A customer calls at twenty past nine and asks a simple question: do you have four of these in stock, and can they be on site Friday?

Nobody in the room can answer it. The stock figure is in the POS, and the person who knows how to read it properly is on a delivery run. Friday depends on whether the supplier’s cut-off is Wednesday or Thursday, which is in an email from last March. Whether this customer gets the trade price is in a contract in a folder somebody named after the customer’s old company name. So the question goes into a WhatsApp group, and the customer is told someone will call back.

By the time the answer comes together it is early afternoon. It took three people, two of whom were doing something else, and the customer has already called someone who answered first.

What makes this expensive is not the one question. It is that everybody has learned not to ask. The stock report gets pulled once a week because pulling it is a chore. Nobody checks whether the supplier changed the cut-off. The trade price gets guessed. The business runs on what people remember, and what people remember is a year out of date.

The business already holds the answer. It is spread across a database, an accounts package, a shared drive and a few inboxes, and the only index is the staff.

We connect all of it to a large language model through MCP — Model Context Protocol, the open standard for giving an AI access to a company’s own systems and files. Two kinds of source, connected the same way.

Nothing moves. No copy is taken, no data leaves your accounts, and no new system is introduced — staff ask in the chat tool they already have open.

Take that Friday question and run it again, with the connection in place.

Someone types it the way they would say it to a colleague: “Do we have 4 of the 90mm brackets, and can they be on site by Friday?”

The agent does not go looking for a document called “90mm brackets”. It reads the map of your database first and works out that this is a question about stock on hand, for a product whose code it needs to find. It runs one read-only query, gets eleven across two branches — seven in Albany, four in Penrose — and holds on to the fact that these four would have to come from Penrose.

Then the second half of the question, which is not in the database at all. It searches the documents it has been given and finds the supplier’s terms: orders confirmed before 2pm Wednesday ship Thursday and arrive Friday. It checks the date. It is Tuesday, so Friday works — as long as the order is confirmed today or tomorrow morning.

The answer comes back in about fifteen seconds, in one message: four are available at Penrose, Friday is achievable if the order is confirmed before 2pm tomorrow, and this customer is on trade pricing under the agreement signed in March. Under each part is where it came from — the stock table with its timestamp, the supplier email with its date, the contract with its file name — so the person on the phone can check any of it before repeating it to the customer.

Two things it does not do. It cannot write: no order is placed, no record is changed, no price is updated. It holds a read-only connection and it only sees the tables and folders you granted it. And when the answer is not in what it has been given, it says so — “I don’t have the freight terms for this supplier” — instead of producing a confident sentence that sounds right. Every question, and every query it ran, is logged.

When an action does need to follow — placing the order, sending the quote — that is a different agent, and it stops in someone’s queue for approval before anything leaves the building.

The change is not that answers arrive faster, though they do. It is that the questions start getting asked at all. Stock gets checked before a promise is made instead of after. The supplier’s terms get read on the day they matter. New staff stop needing the one person who knows where everything is, which means that person gets their week back. The business stops running on memory and starts running on what it actually holds.

Ask a small or medium-sized business how many customers it has and you will get three answers. The till says 6,204. The spreadsheet — the one called customers-FINAL-v3, on the laptop in the back office — says 4,980. The accounts say 5,617, except that includes two companies that closed years ago and one customer who appears four times under four spellings of her own name.

None of this is anyone’s fault. The POS arrived with the shop fit. The online store was added in a hurry one lockdown. The spreadsheet is older than both and survives because one person is fast with it. Each system is right about the part of the business it can see, and no two of them agree about anything.

A person bridges the gaps without noticing. The counter staff know that the “J Wilson” in the spreadsheet, the “Wilson, James” in the till and the company called JW Interiors are the same man, and that the delivery address on file is his old house. An agent knows none of this. It reads what is written, and what is written is three strangers and a wrong address.

So before any agent can take on any job, the business has to be able to answer four questions from one place: who are our customers, with their history; what do we sell, in the codes we actually say out loud; what did we deliver, to whom, and did it arrive; and which of these records do we trust.

We export everything and rebuild it as one structured record, in accounts the client owns — not ours. Customers linked to orders, orders to products, products to suppliers, deliveries to addresses, with ten years of history intact.

The matching is the real work, and it is full of judgement calls that look tiny and are not. A married name and a maiden name. A tradesman with a personal account and a company account. A supplier that renamed itself mid-decade. Everything that can be settled by rule — same email, same phone, same delivery address — is settled by rule. Everything the rule is not sure about is flagged for a person instead of guessed. For Irelax that was 188 flagged rows out of 6,204 customers. For SSA Wheels it was three versions of the same product list, one per branch.

Those 188 rows became one afternoon at a desk with the owner: is this the same person, which of these two price lists is real, do we keep the addresses from 2014. Every decision was recorded on the row itself — matched by rule or confirmed by a person — so that two years from now nobody has to wonder whether a figure was real or inferred.

Then the live systems are connected, so the record keeps itself current instead of ageing from the day it was built.

There is no agent in this scene, and that is the point of it. This is two engineers in your office for a couple of weeks, and it is worth seeing what those weeks actually contain.

