OPSMITHOS. WHAT IT ACTUALLY CHANGED.
Not a feature list. This is a 60 door operation, area by area, with the numbers that came out of it and the mistakes that got caught on the way. Names are out. Nothing else is dressed up.
DEPOSITS
BEFORE
Chasing was reactive. One deposit took 13 separate touches across 23 days. The payment report itself got skipped five times in a single month.
NOW
An audit runs at 08:00 and 16:00 against every arrival in the next 48 hours and flags anything missing or unverifiable, checked against real payment data rather than a field that does not reliably reflect it.
THE DRY RUN
Before it was even fully live, the first pass found five real missing deposits on live upcoming arrivals, two of them checking in that same day. The company had already eaten that exact failure once: no deposit taken, company paid. This time it was caught before anyone walked in.
The check was wrong in both directions before it was right, and both were caught. First it reported no deposit on a booking that held £250. Then, two days later, the opposite. That forced the verdict to become time aware, hold fired against scheduled against lapsed, instead of a yes or no that is right by accident half the time.
IF YOU ARE THE FOUNDER
The check you used to have to remember now happens twice a day whether you are thinking about it or not.
IF YOU ARE ON THE TEAM
A flag, instead of having to already know which bookings are the risky ones.
ENQUIRY TO BOOKING
BEFORE
Every enquiry meant a person running a ten prompt ritual: scan the database, check the calendar for the exact dates, work out capacity, hand copy the right card into the reply. Slow, and unreliable with it. One enquiry got answered fully booked all August, then minutes later in the same thread, wide open from the 3rd.
NOW
One call takes the area, the dates and the guest count and returns every viable property with live availability, price ranges, near misses labelled as near misses, and a pass it on fallback when nothing fits.
THE ONE IT GOT WRONG
A VA asked for something sleeping seven singles. The bot said nothing was in stock. It was wrong: one property split to eight singles and was free for the exact window. The founder would not let it go, the cause was traced, and it was retested against that original enquiry before anyone called it fixed.
Properties checked per enquiry went from 8 to 45, which killed a long tail of 26 that were silently never offered even when they were the best fit. Bed matching is deterministic rather than a judgement call, and was tested against all 44 active properties. 44 of 44 correct.
IF YOU ARE THE FOUNDER
The single biggest time sink in the operation, and it is now one message instead of ten.
IF YOU ARE ON THE TEAM
VAs quote with the confidence the founder would, because they are reading the same live data he would check himself.
MONTH END
BEFORE
Statements were reconciled from memory. Month end meant a morning of six blocked statements and no clear reason why any of them were stuck.
NOW
One sweep across every landlord, each statement either ready or blocked with the exact field that is missing, so the fix is a task rather than an investigation.
THE FEE FORMULA, CHECKED AGAINST REAL INVOICES
The fee maths was not trusted because it looked right. It was run against three landlords and 24 real bookings across every channel, and every single one landed on the exact contracted percentage. That is a formula matched to paid invoices, not a formula that seemed plausible.
The bot quotes only from the database price columns and never invents a number. One long stay instruction was found to be causing genuine under quotes, and the instruction itself was rewritten rather than that one quote being patched.
IF YOU ARE THE FOUNDER
You are no longer the only person who can tell whether a quote is defensible.
IF YOU ARE ON THE TEAM
A floor you can hold in a negotiation, instead of guessing how far you are allowed to move.
MAINTENANCE
BEFORE
A guest reported a fault and a tradesperson was booked. One job replaced a front lock at over £260 when the actual fault was the patio doors, and the spare keys were already in the database the whole time. Nobody looked, because looking was somebody remembering to look.
NOW
It reads the fault, checks the property's own history and hardware, and hands back the troubleshooting steps and the most likely fix first. It reads photos, so a picture of a leak or an error code is diagnosed rather than described. Most jobs close there, with a VA giving the instructions.
THE CHAIN
A fault comes in, so it diagnoses against that property's history instead of guessing, which means the obvious fix is tried first and most jobs never become a callout. When one genuinely does, the tradesperson arrives already knowing the fault, the model and where the isolator is, so it is one visit instead of two. Fewer callouts and shorter visits is a maintenance line that stops climbing with the door count, which is the difference between adding properties and adding overheads.
A standing cost cap lets small jobs proceed without waiting on a signature, so a £40 part does not sit in someone's inbox overnight. When that cap was raised above the contract's pre authorisation clause, the tension was surfaced rather than quietly absorbed.
IF YOU ARE THE FOUNDER
You stop paying for the diagnosis, the wrong part, and the second visit to fix the first one.
IF YOU ARE ON THE TEAM
A VA can actually resolve a fault now, with the steps in front of them, instead of forwarding it and waiting.
URGENCY
BEFORE
Everything landed in the same queue at the same speed, and how fast a dangerous message moved depended on who happened to read it and whether they understood what they were reading.
NOW
Danger is matched on meaning, not on keywords, and jumps every queue: it bypasses the normal path entirely and escalates immediately, at any hour, whether or not anyone is watching the channel.
