Thursday morning in Ōtaki. A fencer is loading rammer and augers into the ute when the phone goes. Woman up the coast wants a quote for "a fence out the back, nothing fancy". He books her in for Tuesday, writes the address on the back of a delivery docket, and gets on with his day.
Tuesday comes. Forty minutes up State Highway 1, down a metalled driveway, and the back boundary turns out to drop about a metre and a half across its length. There's a macrocarpa sitting right where four posts want to go. The customer is thinking aluminium slat, not paling, which is a different quote entirely. And the neighbour, whose half of this she assumed was already sorted, has not been asked.
None of that is a disaster. He'll still quote it, and he'll probably win it. The visit just cost him two and a half hours instead of forty minutes, and the version of the job he had in his head on Thursday bears no resemblance to the one in front of him. Every fencing contractor in the country knows this drive.
We build intake agents that do the asking before anyone gets in the ute. Voice on the phone, text over SMS, a photo request that comes back with four pictures and a dropped pin, all of it structured into something you can price against. Below I've set out what changes the price of a fence, what an AI agent can genuinely capture over the phone and what it can't, how photos and boundary length work in practice, where the Fencing Act 1978 and the Privacy Act 2020 sit in all this, and what a rollout looks like for a two or three crew operation.
What actually goes wrong between the phone call and the quote?
The gap is information. A fencing enquiry arrives with about four useful facts in it and needs about twenty before it can be priced with any confidence. The site visit exists to close that gap, so every visit that starts from scratch is doing work a five minute conversation could have done first.
Most trades have a narrow band of variables. A sparky quoting a new powerpoint knows roughly what that involves before they arrive. Fencing sits at the other end. The same fifty metre run can be a $9,000 job or a $30,000 job depending on ground, access, height, material and whether anything needs to hold back dirt.
That variability is why fencers quote on site and why the drive is unavoidable for a lot of work. The question worth asking is not whether to visit. It's how much you know before you do.
Why does a fencing quote need more information than most trade jobs?
Because the price per metre swings by a factor of ten across ordinary residential work. Rural post-and-wire runs from roughly $45 to $140 a lineal metre installed. Standard timber paling sits around $180 to $350. Aluminium slat and custom metal push $450 to $900 and up, and glass pool fencing starts around $550 and passes $1,200.
Those bands come from published NZ pricing guides and they move around by region and by year, so treat them as shape rather than gospel. The point is the spread. A customer saying "about thirty metres" has told you the job is somewhere between $1,350 and $36,000.
Then there are the extras that never come up on the phone unprompted. Removing an existing fence adds roughly $30 to $100 a metre. A pedestrian gate is $800 to $1,800 in timber, more in aluminium. A driveway gate starts around $3,000 and an automated one can run past $20,000 on its own.
None of that is exotic. It's just a lot of questions to ask someone who rang up while making the kids' lunches.
What does a wasted site visit really cost a fencing crew?
Call it two and a half hours of a principal's time plus the vehicle. At a $95 an hour charge-out equivalent that's around $240 of capacity, and eighty kilometres of ute at roughly $1.10 a kilometre once you count fuel, tyres and wear adds another $90. Before the quote is even written.
So a visit that goes nowhere is a $330-ish afternoon, and that's the polite version. It ignores the quoting time on Thursday night, and it ignores the job you didn't get to because you were up the coast looking at a boundary the neighbour hasn't agreed to.
Run three of those a fortnight and you're giving away most of a working week a year. Each individual drive feels like part of the job, which is why the aggregate is the thing that bites. It's the same quiet arithmetic behind what missed calls cost NZ trades, just measured in kilometres instead of rings.
What information does an AI intake agent need to capture before anyone drives out?
Roughly a dozen fields, gathered conversationally rather than as an interrogation. Boundary length and how it was measured, fence height, material preference, ground slope, surface underfoot, machine access, existing fence and its condition, gates, whether the line is a shared boundary, and the customer's timeframe and budget shape.
The trick is sequencing. A person will answer eight questions on the phone happily if the questions feel like a conversation and each one follows from the last. Fire the same eight as a form and half of them get skipped.
This is where language models earn their place. They convert a rambling answer like "it's the bit down the side by the driveway, maybe as far as the shed" into a structured field with a confidence flag on it, which is the whole reason this is possible now in a way it wasn't five years ago.
Which questions actually change the price of a fence?
Six, in rough order of impact. Material and height set the base rate. Total length sets the multiplier. Slope, ground surface and machine access set the labour loading. Gates and demolition are line items on top. Everything else refines a number you already roughly have.
Slope is the one that gets missed most. A moderate fall across a thirty metre run can add over a thousand dollars in stepping, longer posts and extra concrete, and it's invisible in every phone call ever made. Ask "is the ground flat along that line, or does it fall away?" and most people will tell you accurately.
Ground surface is the second. Sand at Waikanae Beach is a different day's work to Horowhenua clay, which is different again to anything with rock in it. A single question about whether they've dug in the yard before, and what came up, is worth more than it sounds.
Access is third and easiest. Can a machine get down the side, or is everything going in by hand? That one answer can move a quote by twenty percent.