The first days are extraction: the till, the store, the accounts, the laptop spreadsheets, all copied into a working set — nothing live is touched. Then the matching runs, and the flags come out, and there is the afternoon of decisions described above. It is unglamorous in exactly the way that matters: every one of those decisions used to live in somebody’s head, and now it lives in a table.

What comes out the other end is a database with none of the cleverness on show. The tables have names a person can read. The product codes are the ones the counter staff use on the phone. A row carries its source and its confidence. It sits in the client’s own accounts, and any tool can read it — including tools we did not build and will never see.

Every other scene on this page stands on this one. The enquiries agent answers from this record. The win-back list is drawn from it. The pickers’ phones read the catalogue that came out of it. None of them would survive a week on the data as we found it.

The first change is that questions stop having four answers. The catalogue the website shows is the catalogue the warehouse picks from; the customer who rings is the same customer the accounts know. The second change is speed: once the record exists, each new agent takes weeks rather than months, because the hard part is already done. And the record is yours — if we disappeared tomorrow it would still be sitting in your accounts, readable by the next person, with every uncertain row honestly marked.

By ten past eight there are nine messages waiting. Two are “where is my order”. One wants to know whether the leather colour in the photo is the colour in real life. One asks if delivery can move to Saturday. One is a warranty claim on a chair bought four years ago. The rest are variations the shop has answered a thousand times.

None of them is hard. Every one of them needs somebody to open a system and know their way around it — the order in the POS, the stock in the store admin, the warranty terms in a PDF somebody saved in 2021. So the messages queue behind whoever is free, and whoever is free is also serving the showroom. An enquiry that lands after five waits until morning.

The usual fix makes it worse. A chatbot gets bolted on, it answers from a page of FAQs, and within a week customers have learned that it cannot see their order. They type “human” and wait anyway, now slightly angrier than if the bot had never existed.

The real cost is not the queue. It is that the person clearing it is the same person who sells chairs, and every hour spent retyping order numbers is an hour off the floor.

The agent sits on the live record — the one built in the data scene — rather than on a FAQ page. It has four tools: look up an order, check stock, read the warranty and returns terms, and draft a reply.

What it may do with each tool is set by rule, and the line is simple. Reading is free: order status, stock levels, published terms. Anything that reaches a customer or changes a record is drafted and queued for a person. A return is started, never approved. A refund is proposed, never issued.

It also writes like the shop, not like a ticketing system — short, direct, no reference numbers in the first line — because the draft a person has to rewrite is a draft that saved nobody any time.

A message arrives at 8:12: “Ordered a chair three weeks ago, still nothing, getting annoyed.”

The agent finds the customer from the email address, finds the order, and sees it is sitting at the depot waiting on a delivery slot. It also sees this is the second message about the same order — which changes what a good reply is. It checks the delivery calendar and finds two Saturday slots genuinely open.

The draft it queues reads: “You’re right, and I’m sorry — your chair has been at our depot waiting on a delivery slot longer than it should have. I can hold Saturday 9–11 or Saturday 1–3 for you this week. Reply with either and it’s booked.” Under it, for the reviewer, sit the order, both previous messages and the two slots it checked.

The person clearing the queue reads the situation, not just the draft. Most go out as written; some lose a sentence or gain one. Roughly one in fourteen gets pulled out entirely and handled by a person from the first word — usually a complaint that needs an owner’s voice rather than a correct answer.

The split across a normal week: 62 percent of enquiries are answered on the spot, because they are pure reads — where is it, is it in stock, what does the warranty say. Another 31 percent end in a drafted action that waits for approval. The last 7 percent go straight to a person. The queue is cleared twice a day, morning and mid-afternoon, and takes about ten minutes each time.

When the agent does not know, it says so. A question about a part for a model discontinued before the record begins gets “I’ll pass this to the team” — not an invented answer — and lands in the human pile. No record has ever been changed, and no message has ever been sent, without a person’s yes.

About seven enquiries in ten are now answered without anyone opening a system, and the ones that do reach a person arrive with the order, the history and a draft attached — a minute’s work instead of ten. The customer who wrote at 8:12 has an answer before the showroom opens instead of after lunch. And because every reply is drawn from the live record, it is the same answer the warehouse and the accounts would have given — which is not something the old inbox could promise.

Irelax has been selling massage chairs for ten years, which means somewhere in its systems are ten years of people who spent four or five thousand dollars each and then never heard from the shop again.

Everyone knew the list was worth something. A massage chair is a five-to-ten-year purchase: the person who bought one in 2016 is the best upgrade prospect the business has, and the person whose warranty runs out next month is the second best. The names were all there — in the till, the online store, a spreadsheet and the accounts, in four spellings, never merged.

The plan was always the same — “we should really go through the list one day” — and one day never came, because “go through ten years of receipts and work out who to ring” is a week of work nobody has.

So the asset sat there. The business had paid for every name on it — each one is a delivered chair — and the list earned nothing, year after year.

The record came first: 6,204 customers rebuilt from the four sources into one list, with what each person bought, when, for how much, and everything that has happened since — service visits, complaints, returns.

Then the rules, written down with the owner rather than assumed. Who qualifies: bought five or more years ago, no purchase since, no unresolved complaint. Who is excluded, permanently, no matter what the numbers say. And what each group should hear: the ten-year owners get an upgrade conversation, the warranty-ending customers get a service reminder, the office that bought two chairs for a staff room gets neither — it gets a business message.

