A plumber with four vans — call him Steve, because that's a name plumbers actually have — gets 14 enquiries a week through his website form, his Google listing, and WhatsApp. He quotes nine of them. He wins two. He's not bad at quoting. He's bad at knowing, from the first message, which nine were worth his time.
Here's one he quoted last month: "Hi do you do boiler repairs? Not urgent, just want a price for when convenient." He spent forty minutes on that quote — checked his diary, called the customer back, talked through options, sent a written estimate. Never heard back. Compare that to the message he almost skipped: "leaking tap, need someone today, will pay whatever, three properties on this street all mine." That one turned into a same-day job and two more bookings the following week.
The cost here isn't the two jobs he lost. It's the hours spent on the seven that were never going anywhere — hours he could have spent on the jobs that were. A spam filter can't catch this, because none of these messages look like spam. A keyword rule can't catch it either, because "not urgent" and "whatever it costs" aren't keywords you can build a filter around reliably — they're tone, and tone only means something next to the rest of the message and what you know about the sender.
What does 'a lead worth taking' actually mean for your business?
Before anything gets built, you have to answer a question nobody's ever asked Steve directly: what does a good lead actually look like, for him, specifically?
Not in general. Not "profitable jobs." For Steve it might be: anything within eight miles, anything over £150, no first-time enquiries asking for a price before a description of the fault, and anything from a landlord he's worked for before goes straight to the top regardless of size. For a landscaper it might look completely different — season matters more than price, and a scruffy garden with a big budget beats a tidy one with a small one.
This isn't something an agent — a system that reads the situation and decides what to do, rather than just following a fixed script — can work out on its own. It has to come from you, written down. A useful exercise: list the last five leads you regretted taking, and the last five you were glad you took. Not from memory as a vague feeling — actually write the messages down, or as close as you can recall them, next to what happened after. Patterns show up fast. Maybe every regretted job came from someone who haggled before you'd even quoted. Maybe every good one mentioned a specific brand of appliance, because that's the make you're fastest at.
That list is the thing the agent reasons over. Without it, there's nothing to build — you'd just be asking a computer to guess at taste you haven't articulated yourself.
Where the obvious rule declines the job you'd have wanted
Say Steve tried the simple version instead: reject anything under £200 automatically. Clean rule, easy to set up, no judgment required.
Except one Tuesday a £150 tap job comes in from three doors down from a £4,000 bathroom refit he's already got booked for that afternoon. Under the rule, that lead gets declined — it's under threshold. An agent that can see both the message and Steve's calendar for the day treats it differently: same street, same day, marginal extra time. It doesn't book the job itself, but it flags it back to Steve: "This one's below your usual minimum, but it's on the same road as the Whitfield job at 2pm — might be worth a quick add-on." That's not a rule finding a keyword. That's reading two separate pieces of context — a price and a route — that only mean something together.
Same problem the other direction. A message comes in: "Budget's tight, I know it's a small job, but wondering if you can help — I've got a few rental properties and this is the first one that's needed anything." A price threshold sees "small job, tight budget" and declines it. A system reading the actual wording sees "a few rental properties" and treats this as the start of a relationship, not a one-off. It doesn't know for certain it'll pay off — but it flags the signal instead of burying it.
A fixed number can't hold either of those ideas at once. It has no way to notice the calendar or the sentence that hints at six more jobs behind this one. That's the whole case for something that reads the specific enquiry in front of it rather than checking it against a line.
What the agent should decide, and where you stay in the loop
The judgment doesn't have to sit in one place. It should sit in three, sorted by how much damage a wrong call does.
| Tier | What happens | Example |
|---|---|---|
| Auto-send | Obvious no-fits get a polite decline; obvious strong fits get an acknowledgement and a booking link | "Thanks for reaching out — this one's outside the area we currently cover, but I'd recommend [local trade body] for a recommendation." |
| Draft for you | The in-between leads are sorted, with the agent's reasoning attached, ready to send or bin in one tap | "This one's borderline — £180, four miles out, but they mention a boiler service contract with three other flats. Suggest booking a call." |
| Never automatic | Any actual price, date, or commitment | Nothing gets promised until you've said so |
The third row matters more than it looks. The agent can read a message, cross-reference your calendar, and draft a response — but a specific price or a booked slot is something the other person will treat as binding. That's your call, every time.
Worth noticing which direction carries more risk here. In most workflows, sending something wrong is the scary outcome. In lead triage, declining wrongly is worse — because a bad send just gets ignored, but a wrong decline is a customer you never hear from again. You won't get a complaint. You'll just quietly lose them to whoever they call next. That asymmetry is why the middle tier — draft, don't auto-decide — should be wider here than you'd set it for, say, chasing an overdue invoice.
How this goes wrong: the failure modes to design against
The most common failure is over-confidence on declines. A terse message — "quote for a fence, 20m, ASAP" — reads as low-effort to a system trained to value detail. It might actually be a busy tradesperson himself, buying from you wholesale. Terse isn't the same as low-value.
Related: a bias toward polite, well-punctuated enquiries. Plenty of your best customers write like they're in a hurry because they are. Don't let "sounds professional" stand in for "is a good fit" — that's a bias worth checking for explicitly, not assuming away.
Then there's staleness. If you've just taken on a second van, or a second crew, your capacity has changed and the agent's picture of you hasn't — unless you tell it. It'll keep declining jobs you can now easily take, because nobody updated the one thing that changed.
And the seasonal shift: a landscaper's £120 hedge trim is a decline in July, when the diary's full of bigger jobs, and a decent filler in December, when it's quiet. The criteria you wrote down aren't fixed. They need revisiting a few times a year, not set once and left.
What it takes to run — and whether you're ready yet
None of this works if your enquiries are scattered — some in a personal WhatsApp, some in a paper notebook by the phone, some in a website form nobody checks on weekends. The agent can only read what actually reaches a place it can see. If that's your situation, the real first project is getting your enquiries into one place the agent can read, and it's worth doing on its own merits even before any triage gets built on top of it.
You also need the fit list from earlier actually written down, not just felt. Most owners haven't done this, because they've never had to explain their own judgment to anyone before — it's lived in their head and worked fine there. That's not a criticism. It's just the real prerequisite, and skipping it is the most common reason these builds disappoint.
If your leads come almost entirely through word of mouth, straight to your mobile, at a volume you can genuinely keep in your head — five or six a week, say — a simple rule or just your own judgment may serve you better than any build. The agent earns its keep on volume and variance. Below a certain threshold of either, it's solving a problem you don't quite have yet.
How would you know it's actually working?
Track your quote-to-win rate before and after. The number you want to see move isn't more leads handled — it's fewer quotes sent for the same number of jobs won. If you're quoting nine and winning two now, success looks like quoting five and still winning two.
For the first fortnight, spot-check every auto-decline. Read the message yourself and ask whether you'd have said yes. If you disagree with more than a handful, the criteria need adjusting before you trust it further — that's the calibration step, not a sign it's broken.
The number that actually matters at the end of it isn't leads processed. It's hours back in your week, and quotes you're not writing at ten at night for jobs that were never going to happen. If you get that far, the same instinct — sorting the promising from the not-yet, without a fixed rule doing the deciding — is what carries into following up on the quotes that go quiet, which is the next place this same judgment pays off.
What to do next: write the two lists — five regretted leads, five good ones — before you talk to anyone about building this. If you want to understand the distinction this whole article rests on, read what actually makes this an agent and not a filter. And when you're ready to work through what the agent should decide, draft, and never touch for your specific business, that's exactly what it means to map out these decisions before you build.