AI is clever, but on its own it cannot see your business or reach your systems. Three things change that. We build them with you, in this order, and every offering below builds one of them.
Ninety minutes, a fixed fee, a written page. Half the time the answer is: not yet.
Before Anyone Builds Anything.
An owner has been pitched three times this quarter. One quote is $40,000 and eleven pages long, one is a licence reseller, one is a nephew. None of them can be judged from the inside, because judging them requires knowing which parts are genuinely hard. Meanwhile twelve Copilot licences were bought in March and nobody has opened one since April.
The most expensive AI mistake a small or medium-sized business makes is not choosing a bad vendor. It is building the right-looking thing for a job that did not need it.
0190 minutes, fixed fee, a written page within two working days.
02The fee comes off the first build, in full, for three months.
We charge for this, and it is the cheapest thing we sell. A free session is a sales call and both sides know it — the advice bends, quietly, toward the thing the adviser builds. Ninety paid minutes buys an answer we have no reason to steer.
The fee is credited, so the incentive is visible. If you go on to build with us within three months, the advisory fee comes off the first invoice. That means we are not trying to sell you a second session. The cost: you pay before you know whether we are any good.
Half of what we say is “not yet”. A job that happens ten times a month is cheaper done by a person for years. Data spread across three systems with no owner will defeat any agent, regardless of who builds it. Both of those go in the written page, in those words. The cost to us is obvious and we would rather write it down.
Bring the other quote. We will read a competitor’s proposal and mark which lines are real engineering, which are licence resale, and which are margin. Sometimes the answer is that their quote is fair and ours would be higher. That has happened.
One page, not a deck. A forty-slide strategy document is a way of charging for volume. You get one page: the three jobs ranked by hours saved, what each needs in place first, a rough number, and which one to do in the next sixty days. The cost: it does not look like much work. It is.
90 minutes, fixed fee. At your place or online. Bring whoever actually does the work, not just the owner.
Written answer in two working days. One page. Yours, whether or not you ever call us again.
Credited against a build. In full, for three months.
Your numbers stay yours. Nothing recorded, nothing used as a case study without asking.
Monthly advisory afterwards, if useful. Fixed hours a month, ask anything, shares the same retainer pool as upkeep.
For an owner who has been told to do something about AI and has no way to judge what. Fee credited against any build.
Half a day, by role, worked on your own files. Everyone leaves with three workflows.
Teaching the Tools They Already Have Open.
Half your staff are using ChatGPT already. On personal accounts, on their phones, pasting in customer names and quotes, and not telling anybody — because nobody told them whether it was allowed. The other half have read that AI is coming for their job and have decided not to touch it. Neither group has been trained, and both are now a risk of a different kind.
01Half a day, one role group at a time, worked on the business’s own files.
02Everyone leaves with three workflows they run the following week, written down.
03A one-page note to the owner afterwards: the jobs people are doing by hand in a chat window every day, which are the jobs that should be agents.
Most AI training teaches prompting in the abstract. People nod, the day goes well, and within two weeks everybody is back to their old method — because nothing they were shown used a document from their own desk.
We train on your files. The quote template, the supplier email that always takes twenty minutes, the complaint nobody wants to answer. Real ones. The cost: we need them a week ahead, some of them are confidential, and we sign whatever your lawyer wants before we see them.
We teach by role, and roles do not mix. An estimator’s morning and the front desk’s morning have nothing in common, and a session that serves both serves neither. The cost: twenty staff means two or three sessions, not one, and it costs more than a single all-hands.
The first hour is what not to paste. Customer lists, staff files, bank details, anything under the Privacy Act. Most of the room is already doing some of this on a personal account. It is the least interesting hour of the day and we still put it first, because the alternative is finding out later.
The session doubles as a shopping list, and we say so on the day. The tasks people are doing by hand in a chat window, every day, are exactly the tasks that should be agents. We write them down and hand the list to the owner. That is a sales document. We would rather name it than pretend.
Half a day, on site, per role group. Fixed price per session.
Your files, returned, not retained. We work on copies and delete them after.
Three written workflows per person. On paper, in their words, for their job.
A note to the owner within two days. What got faster, and what should not be done by hand at all.
Company accounts, if you want them. We will set up and manage ChatGPT Teams, Claude or Copilot billing so it is not on someone’s personal card.