Then the agent, whose entire job is to draft. It has no send button.

The agent runs on Monday mornings. It reads the record, applies the rules, and finds who qualifies this week. The first run found 4,412 of the 6,204 had not been back in five years or more — which is the number that made the owner sit down.

Each week’s batch is deliberately small and deliberately specific: not “everyone lapsed”, but the people for whom something is true this week — a warranty ending this month, a chair passing its ninth year, a model whose replacement just arrived in the showroom.

A draft reads like this: “Hi Marie — the recliner you bought from us in 2016 is coming up on nine years, and the model that replaced it does things yours can’t. If you’d like to try it, come in any Saturday — and we’ll take the old one off your hands either way.” The specific detail is the whole trick. “We miss you” gets deleted on sight; “your 2016 chair, here is what replaced it, we’ll take the old one away” gets a reply.

The drafts land in a queue, and on Monday the owner reads them with a coffee. Most get sent as written. A few get a line changed. And some get killed for reasons no record could know — the customer who had a rough service visit, the one who mentioned they were moving to Australia. That judgement stays where it belongs.

The first month put 212 drafts in the queue, of which the owner approved and sent 41. That ratio is not a failure — it is the design. The agent proposes widely so the owner can choose narrowly, and nothing has ever gone out without his yes.

When a customer replies, a person answers. The agent’s job ended at the draft.

The job that needed a week nobody had now takes twenty minutes on a Monday. The business contacts customers on purpose — because a warranty is ending, because a chair is nine years old — instead of when somebody remembers, and every message carries a detail that proves it is not a blast. And the record underneath is the same one the enquiries, the deliveries and the accounts run on, which is why the agent knows about the rough service visit’s complaint ticket — even when it does not know what the owner knows.

The four-message dance goes like this. “What day suits you for delivery?” — “Thursday?” — “Morning or afternoon? The truck’s in your area in the morning.” — “Actually, Thursday’s no good, my husband works from home Thursdays.” By the fourth message the morning slot has gone to someone else, and the conversation starts again.

Meanwhile the run itself lives on a whiteboard. The order of the stops is whatever made sense to the person holding the marker at eight in the morning, and a job’s status is whatever the driver last said on the phone. When a stop runs long — and a two-seater up three flights always runs long — the rest of the day is repaired by ringing people.

The customers at the end of the run find out they have been pushed when they call to ask where the truck is. That call, more than the delay itself, is what they remember about the shop.

The obvious answer is scheduling software, and it fails in a business this size for a reason that has nothing to do with software: the jobs were never defined. Nobody had written down that a standard drop takes thirty minutes and a stairs job ninety, that a crew tops out at six jobs a day, or which suburbs can sensibly share a run. Any calendar assumes those numbers exist. Here they lived in the crew’s heads.

So the first fortnight produced no software at all. It was spent with the crew, turning what they knew into rules: the job types and how long each really takes, the six-job cap, the three statuses a job can be in — booked, on the truck, done — and which jobs can share a day.

Only then does the agent earn its place, because only then can it offer slots that are real. It reads the calendar, the crew’s remaining capacity and the addresses already on each run, and proposes times that will actually hold — rather than times that merely look empty in a diary.

A customer buys a chair and wants it delivered next week, somewhere in West Auckland.

The agent looks at the runs that already exist for those days. Thursday morning has two stops within ten minutes of the address and room under the cap; Friday afternoon works too. So the first message the customer gets is also the last one: “We can do Thursday 9–11 or Friday 1–3 — which suits?” One reply books it. The job lands on the run with its type, its real duration and its status, and the crew sees it on a phone instead of a whiteboard.

Then Thursday does what Thursdays do: the stairs job runs twenty minutes over. The agent re-sequences the remainder of the run on the spot, works out that two customers will now fall outside their promised windows, and drafts their messages: “Running about 25 minutes behind this morning — we’ll be with you closer to 11:45. Sorry for the wait.”

The crew leader confirms with one tap. He is standing in a driveway holding one end of a recliner, so one tap is the entire interaction budget. The messages go, the board in the office updates itself, and the customers at the end of the run hear about the delay from the shop instead of discovering it.

The agent never contacts a customer on its own, and never moves a window that was promised without a person confirming. The plan is its job; the promise stays human.

Agreeing a slot went from four messages to one, and a broken day is now repaired in minutes instead of an afternoon of phone calls — with the affected customers told before they have to ask. The office can see where the run is without ringing the driver, which sounds small and changes the whole day’s temperature. And none of it came from clever routing, which is the easy part. It came from the fortnight of writing down what the crew already knew.

It is Sunday evening and the owner is doing what he does one Sunday a month: the bank feed on one screen, the orders on the other. A payment of $1,850 arrived on Tuesday with the reference “chair”. Which chair. Whose chair.

Half the overdue list turns out to be wrong, and it is wrong in three different ways at once. A customer paid a deposit in March and the balance on delivery in May, but the two payments carry hand-typed references, so the order still shows unpaid. Another customer paid the right amount against the wrong invoice. A third genuinely has not paid. From the outside, all three look identical.

Telling them apart means opening the order, the bank line and the delivery note side by side, for every line — which is why it happens once a month, and why it happens to be the owner’s Sunday.