How do you capture terrain and access without being there?
You ask, then you verify with photos. The agent collects the customer's description in the call, then sends a text with four specific photo requests: along the fence line from each end, the side access, and anything currently there. Photos come back to the same thread and get attached to the job.
Specific beats general every time. "Send us a photo of the fence line" produces a picture of a lawn. "Stand at the far corner and take one looking back along where the fence goes" produces something you can read slope off.
We build these with a vision model reading each photo as it arrives, tagging obvious features like existing fence type, ground cover, visible slope, and whether there's a retaining situation developing. That's a filter, telling the estimator which jobs need eyes on them this week and which can be quoted from the desk.
And when the photos show something the model can't call confidently, it says so. A flag reading "possible retaining requirement, low confidence" is more useful than a clean answer that's wrong.
How does an AI agent handle photos, maps and boundary length?
It combines three sources and scores the agreement between them. The customer's own measurement, a satellite or parcel boundary estimate from the address, and what the photos suggest. When those three broadly line up, you can quote from the desk. When they don't, the job routes to a site visit.
That routing decision is the actual product. Everything upstream is data collection, and data collection is only worth doing if something downstream changes because of it.
For a good share of straightforward residential work the three sources agree closely enough to send a priced quote the same day.
Can a customer measure their own boundary accurately enough to quote from?
Often, yes, if you tell them how. A pace count is accurate to within about ten percent for most adults, which is enough to bracket a quote. A tape or a phone measuring app gets you closer. The agent asks which method they used and records it, so the estimator knows how much to trust the number.
The failure mode is silent inaccuracy. Someone says "about forty metres" and nobody knows whether that's a measured forty or a guess, so it gets treated as gospel and the quote goes out wrong.
Asking one follow-up fixes it. "Did you measure that or is it an estimate?" costs three seconds and changes how the whole quote is handled. It's the sort of question a good estimator asks instinctively and an intake form never does.
Parcel boundaries from the address give you a cross-check. Council geographic information system layers and Land Information New Zealand data can put a rough length on a boundary run, and if the customer's forty metres comes back as a fifty-eight metre parcel edge, that discrepancy is worth a phone call before it's worth a drive.
What can you tell from four photos and a dropped pin?
More than most people expect and less than a site visit. From four well-framed photos you can usually identify existing fence type and condition, ground cover, rough slope direction, obvious access constraints, and whether there's a level change that implies retaining. You cannot reliably read soil, services, or the exact boundary line.
That's the honest split, and it's worth being clear about with customers. The agent frames the photo request as "helps us give you a better indication before we come out", not as a replacement for the visit. Nobody feels sold a shortcut.
A dropped pin adds the site's actual position, which matters more than it sounds on the coast. A job in Te Horo and a job in Paekākāriki are twenty minutes apart and can be run in the same afternoon. Two jobs booked without location awareness get spread across two days.
Route clustering falls out of this almost for free. Once every enquiry has a pin on it, the week's site visits can be grouped by geography instead of by the order the calls came in.
Where do the Fencing Act 1978 and the Privacy Act 2020 fit in?
Both sit at the intake stage, which is exactly where an automated agent can help. The Fencing Act governs who pays for a boundary fence and how neighbours are notified. The Privacy Act governs how you collect and hold the customer's details and photos of their property. Neither is complicated, and both are easier to get right when the questions are asked the same way every time.
Contractors aren't lawyers and shouldn't be giving legal advice. What the agent does is flag the question early and hand over a plain-English explainer, so the customer knows there's a conversation to have with the neighbour before the quote turns into a booking.
That's a customer service improvement dressed as compliance. Jobs that stall three weeks in because a neighbour objected are the ones nobody enjoys.
Should the AI agent ask who owns the boundary?
Yes, and early. One question, phrased plainly: "Is this fence on a boundary you share with a neighbour, or is it inside your own property?" The answer determines whether the Fencing Act applies, whether costs are shared, and whether a fencing notice needs to go out before work starts.
Under the Act, a boundary fence between two occupiers is generally a half-share proposition. You serve a fencing notice on the neighbour setting out the boundary, the proposed fence, the cost and the timing, and they have 21 days to respond with a cross-notice. No response inside 21 days and they're deemed to have agreed.
There's a standard in there too. Neighbours share the cost of an "adequate fence", meaning one reasonably satisfactory for its purpose. If your customer wants powder-coated aluminium slat where paling would do the job, the difference is usually theirs to carry.
Fencing Act disputes can go to the Disputes Tribunal up to $60,000, which covers nearly all residential fencing work. Worth knowing, less worth relying on. A quote that mentions the notice process up front prevents most of the arguments that would otherwise end up there.
What are the privacy obligations when you're collecting photos of someone's property?
Standard Privacy Act 2020 collection principles apply. Collect only what you need for the quote, tell the person why you're collecting it, keep it secure, and don't hold it longer than you need to. Photos of a property, tied to a name, address and phone number, are personal information.