For teams of five to fifty where AI has already arrived informally. Priced per session, per role group.
One record of your business — customers, products, orders, deliveries — in accounts you own.
One Record of the Business.
The customer list is in the POS. The invoices are in Xero. Half the conversations are in WhatsApp and the rest are in somebody’s Outlook. Deliveries are on paper in the van. Asking how many customers the business has is a week’s work, and the answer is an estimate.
Every other thing on this page sits downstream of this one. An agent cannot work on data it cannot reach, and no amount of model quality substitutes for a record that does not exist.
01Irelax: four sources in, 6,204 customers out, one row each, with every order, service and message attached.
02188 rows the AI would not guess at were flagged for a person. A staff member cleared them in an afternoon.
03Two weeks. It is usually weeks one and two of every engagement.
The failure here is not a missing warehouse. It is a quiet wrong merge — two customers with similar names folded into one, discovered eighteen months later by the one of them who gets the other’s warranty letter.
We do not guess a match. When the system cannot establish that L. Chen and Lily Chen are the same person, it flags the pair rather than merging them. On the Irelax build that was 188 rows out of 6,392. The cost: somebody at your business spends an afternoon on them, and that afternoon is not optional.
It goes into your accounts. Postgres in an account the business owns and pays for, on a monthly bill you can see. The cost: it is one more subscription with your name on it, and you can be talked out of it by anyone offering to host it for free. They are not doing you a favour.
We take only the fields a job needs. The win-back agent needs purchase date, model, contact and opt-out status. Not “everything, just in case”. The cost: the second agent usually needs a second pass at the sources, and we quote it separately rather than pretending the first pass covered everything.
Your POS keeps running. We do not replace it. It stays the system of record for what it is good at and we read from it. The cost: it also stays the constraint. If the POS has ten years of bad address data, we clean it once, and it drifts back unless we are allowed to write corrections home.
Weeks one and two. Before any agent is built, on almost every engagement.
A read-only login from day three. You watch it fill.
Every match rule written down. In English, reviewable by someone who is not technical.
Export to CSV at any time. No lock-in, no export fee, no conversation about it.
Kept matched by monthly upkeep. New data arrives daily; matching is not a one-off.
For any business whose customers, products and orders live in more than one place. Usually the first two weeks of every engagement.
It reads and writes where your people already work: the POS, Xero, Shopify, WhatsApp, WeChat.
Where the Work Already Happens.
Not because the automation was wrong. Because the result landed somewhere nobody opens. Within a month somebody is copying it into the real system by hand, and two months after that the project is described as “something we tried”.
01Results return to the system people already open. No second login, no new habit.
02Systems without an API are handled too, with the ceiling documented before you pay for it.
03Every write is logged: what changed, when, and which rule caused it.
We write back, not only read. Anything that only reads produces a report, and reports get ignored by week three. The agent has to be able to put the order into Xero, move the stock in the POS, mark the message answered. The cost: write access is a real conversation with your bookkeeper, and it should be. We scope the keys tightly and we expect to be asked why.
Where there is no API, we tell you what will break. Some New Zealand systems have none. We can drive the screen or exchange files on a schedule, and it works — until the vendor changes a layout. Those integrations are named as fragile in the quote and they carry higher upkeep. Everyone finds this out eventually; we would rather you found out in week one.
Rate limits and usage costs go in the quote. If a vendor allows a set number of calls a day, the sync is hourly and not live, and we say which. The cost: our quote reads as less impressive than one that promises real-time and discovers the limit in month three.
Credentials stay yours. Your accounts, scoped keys, every key listed in one document with what it can and cannot do. Revoke any of them without calling us. The cost: a slower first week, because getting access properly takes longer than being handed an admin password.
Priced per system. Listed separately, so you can drop one.
An integration map before any code. One page per system: what is read, what is written, how often.
No silent failures. A failed sync messages a named person at your business, not just us.
Keys documented and revocable. By you, at any time, without notice.
Covered by monthly upkeep once live. Vendors change endpoints. That is what the monthly is for.
For a business running four systems that have never spoken to each other. Listed and priced per system.
The screen your team opens on Monday: a queue, the reason each line is on it, and two buttons.
The Screen Your Team Opens on Monday.