The real damage is quieter than the lost evening. When every second line on the overdue list is a matching error, nobody trusts the list, so nobody chases anything until a customer is embarrassingly late. And once or twice a year the shop chases someone who paid months ago — which costs more goodwill than the invoice was worth.

Payments are now matched against orders every morning rather than once a month: amount, date, reference and customer, with the order and its delivery status alongside.

The rules that matter are the awkward ones, and they were written down with the owner before anything ran. A deposit is not a debt while the delivery has not happened. A payment short by the exact value of a discount code is a match, not a mystery. A payment from a name that does not match the customer — a spouse, a company account — is proposed to a person, never assumed.

Only exact rule-hits close themselves. Everything else is a proposal with the evidence attached. And the roughly 140 historical lines that no rule could settle went to a person, one by one, rather than into a guess — because a ledger you have guessed at is worse than a ledger you know is incomplete.

Tuesday’s feed has a line the rules cannot close: $2,140, no reference, from a name the record does not know.

The agent looks for footholds. There is exactly one open order for exactly $2,140, delivered last week to a suburb that matches the payer’s bank branch. It proposes the match — this payment probably belongs to this order; the payer may be a partner or family member — and shows its reasoning. A person confirms it in five seconds, and the couple’s next delivery will not be chased for money they have already paid.

What comes out each morning is one short list in order of attention needed. This order is genuinely overdue, by this many days, and here is the customer’s history. This one is waiting on a delivery — leave it alone. This payment is $40 short, and the $40 is the loyalty discount, so close it.

For the genuinely overdue, the agent drafts the reminder: “Hi James — invoice 4471 for the recliner delivered on 12 August is still showing unpaid ($2,300). If it’s crossed with a payment, ignore this; otherwise here are the details…” — the shop’s voice, the order attached, an exit ramp in the first line for anyone who has already paid.

The draft waits for approval, and this queue gets read more carefully than any other in the business, because the cost of a wrong chase is a insulted customer. No reminder has ever left without a person reading it. The agent cannot mark an order paid on its own, and it cannot send anything.

The overdue list went from monthly and half-wrong to daily and true, which means it finally gets acted on: real debts get chased inside a fortnight instead of at six months, and nobody who has paid gets a letter. The unmatched pile stopped growing — new payments are settled the morning they land, not archaeologically at month’s end. And the owner got his Sunday evenings back, which he mentions more often than the cash flow.

The reorder at SSA Wheels ran on the memory of whoever had been there longest. He knew which lines moved in winter, which supplier was slow, and which tyre the counter called by a name that appears nowhere in any system. When he was on leave, the ordering waited for him.

Three branches made it structural. Three versions of the product list, the same tyre written three ways, so nobody could say how many of anything the business actually owned. The company was paying for stock twice: once when a branch bought a line that another branch had on the shelf, and again in the sale lost when a line ran out and the first person to notice was the customer asking for it.

Even the fees were folklore. The Tyrewise disposal levy had been settled three different ways by three different people, all of them sure they were right.

Inventory software cannot fix this, because the problem is underneath the software. Feed three versions of the same product list into any system on earth and you get a very confident view of three products that are one product.

The catalogue came first: 872 tyre lines reconciled into one record, under the codes the counter staff actually say on the phone, shared by all three branches. That was the slow part, and the part that made everything after it possible.

Then the rules, which are more particular than “order when low”. Stock you own beats stock you buy: if Penrose is holding four and Albany needs them, that is a transfer, not a purchase order. Seasonal lines move on last year’s same-month sales, not last month’s — April’s winter numbers would starve June. And the Tyrewise fee became one three-step rule that every branch now settles the same way.

Every morning the agent reads sales and stock across the three branches against those rules, and produces a proposal. Not an order — a proposal.

On a given Tuesday it reads like this: six lines need attention this week. Three are transfers — Penrose holds nine of a winter 205/55R16 that Albany has nearly sold through, so move six across before anyone buys anything. Three are purchase orders, each with the supplier, the quantity and the reasoning attached: last June sold thirty-one, current stock across branches is eight, lead time is two weeks, so order now rather than in two weeks.

The person with ordering authority reads it with the reasoning open and decides in minutes. Approve, adjust the quantity, or ignore — the agent accepts all three without argument.

When a line has no history — a range the business only started stocking in spring — the agent says exactly that: no basis to forecast this line yet, and asks, instead of dressing a guess in decimals. That sentence is why the proposals are still being read a year on.

Nothing is ever ordered automatically. No supplier has received an order the business did not intend, and the agent cannot be talked into urgency by a low number alone — low against what, is the whole question, and the answer is in the rules.

The reorder stopped depending on who was rostered, and the longest-serving man’s knowledge is now written down where the business can keep it — which he regards, correctly, as a promotion rather than a threat. Stock moves between branches before money moves to suppliers, which is the cheapest saving in the company and was previously invisible. And the catalogue underneath it is the same one the pickers’ phones read, so the counter, the racks and the reorder finally describe the same shop.

The monthly review, in most small and medium-sized businesses, is a spreadsheet that arrives from the accountant three weeks into the following month. By then it is history — accurate, and useless, because everything it might have changed has already happened.