Three practical things matter. First, the agent states the purpose when it asks: "so we can size the job before we come out". Second, the photos live in the job record inside your job management system, not in someone's personal phone gallery. Third, there's a retention rule, so quotes that don't convert age out after a set period rather than sitting on a server forever.
There's a wrinkle specific to fencing. A photo taken along a boundary line very often includes the neighbour's yard, sometimes their windows. That's not automatically a breach, but the agent's photo instructions should say "frame along the fence line, no need to capture next door", and the retention policy should be short.
None of this needs a compliance project. Ask the questions consistently, set the storage up properly once, and it's handled. That's the sort of thing built in at the start or bolted on painfully later.
How do you put this in front of a fencing business without wrecking how they already work?
You start with overflow. The agent picks up the calls that were going to voicemail anyway, the ones at 4:50pm when the crew is loading out and the ones on a Saturday. Nothing about the existing process changes on day one. The team just stops losing the enquiries they were never going to get to.
Most fencing outfits I've worked with already run a job management system. Fergus, Tradify and NextMinute all show up regularly, feeding into Xero for invoicing. The intake agent writes into that, so quotes appear where they've always appeared. It's the same shape as the rest of our work with trades and SMEs.
Embedded, not bolted on. If the crew have to check a second inbox, it's failed.
What does a two-week rollout look like?
Week one is capture and configuration. We sit with the estimator, work through how they actually price, and build the question set from their logic rather than a generic template. Week two is a live pilot on after-hours calls only, with every transcript reviewed daily and the question set tuned as it goes.
The estimator session is the part that matters. Every fencing business has its own rules about what triggers a site visit, and those rules live in one person's head. Getting them out is most of the build.
Technically it's not a heavy lift. A phone number, a voice agent, a photo collection flow over text, and a write-back into the job management system. We use Vapi and Twilio for the call layer and Supabase to hold the job records, and the whole thing runs from a number that forwards on no-answer.
By the end of week two you have a fortnight of transcripts, which is the first time anyone can see what customers are actually ringing about. That data tends to be more interesting than the automation.
How do you know if it's earning its keep?
Four numbers, measured against the fortnight before you started. Enquiries captured after hours. Percentage of quotes issued without a site visit. Average days from enquiry to quote sent. And quote conversion rate, which should hold steady or improve, never drop.
If quotes are going out faster and converting at the same rate, the automation is working. If conversion drops, the desk quotes are missing something and the routing threshold needs tightening. That's a tuning problem, and it shows up inside the first month.
The number I'd watch hardest is enquiries captured after hours. It's the cleanest one, because those calls were previously going to voicemail and voicemail-to-callback conversion in the trades is grim.
For a business doing forty enquiries a month, recovering even six after-hours calls that would otherwise have gone elsewhere pays for the whole thing several times over on a single won job. This is the same logic behind CallCover, which we built for exactly this pattern in NZ trades.
Practical Takeaway
Start by writing down the questions your best estimator asks before agreeing to a site visit. That list, in their words, is the specification. Everything after it is plumbing, and the plumbing is the easy part.
Then split your enquiries into two piles for a fortnight. Jobs you could have quoted from a good description and four photos, and jobs that genuinely needed you on site. Most fencing businesses find the first pile is bigger than they expected, and that ratio tells you whether this is worth doing at all before you spend a dollar.
If the split is meaningfully in favour of desk-quotable work, the build is straightforward. Voice agent on overflow, structured photo request over text, write-back into whatever job system you already run. Two weeks, live on after-hours calls first, tuned from real transcripts rather than assumptions.
If the split says nearly everything needs a visit, you've still gained something. Use the same question set to make the visits smarter, so you arrive knowing the material, the slope, and whether the neighbour is on board. Same information, different destination.
Frequently asked questions
- Can an AI agent quote a fence without a site visit?
For straightforward residential runs, often yes. When the customer's measurement, the parcel boundary estimate and the photos all agree, a priced quote can go out the same day. Anything showing slope, retaining, difficult access or a disputed boundary should still route to a site visit. The agent's job is deciding which pile a job goes in.
- How accurate are customer-supplied fence measurements in New Zealand?
A pace count is typically within ten percent, which is enough to bracket a quote but not to order materials. Tape or phone app measurements get closer. The important thing is recording which method was used, so the estimator knows how much confidence to put on the number rather than treating a guess as a survey.
- Does an AI intake agent need to know about the Fencing Act 1978?
It needs to ask one question: is this a shared boundary. That single answer determines whether cost sharing and the fencing notice process apply, including the 21 day cross-notice window. The agent flags it and hands over a plain-English explainer. Contractors shouldn't be giving legal advice, and a good agent doesn't either.
- What happens to customer photos under the Privacy Act 2020?
They're personal information and need to be treated that way. Collect only what the quote requires, state the purpose when asking, store them in the job record rather than a personal phone, and set a retention period so unconverted quotes age out. Photo instructions should also steer framing away from the neighbour's property.
If you run a fencing business on the Kāpiti Coast or anywhere else in New Zealand and this sounds like your Tuesday, get in touch. We can usually tell inside one conversation whether the numbers stack up for your operation, and we'll say so if they don't.