Somewhere a person can act on it in ten seconds. A chat window is not that place, and neither is a daily email digest. What the work needs is a queue: what to do, why it is on the list, and two buttons.
01Picking across three SSA Wheels branches, on staff’s own phones. No scanners bought, no system replaced.
02Claim locking: two pickers never take the same order.
03Win-back queue at Irelax: 212 drafts, approve the lot in one action or edit the three that need it.
Internal tools do not fail by being wrong. They fail by being routed around. If the warehouse is back on paper in week three, the build failed, regardless of what the tests say.
It runs on the phones already in the building. No hardware purchase, no scanner fleet, no tablet order. The cost: we design to the oldest phone on the floor, which rules out interfaces that would look better in a screenshot.
We do not replace the system people already know. The POS stays. This sits beside it and does the one thing the POS does not. The cost: two apps open instead of one — accepted, because the alternative is a migration nobody asked for and a month of everybody being slower.
Two buttons, not a form. Approve and Edit. When a screen seems to need a third decision, that is usually a missing rule in the agent, and it goes back there instead of becoming another dropdown. The cost: we push back on feature requests, in writing, and some of them are yours.
We stand in the warehouse in week one. On site, behind someone using it, watching what they do instead of what we expected. Whatever is wrong is fixed that week. The cost: none to you — but it means we do not consider the build finished at handover, and neither should you.
Quoted as part of the agent. Not as a separate app project.
Role-based logins. The warehouse floor does not see margin.
Built for the phones you have today. Tested on them, not on ours.
One week on site after launch. Watching people use it, fixing what we got wrong.
Changed by message, same as everything else. No change-request form.
For teams who will route around anything slower than what they do now. Running across three SSA Wheels branches.
A site drawn for your business, live in two weeks, changed by sending a message.
Rebuilding the New Zealand Web.
Most of them were built between 2015 and 2018, on WordPress or Squarespace, by an agency that no longer answers email, with a login nobody in the business still has. They are not broken. They are describing a company that has since changed, at a speed phones no longer tolerate, in a structure language models cannot read.
We rebuild them as software rather than as documents.
01Two weeks from kickoff to live. An agency quote for the same scope runs six to twelve.
02A finished homepage, your products and your words, on a URL you can open on your phone, before you are invoiced.
03Content changes ship the same day, by message, at no cost per change.
Website projects rarely fail at launch. Launch is the easy part. They fail four months later, when the prices are wrong, the new model is not on it, and the last person who knew the login has left. The site stops being true, and a site that is not true stops producing enquiries.
We build against month nine, not against launch day.
We do not give you a CMS. Every small and medium-sized business we have worked in has a content management system nobody has opened in a year. The self-service is theoretical. The staleness is real. So we removed it. You send a message — new model on the front page, hours change from Monday — and it is live that day, included in the monthly. The trade is honest: you cannot go in and change it yourself at 11pm. What you get for that is a site that is still correct in month nine.
The site is a view of your database. Price, stock, lead time and branch hours render from the record, not from text somebody typed into a page. Nobody updates the website when a price changes, because there is nothing to update. This is also the product’s honest limit — if that data still lives in a spreadsheet on one laptop, this part does not work until the data foundation is in.
Designed, not assembled. No theme is bought and modified. Type, grid and motion are drawn for the business; the site you are reading now is our own. Two engineers with AI doing the production get that to live in one to two weeks. What we cannot do is a $3,000 site. Below that a template is the right answer, and we will tell you so.
Built for the reader who is not a person. A growing share of buyers ask a model instead of searching. We write clean HTML with schema.org markup, one real page per product, and no text trapped inside images or rendered only by JavaScript. Nobody can promise a model will cite you. We can make sure it is able to read you, which a 2016 site generally is not.
You see it before you are invoiced. One real page — your homepage, your products, your words — on a URL you can open. If it is not clearly better than what is live, we stop and you owe nothing.
Two engineers, in your building. Not a ticket queue in another timezone.
Old URLs mapped before launch. Every address on the current site gets a 301. Your rankings, and the links other people have to you, survive the move.
Everything in your accounts. Domain, host, repository, analytics. If you leave, nothing has to be handed over, because none of it was ever ours.
English and Mandarin from one source. Both written, neither translated. Same structure, same launch day.