Until it arrives, the month runs on feel. Which branch had the better month — the owner has an opinion. Whether the winter lines paid for themselves — probably? Whether February’s price change helped or quietly gave away margin — nobody has looked, because looking means joining sales to products to branches to costs, and the one person who can do that join has a full-time job as well.

So the owner’s view and the accountant’s number meet once a quarter, disagree politely, and nothing changes. It is not a lack of curiosity. It is that every question costs an hour, and an hour is exactly what nobody has.

This scene exists for the owner. It is deliberately not the everyday ask-anything of scene one — it is the discipline of reviewing the month, the same way, every month, with the rows on the table.

On the first working day of the month, the agent prepares the review from the business’s own record — and only that record. Nothing is pulled from outside the business, and every figure it reports arrives with the rows behind it, so it can be checked rather than believed.

The pack is the same shape every month on purpose: revenue by branch and by category against the same month last year; the lines that moved most in either direction; margin wherever cost data exists; and the three or four questions the agent thinks the numbers are asking.

The review lands and the conversation starts, in plain English, and this is where the scene earns its keep.

“Why was Penrose down?” — and the answer is not a shrug or a chart, it is the two categories that fell, by how much, against last year, with everything else flat. “Show me the customers behind that.” — and the fifteen accounts appear as rows: two big workshops that bought nothing this winter, and a dozen small drops that add up.

It goes on like that — about thirty questions in a normal week, each answered in the time it takes to read the answer, each with its rows attached. The question that used to cost a week of waiting — ask the accountant, wait for the export, discover the export answered a slightly different question — now costs minutes, which means it actually gets asked.

When the record cannot support an answer, the agent says so in so many words. Cost prices are missing for one supplier’s ranges, so for those lines it reports revenue and states that margin is unavailable — it does not produce a number that looks authoritative and is not. Ask it something outside the business’s own data — how a competitor is doing — and the answer is that it does not know, because it genuinely does not.

That honesty is the entire difference between a tool an owner still uses in month six and one that gets quietly abandoned in month two, the first time a made-up number gets caught.

The review that used to arrive three weeks late now exists on the first working day, and the questions it provokes are answered while they are still worth asking. The owner and the accountant have stopped arguing about which number is right — they are looking at the same rows — and started arguing about what to do, which is the argument a business is supposed to have. Around thirty questions a week get asked that previously died of friction, and every answer can be traced to its rows by anyone who doubts it.

The ad budget gets set in a January meeting and then looked at when somebody remembers. A campaign that stopped working in its first week is still spending in week five. The campaign that is quietly working never gets another dollar, because nobody noticed it working.

The platforms do not help, because they report success in their own currency. A form fill is a “conversion”. But Maria Learning Centre does not sell form fills — it enrols students — and the distance between an enquiry and an enrolment is where ad money goes to die without anyone seeing it happen.

The classical answer is an agency, and for a business this size the mathematics are unkind: the retainer costs more than the budget being managed, and the report still arrives monthly.

So the spend sits in the worst possible position — small enough that nobody senior watches it, large enough to matter.

The agent reads both ad platforms every morning and joins them to the thing the platforms cannot see: the centre’s own record of what each lead became. An ad set that produces plenty of enquiries and no enrolments is now visible as what it is — a failure with good numbers.

For its first fortnight it was allowed to change nothing. It watched, wrote its morning notes, and its recommendations were compared against what the owner would have done. That trial period was not caution theatre — if you cannot check an agent’s judgement against your own for two weeks, you have no business letting it near a budget.

The morning note is short and the same shape every day: yesterday’s spend by campaign, what each produced — in enquiries and in enrolments — and what changed since the day before.

Then the recommendations, rarely more than two or three, each carrying its arithmetic: “Pause ad set B — $412 over eleven days, nine enquiries, no enrolments. Move that budget to the school-holiday campaign, which is enrolling at roughly half the cost per student.”

The owner reads it over coffee and approves, edits or ignores. An approved change is made in the platform by the agent that morning; an ignored one is dropped without argument. Nothing else happens — the agent cannot raise a budget, launch a campaign or widen an audience on its own initiative, and it has no opinion about creative.

What it will not do is decide strategy. It does not write the offer, pick the audience or invent the campaign — those are decisions about what the business wants to be, not patterns in yesterday’s numbers. It watches what is running, against enrolments rather than clicks, and says plainly what it sees.

Every change on the account now has a name and a reason attached — the agent’s reasoning, and the owner’s yes. When something looks odd a month later, the answer to “who did this and why” takes one look, not an argument.

The spend is read every day instead of when somebody remembers, and it is judged on enrolments — the only number the centre actually banks. Dead campaigns now die in days instead of surviving five weeks on momentum, and the quiet winners get fed. The owner’s part takes about five minutes a morning, and not one dollar has moved without her approving the move.

The account is busy for a fortnight after every promotion, and then the person who was doing it in the evenings gets a real deadline, and the page goes quiet for a month. Scroll any small or medium-sized business’s feed and you can read its staffing history in the gaps.

The work is not hard — it is relentless, which is worse. Decide what to post. Write it. Find or make the image. Resize it for the feed, again for stories, again for the web banner. Write it a second time in the second language. Schedule it. Then do all of it again next week, forever, in a business where nobody’s actual job is social media.

