessons below are the ones that survived contact with actual users.
Our earliest agents made a classic mistake: they asked for too much before giving anything back. Name, email, a few qualifying questions — all up front, all before the visitor had received a single useful thing. Conversion was poor and we initially blamed traffic quality. The traffic was fine. The order was wrong.
The moment we flipped it, the numbers moved. Lead with a useful, specific result the visitor can get in two or three taps, and only then ask for the email to send or save it. People will happily tell you what they need in order to get an answer. They will not fill in a form to earn the privilege of being helped.
The mental model that stuck: every question you ask before delivering value is a withdrawal from an account the visitor hasn't let you open yet. Make a deposit first — answer the thing they came for — and the ask that follows feels fair instead of intrusive.
Across our marketplaces, there's a single highest-intent moment in almost every session, and it's the same one every time: when the agent presents a genuinely relevant match or a scoped recommendation. That's the instant the visitor leans in. It's the natural place — the only good place — to ask for a contact detail.
Ask before the match and you're interrupting someone who hasn't yet decided you're useful; the request feels like a tollgate and they bounce. Ask at the match and the same request reads as the obvious next step: 'Here's the thing you wanted — where should we send it?' Identical ask, completely different result, purely because of timing.
This reframed how we think about lead capture generally, including for clients. Capture isn't a fixed gate you bolt to the top of a funnel. It's a moment you earn, and the agent's job is to reach that moment reliably and then ask. Get the timing right and capture stops feeling like friction at all.
The instinct with conversational agents is to make them impressively open-ended — let people type anything, handle it all, feel magical. On our own traffic, that consistently lost to something far more boring: a tight, button-led flow that always reaches an outcome.
Open-ended chat wanders. It mishears, loops back on itself, and occasionally strands people in a conversational cul-de-sac with no obvious way forward. Every one of those moments is a visitor lost. A predictable flow with clear options gently rails people toward a useful end state, and you can actually see and fix exactly where it leaks because the paths are finite.
It's a humbling lesson for anyone who loves the tech: predictable beats impressive on the only metric that pays the bills. Visitors don't reward you for how clever the agent feels in the abstract. They reward you for getting them somewhere useful without making them work for it.
Running our own products forced us to be honest with instrumentation. Opinions about why a flow converts are cheap; a funnel that shows where people actually drop is the only thing that settles an argument. We measure each step — arrived, engaged, reached a result, gave a contact detail — and we let the drop-offs tell us where the next fix lives.
Two habits matter here. First, change one thing at a time. If you reorder three questions, add a new result screen, and reword the ask all at once, a moved number tells you nothing about which change did it. Second, watch real transcripts, not just charts. The aggregate tells you where people leave; the transcripts tell you why, in their own words.
This is the opposite of launching and hoping. The funnel is the feedback loop, and a tight loop is what lets a mediocre first version become a strong one without guesswork.
None of this stays locked in our own products — it's the playbook we bring to client builds. We design the agent to give before it asks, to capture at the genuine moment of intent rather than at the door, to favour a reliable button-led path over an open-ended chat that can wander, and to be instrumented from day one so it can be improved against evidence.
It's also why we don't just hand over an agent and walk away. The first version is a strong starting point informed by what we've already learned; the real gains come from running it against your actual traffic and tightening the funnel where your customers — not ours — actually hesitate.
The throughline is simple, and it's the same standard we hold our own businesses to: the agent must always reach a useful end state, every time. Everything else — the order of questions, the timing of the ask, the shape of the flow — is in service of that.
One quick brief and we'll point you at the agent that'll pay for itself fastest — and tell you straight if you don't need one yet.
Get a fixed quote →