For a business whose price and service are fine and whose website is eight years old. The site you are reading is the first one.
The same twenty questions, answered from your own stock, prices and order status.
The Twenty Questions That Arrive Every Day.
Do you have it in stock. What does it cost. Do you deliver to Hamilton. Is mine still under warranty. What time do you close. Each one costs a person a few minutes and a look in a system somebody has to know their way around. They arrive on the website, on WhatsApp, on Facebook, on WeChat and on the phone, and as many of them arrive after five as before it.
01Answers come from the record — stock, price, order status — not from a document of frequently asked questions.
02When it cannot answer, it says so and hands over, with everything already said attached.
03An enquiry that arrives at 9pm is captured, answered where it can be, and waiting in a person’s queue at eight.
The two ways this goes wrong are well known. It invents a price with total confidence. Or it traps someone in a loop with no visible way to a human. Either costs more than the receptionist it replaced, which is why most owners here have already decided chatbots do not work.
It answers only from your data. Stock, price, order status, hours, warranty — read live from the record. Anything not in the record, it does not answer. The cost: it will decline questions a person would have handled, until we widen what it can reach, and month one involves a list of those.
There is a way to a person on every single turn. Not buried three levels down. The cost: some conversations hand over that could have been finished, particularly early.
It does not quote outside the price list. Discounts, specials, anything negotiable goes to a person with the context attached. The cost: this does not replace a salesperson, and we will not claim it does.
Every conversation becomes a row on the customer. The enquiry that would have died in a Facebook inbox is attached to a person, and the win-back agent can act on it a year later. The cost: none — but it is why we build this rather than resell a chatbot, and it only works if the data foundation is in.
On the site you have, or the one we build. It does not require a new website.
One agent, both languages. English and Mandarin, not two bots.
Escalation to a named person. With the thread, not a notification.
Every conversation reviewable. You can read what it said, any time.
Monthly upkeep. Your prices and stock change; so must it.
For anyone whose staff spend the morning answering questions a system could look up. Needs the data foundation first.
One agent, one job, carried out the way the person who does it today would carry it out.
One Job, Done the Way You Would Do It.
Never in the model. The difficulty is that the way the job is done lives in one person’s head — which case goes first, how the figure is arrived at, what gets checked before anything leaves, when an exception is allowed — and has never been written down. That person is usually your busiest.
01Win-back at Irelax: 6,204 customers read, 4,412 not back in five years, 212 drafted, 19 opted out and 7 with open tickets held back.
02Takeoff at Archimax: a drawing that cost an estimator a day comes back as an order in their own item codes before lunch.
03Nothing that reaches a customer or moves money leaves without a person’s yes.
An agent that is right ninety percent of the time is not ninety percent of a good outcome. One wrong message to a customer, or one wrong quantity on an order, costs more than the week it saved. So the shape of the product is set by the ten percent, not the ninety.
One agent does one job. Not an assistant for the business. A larger job is two or three narrow agents handing to each other, each with its own rules and its own data. The cost: you end up with several things to maintain rather than one, and the seams between them are ours to keep clean.
The rules are written down before anything is built. Two or three sessions with the person who currently does the job. This is the slowest part of the project and the part worth the most — most of what we write down is useful to you even if the agent is never built. The cost: several hours of your best person’s time, in the weeks they can least spare them.
It stops when a rule runs out. No improvisation. When a drawing has a condition the rules do not cover, it asks rather than assuming. The cost: an agent that asks too often is irritating, and the first month is spent moving that line to the right place.
Every action is on record, including the wrong ones. Input, rule followed, output, who approved. The cost to us is that our mistakes are visible and countable, which is not the industry norm. It is the only way you can tell whether the thing is actually working.
Rules first, and they are yours. Written down, in plain language, whether or not you build.
Fixed quote per agent, per phase. One to two weeks a phase. You see each phase running before the next starts.
It runs in shadow first. It drafts, nothing sends, you compare against what your person would have done.
An approval queue for anything that matters. Customers, money, your records. Everything else runs through.
Monthly upkeep, because rules change. A rule correct in March is wrong in June, and nobody notices without someone reading the numbers.
For a job that costs someone hours every week and has a right answer. Running at Irelax, Archimax and SSA Wheels.