And when it does get done in a hurry, it looks done in a hurry: the logo stretched to fit, an orange that is nearly the brand orange, a headline set in whatever font was open.

The page matters more than its follower count suggests, because it is the shop window customers check before they visit. A feed that died in June tells them something, and it is not the thing the business wants told.

Two documents had to exist before any agent touched this, and neither is a prompt.

The brand, rewritten as constraints rather than as a style guide: the exact colour values, the logo’s clear space, which typeface at which sizes, what the product may never appear next to. A style guide is written for people who already have judgement. An agent needs rules it can obey mechanically — and so, it turns out, does a rushed human on a Thursday night.

And the content plan: what this shop posts about and in what mix — product, customer stories, the showroom, offers — what it never posts about, and what the two languages each carry. With those written down, “a week of posts” stops being a blank page and becomes a task with a right answer.

On Monday the agent proposes the week: five posts, what each one is and why it is this week — the winter promotion that is actually running, a delivery photo from Thursday worth using, the massage chair whose replacement just landed and needs introducing.

The owner strikes one, swaps another, approves the rest. That is the strategy conversation, and it takes minutes because it is a reaction to a concrete plan rather than an invention from nothing.

For each approved post the agent then writes the copy in both languages — written for each language, not translated at each other — and lays out the artwork in every size at once: feed, story, banner. The brand rules are enforced in the layout itself: the orange is the orange, the logo sits in its clear space, the type is the brand type. The files come out layered, so when the designer wants the headline two millimetres left, she moves it — she does not rebuild the artwork.

Everything queues for one review, copy and artwork side by side. What used to be a day of production is now about an hour of judgement: is this the right thing to say this week, and does it look right. Then it schedules.

Nothing posts to a live account on its own — the queue is the only path out. And the replies stay human: comments and messages are customers talking to the shop, and a wrong post is embarrassing, but a wrong reply to a customer is a different category of problem. The agent fills the feed; it does not speak for the business in conversation.

The account stopped going quiet, which was the entire brief — on a page nobody runs full-time, consistency beats brilliance, and the algorithm agrees. Every post looks like the brand because the rules are enforced rather than remembered, in both languages, in every size. The designer’s hours moved from resizing the same artwork five ways to the work that actually needs taste — and the owner’s Thursday evenings stopped belonging to the feed.

Today’s run exists in two places: the whiteboard by the roller door, and the driver’s head. The whiteboard says six addresses in the order somebody wrote them at eight o’clock. The driver’s head holds everything else — which customer is never home before ten, which street floods, which job is secretly a stairs job.

Most days this works, because the driver is good. What does not work is the day going wrong: the two-seater that takes an extra half hour, the customer who is not home, the truck held at the depot. Then the office rebuilds the afternoon by telephone, one call at a time, while the driver keeps driving.

The customers at the end of the run find out they have been pushed when they ring to ask where the truck is. That phone call — not the delay — is what they tell their friends about.

Routing software has existed for forty years, and the reason this business did not use it is the honest one: routing needs addresses, promised windows and job durations as data, and until the record was built, all three lived in people’s heads and on the whiteboard.

Once the orders, addresses and job types exist as one record — the data scene, again — sequencing a day is nearly free, and the agent’s plan is built from what is actually on the run: six stops, each with its real duration by job type, the window that was promised to each customer, and the roads between them.

The crew carries the run on a phone. Every stop has a status — booked, on the truck, done — so the board in the office is now a screen that tells the truth on its own, and nobody has to ring the driver to ask where he is.

The run is sequenced before the truck leaves — not in booking order, but in the order that lands the most stops inside their promised windows for the least driving.

Then comes the day the whole thing exists for. The stairs job at stop two runs eighteen minutes over. By the time the crew is back in the cab, the agent has re-sequenced the remaining stops, worked out that two customers will now fall outside their windows, and drafted both messages: “Running about 25 minutes behind this morning — we’ll be with you closer to 11:45. Sorry for the wait.”

The crew leader’s entire involvement is one tap on Confirm. He is standing in a driveway with one end of a recliner in his hands; one tap is the whole interaction budget, and the system was designed around that fact rather than around a dashboard nobody in a truck will ever open.

The messages go, the statuses update, the office watches the day repair itself on the board. The customer at stop five, who under the whiteboard regime would have rung at noon and been told “he’s on his way”, instead got told the truth before she wondered.

The agent never contacts a customer of its own accord, and never moves a promised window without a person confirming. It plans; the promise stays human.

Six stops a day sequence themselves, a re-planned run saves around eighteen minutes of driving, and ninety-four percent of arrivals now land inside the promised window — but the number the owner quotes is none of these. It is the silence: the office phone has stopped ringing at noon with customers asking where the truck is, because the ones affected by a delay have already been told. The whiteboard is still by the roller door. Nobody has written on it in months.

A pick at SSA Wheels started as a printed list and a walk through the racks. What came back had the right number of tyres on the trolley, and, often enough to hurt, one of them was the wrong tyre — a 205/55R16 where the order said 205/50R16, or the right size in the wrong brand, because the two sat a shelf apart and the list was on paper.

A wrong pick is not discovered in the warehouse. It is discovered three days later at a workshop, on a car that is already up on the hoist. Then it comes back as a return, and a returned tyre costs twice — the freight in both directions, and the customer who now waits another week and remembers it.

Three branches deepened the problem in a way nobody chose: three versions of the product list, so the same tyre had three names, and a picker covering a shift at another branch was effectively working in a foreign language.

The obvious fixes were scanners and warehouse terminals, which is to say: hardware to buy, hardware to break, and a system to train three branches on. The business kept not doing it, and the pick errors kept coming back on trucks.

The single catalogue came first — the same 872 reconciled lines the reorder runs on, in the codes the counter staff say out loud, shared by all three branches. Without it, any picking system would simply have automated the confusion.

Then the app, which runs on the phone already in the picker’s pocket. No scanners, no terminals, nothing new to buy or break. It shows one job at a time: the product, the quantity, and where it lives.

An order lands and becomes a pick list on the phone — sequenced in the order the racks are actually walked, not the order the customer typed the lines.

At each line the picker confirms what they took, and the check happens on the spot. If the code in hand does not match the line, the phone says so in the aisle, where the fix costs ten seconds — not three days later on a hoist, where it costs a return, a truck and an apology.

The near-misses get special treatment on the screen, deliberately: two sizes one character apart are displayed as emphatically different things, because one character apart is exactly where the returns lived.

When the rack disagrees with the record — the system says four, the shelf holds three — the picker records what is actually there, and the discrepancy goes onto a list for a person instead of into a silent guess. The record gets truer every time someone touches it, which is the opposite of what paper did.

The moment the pick completes, the order updates for everyone at once: the counter sees it is ready without walking out the back, the delivery run receives it, the accounts see it moving. Nobody re-keys anything, because there is nothing to re-key.

Pick errors fell by eighty-three percent and time per order by thirty-five, with not one piece of new hardware — the phones were already in everyone’s pockets. A picker can now cover a shift at any branch and read the same catalogue in the same codes, which quietly fixed the rostering problem nobody had filed under “data”. The app itself is almost embarrassingly simple, and that is the lesson the owner repeats: the app was never the hard part. The catalogue underneath it was.

A drawing arrives from a builder mid-morning: a PDF of the elevations with a few pencil marks and a note that says “price these please”. At Archimax, turning that into a quote takes an estimator most of a day — count the openings, size each one, work out the componentry, look up the codes, price it, type it.

The counting is not the hard part. The hard part is the componentry: which head track goes with which panel system, how many brackets at what spacing, what changes when an opening crosses a certain width, which seal a coastal job needs. Those rules lived in the heads of two people, and in twenty years nobody had written them down — because the two people were always in the building.

So the quote took a day when the builder wanted it by lunch, and the order it became was occasionally wrong in ways nobody could see on paper. A wrong bracket count does not fail in the office. It fails on site, on install day, which is the most expensive place in the industry to discover anything.

And under it all sat the quiet business risk: the company’s real asset — how to turn a drawing into a correct order — existed nowhere except in two salaries.

The build began with a day that involved no software at all: sitting with the senior estimator and writing the rules down, item by item, as data rather than prose. This panel system requires this head track, one bracket per six hundred millimetres plus one, this seal by exposure grade, and here are the exceptions by size and configuration.

That rule set is the asset, and it now belongs to the business. The agent is what makes it fast: it reads a marked-up drawing, finds the openings and their sizes, applies the rules, and produces a line-by-line order — in Archimax’s own item codes, the ones the warehouse picks by and the accounts invoice from, so nothing needs translating afterwards.

Because the rules are data, a new product range is a configuration exercise, not a rebuild — add the range’s rules, and the agent quotes it the same afternoon.

The builder’s PDF goes to the agent. It reads the sheets, finds the openings, sizes them, and builds the schedule: this many panels at this size, therefore this head track, twelve brackets — not eleven, because the 2,400mm opening crosses the width threshold and the over-width rule adds a pair — and this much seal.

Every line on the schedule shows its working: which opening, on which sheet, produced by which rule. The estimator is never asked to trust a total. He is shown the reasoning the way he would show an apprentice his own, and that is precisely what makes a forty-minute check possible where a day of counting used to be.

Anything the drawing does not settle comes back as a question, not a guess. A dimension the scan made unreadable. A configuration the rules do not cover. A pencil mark that contradicts the printed plan. Each one is flagged with its location — sheet 4, the western slider, the mark-up says 2,100 but the elevation reads 2,400 — because a wrong quantity buried confidently in a total is exactly the failure this system exists to end.

The estimator answers the questions, adjusts what he wants to adjust, and approves. The order that leaves the building is the one he signed — the agent has no path around him.

Then the order flows in the company’s own codes to the warehouse and the invoice, with no re-keying — which closes the second place errors used to creep in.

Quote turnaround went from most of a day to about forty minutes, which changes who wins the job: the builder who wanted a price by lunch now gets one, and the competitor quoting tomorrow is quoting too late. Ordering errors reaching site fell by roughly seventy percent, which the installers noticed before management did. And the rules that lived in two heads are now written down and owned by the company — which matters on the day one of those men is on holiday, and matters completely on the day he retires.

A consent set goes to council on a Thursday afternoon carrying whatever nobody caught. A stair tread two millimetres under the minimum. A balustrade drawn at 980 where the code wants a metre. A door that met its clearance in July — until the wall build-up changed on a different sheet in August, and nobody re-checked the door.