The parts nobody should have to remember, running on a schedule or on an event.
The Parts Nobody Should Have to Remember.
The Monday report. Chasing the deposit. Reordering at the reorder point. The follow-up two days after delivery. None of these fail loudly. They simply do not happen the week that person is on leave, and nobody can say afterwards what it cost.
01Dispatch at Irelax: stop three ran late, the remaining stops re-planned at 09:48, total drive time down 18 minutes, confirmed with one tap.
02Runs on a schedule or on an event. Nobody starts it.
03When something falls outside the rules it stops and asks, with the history there to read.
The characteristic failure of automation is silence. It ran, it half-worked, nobody was told, and the discrepancy surfaces as a customer complaint six weeks later. An automation that fails loudly is worth more than one that is slightly more capable.
Every automation has an owner’s name on it, and the name is not ours. Someone at your business is told when it breaks. An automation nobody owns becomes a mystery within a year, and mysteries get switched off. The cost: you have to nominate a person, and they have to agree.
It fails loudly and stops. If a sync breaks or a rule cannot be applied, it halts and messages that person rather than doing its best guess. The cost: more interruptions in month one than you expected, while the rules are still loose.
We will not automate a process that is wrong. Automating a bad process produces wrong answers faster and with more authority. Sometimes the right outcome of a fortnight is that you change how the job is done and we build nothing. The cost to us is a project that ends without software, and that has happened.
No drag-and-drop layer. These are built as code in your own infrastructure, not as a Zapier account nobody can read. The cost: you cannot build a new one yourself on a Friday afternoon. Changes come through us, or through your upkeep hours.
Priced per workflow. Small ones are small money.
Shadow runs before it acts. It logs what it would have done, you read a week of that.
Alerts to a named person. Not to a shared inbox.
Every run logged and readable. What it did, what it skipped, and why.
Included in monthly upkeep once live. Triggers break when the systems around them change.
For the jobs that only happen when one particular person remembers. Running at Irelax on the day’s delivery run.
Someone keeps it alive: data drifts, rules change, the POS updates. From $120 an hour.
Month Nine.
A supplier renames a product line. The POS updates an endpoint. A rule that was correct in March is wrong by June. None of it announces itself. The agent keeps running, quietly producing worse output, and the first person to notice is a customer.
01A monthly report: what ran, what it did, what it declined to do, and what changed underneath it.
02From $120 an hour, billed monthly. Cancel with a month’s notice.
03Unused retainer hours go to the next agent rather than expiring.
The normal ending of a software project at a small or medium-sized business is that the vendor builds, invoices, and is never heard from again. That is why owners here are sceptical before you open your mouth, and it is a reasonable position.
We price this before the build, not after. You know the monthly before you commit to the project. A build we cannot maintain at a price you will actually keep paying is a build we should not take. The cost: our proposal looks more expensive on day one than a competitor’s quote that leaves the number out.
We watch rather than wait to be told. Failed syncs, match rates dropping, an agent declining more often than last month — those reach us before they reach you. The cost: this is most of what the monthly pays for, which is why the monthly is not a token amount.
Cancel with a month’s notice, and everything stays where it is. Database, code, keys, domain, analytics — all in your accounts already. The cost to us: it removes the lock-in that most of this industry depends on. We would rather you stayed because it works.
AI usage is passed through at cost, with a cap you set. You see the actual bill. As usage approaches the cap it tells you, rather than running past it. The cost: a cap can stop an agent mid-month, and we would rather that than an invoice you did not expect.
From $120 an hour, billed monthly. Quoted before the build starts.
A monthly report. One page, readable by someone who was not in the room.
Retainer hours roll forward. Into the next agent, not into nothing.
Cancel with a month’s notice. Nothing to hand over; it is already yours.
AI usage at cost, under your cap. With a warning before it is reached.
For anyone running one of the above after launch. Monthly report, and retainer hours roll into the next agent.
Data foundation
System integration
AI apps & internal tools
AI website
Agent build
Operate & upkeep
You don’t need a strategy deck or a big budget to begin. Pick the one job that costs your team the most time — we’ll come to you, look at it together, and tell you which of these ten offerings fits, with the price fixed before anything starts. The first conversation is free.
Four steps, two minutes. Within two working days: can AI take them on, what it needs, roughly what it costs.