None of these mistakes is hard to find. Finding them just requires a person to hold eighteen sheets against a code document, clause by clause, on the exact afternoon the set is due — which is when the checking is at its thinnest and the coffee has stopped working.

The price of a miss is not embarrassment. It is a request for further information: weeks added to the consent, a client asking what happened, and a correction that now has to be chased across every sheet the change touches. One missed clearance can cost a project more calendar than the entire documentation phase saved.

The bitter part, every architect knows: the errors that trigger RFIs are almost never interesting. They are the mechanical ones — dimensions against limits, heights against minimums — the kind a tired expert skims past precisely because they are beneath his attention.

The agent reads a drawing set the way a reviewer does: against the clauses that apply to this building type — dimensions, heights, clearances, and the relationships that have to hold between elements on different sheets, which is where the August wall meets the July door.

It is built to make the architect’s review shorter, not to replace it, and the distinction is structural: the agent cannot issue a set, cannot sign anything, and its findings are an input to the architect’s judgement, never a substitute for it.

It is in build with Approach Architecture — running on real sets, not yet finished — and this page says so plainly rather than dressing a pilot up as a product.

A set is handed over and the agent reads every sheet, then returns a findings list ordered by severity.

Each finding carries three things or it is not reported at all: what is wrong, where it is, and the clause it fails. Not “balustrade may be non-compliant” but “balustrade on detail 7, sheet A-402: drawn at 980mm; D1 requires 1,000mm for this fall height.” The ten-second rule governs everything: if the architect cannot verify a finding in ten seconds, the finding costs more attention than it saves, and the agent stays silent instead.

The cross-sheet checks are where it earns its keep, because they are what a human review does worst at 4pm: the wall build-up that changed on one sheet and quietly broke a clearance on another, the level that moved on the section but not on the stair detail.

On the first eighteen-sheet set it returned a short list. Four of the findings were real and had not been caught — four RFIs that never got written. Several others turned out to be fine on inspection, which is expected and is exactly why a person reviews the findings rather than the agent issuing them: the architect burns a few ten-second checks; the client never sees a false alarm.

The architect works the list — accept, reject, amend — fixes what needs fixing, and the set goes out. What council receives is what the architect approved. The agent’s name appears on nothing.

Review time per set is down about sixty percent, and the four uncaught findings on the first set are the real ledger — each one an RFI that did not happen, weeks that did not vanish, a conversation with a client that never had to occur. The value concentrates exactly where the risk did: the Thursday afternoon when a deadline and a tired reviewer used to decide, between them, how thoroughly eighteen sheets got read. The reviewer still decides. He just no longer does it alone, at 4pm, from memory.

A surprising fraction of an architectural practice’s week is spent drawing things the practice has already drawn. Sheet setup for a new project. Title blocks. The standard threshold detail, the standard balustrade fixing, the standard wall junction — each appearing on every project with small variations, each rebuilt by hand every time.

It is work that demands real Revit skill and zero design judgement, and by a cruel default it lands on the most design-capable person in the room, because they are the one who knows the practice’s standards. On a typical project it runs to a dozen hours that nobody enjoys doing and nobody is proud to bill.

Every practice knows this, and the standard remedies each fail in their own way: the template library covers the cases that never vary, the junior needs the standards explained — by the senior, which was the time being saved — and the script breaks on the first project that differs from the one it was written for.

So the repeats persist, absorbing afternoons that were supposed to be spent designing.

The agent works inside the practice’s own Revit file, on the practice’s own standards, from an instruction written the way a senior would brief a graduate.

The interface is two tabs, and the second one is the important one: the conversation, and the output the agent proposes to make — shown before anything lands in the model. The architect reads what is about to happen, element by element, and nothing enters the file unreviewed.

And the honest line, stated as plainly here as on the product page: this one is in build with Approach Architecture, first release planned for Q1 2027. Nothing in this scene is running in production yet, and the figures below are targets, not results.

The instruction reads like a note to a graduate: “Set up the consent sheets for a single-storey addition — our standard set — and place the D-11 threshold detail wherever the new floor meets existing.”

The agent builds it in the model and then shows its working: these eight sheets created from the practice template, numbered to the office standard; D-11 placed at these three junctions, referenced on these sheets; this schedule updated to match. Every element is listed, checkable, and undoable — review is a scan of what changed, not a hunt through the file for what might have.

Where the standards do not cover a condition, it stops and asks rather than improvising — “the garage junction isn’t covered by D-11 or D-12; which do you want adapted?” — because an invented detail is worse than a missing one. The missing one announces itself. The invented one looks finished, and gets built.

The boundary is drawn in one sentence: it repeats decisions the practice has already made, and makes none of its own. It does not design, it does not choose between options, it does not touch anything it was not pointed at. The judgement stays in the chair it was always in — the work under the judgement is what moves.

The target is about twelve drafting hours returned per project — the afternoons currently spent rebuilding what the practice has drawn a hundred times. The aim is not more projects through the same office; it is that the person hired to design spends Thursday designing instead of numbering sheets. That is the promise on the table for Q1 2027, and this page will carry the real figures — not the targets — once the first release has run on real projects.