THE CHAIN
A guest writes that they can smell gas, so it is recognised as an emergency in the same second rather than sitting behind eleven routine questions, which means the right person moves in minutes instead of the morning. That is the difference between an incident and a claim. One of those costs a callout; the other costs a property, a licence and a reputation you cannot rebuild by working harder.
This was tested before it went near a live channel, and the test caught a real hole: the matcher had been built from equipment words, so a smell of gas and water coming through a ceiling matched nothing at all. The two most dangerous phrasings would have received a calm, unhurried answer. Rebuilt to match intent, then retested 18 for 18 before shipping.
IF YOU ARE THE FOUNDER
The one category where being slow is unrecoverable is the one category that no longer depends on who is awake.
IF YOU ARE ON THE TEAM
Nobody has to decide whether something counts as an emergency at eleven at night.
IT COMPOUNDS
BEFORE
Whatever a VA worked out in a thread stayed in that thread. The same question got answered from scratch every time it was asked, by whoever happened to be on, at whatever standard they happened to have. Training was the founder answering the same thirty questions on a loop.
NOW
It teaches while it answers, and it learns while it teaches. Correct it once and that correction is permanent, from anyone, not just a founder. A nightly pass pulls durable rules out of the day's real conversations, so the operation's own standard is captured as it is set rather than reconstructed later.
THE CHAIN
A VA corrects it, so that answer is right for every person who asks it from then on, which means the standard only ever travels upward and nobody re-learns what the company already knows. Ask the same thing twice and it teaches the pattern harder the second time, so people stop needing to ask. Every month it needs less input and gives back more, and a team that is getting better on its own is the only kind you can grow without growing the payroll to match.
It was not always this way, and the gap was caught in a night. A VA taught it a real, durable fact that it used correctly that same day and would have forgotten by the next, because only founder corrections were persisting. Fixed the same evening. The same night turned up a second gap behind it: the founders' own direct messages, the single richest teaching signal in the system, were not being captured either.
IF YOU ARE THE FOUNDER
You stop being the ceiling. The operation gets sharper on days you are not in it.
IF YOU ARE ON THE TEAM
You are taught the reasoning, not handed the answer, which is the difference between asking again next time and knowing.
ESCALATIONS
BEFORE
The founders were the whole safety net. A booking taken at a fraction of the right rate, a two night stay five weeks out, a double accept that ended in a full refund: none were caught by a system, all three were caught because someone happened to be looking. A finished offer worth four figures a month sat unsent for 13 hours because nothing was pushing it.
NOW
Nothing arrives blank. Every escalation comes with the situation, what it has already checked, and the call it would make, so the founder is agreeing or disagreeing rather than starting from nothing.
THE CHAIN
An awkward call comes in, so it is escalated with a recommendation already attached, which turns a ten minute investigation into a five second yes. Twenty of those a week stop being twenty interruptions and become one short list. The founder's attention moves off the operation and onto the things only a founder can do, and an operation that no longer needs its founder in the daily loop is one that can take on more doors without taking more of the week.
The mandate is widened deliberately, one ratified piece at a time, with a worked example reasoned through properly rather than a blanket instruction. And a live prompt injection risk was closed before it was ever used: guest written text could previously reach write tools through the escalation path. It is read only now, found by asking what it could reach rather than by an incident.
IF YOU ARE THE FOUNDER
The most exhausting part of this business is being the only person allowed to say yes.
IF YOU ARE ON THE TEAM
Something to agree with, instead of a blank escalation and no steer.
CLAIMS AND DISPUTES
BEFORE
A guest claimed, and you argued back from memory under a deadline. Entry details could even go out before anyone had confirmed the deposit was secured, which happened on a live booking.
NOW
A release gate blocks door codes until the deposit is proven, and every interaction leaves a timestamped record: entry activity, staff messages at the time, mid stay check ins where the guest was asked and raised nothing.
THE CHAIN
A guest claims he could not get in, so the entry log and a cleaner's message from that hour are produced instead of a recollection, which means the platform decides on facts rather than on who complained hardest. That is a refund refused and a review removed. Do that consistently and disputes stop being a cost of doing business and start being something the operation reliably wins, which is margin you were previously donating.
Mid stay check ins are understood as evidence as much as service: a record that the guest was asked and had nothing to raise at the time. Guest facing sending is still shadow only, drafting but never sending, and stays that way until the ledger has been reviewed and the founder says otherwise.
IF YOU ARE THE FOUNDER
Revenue and reputation defended on the record, not on your word against theirs.
IF YOU ARE ON THE TEAM
Nobody is deciding whether to release a code against an unverified deposit. The gate decides first.
YOU CAN SEE IT WORKING
BEFORE
Nobody could say whether things were improving. The founders were intervening more than 20 times across three days and the only evidence was how tired they felt.
NOW
Every intervention is classified by root cause and reported daily: what it answered, what it assisted with, what still needed a founder and why. The target is stated out loud and progress against it is visible rather than felt.
THE CHAIN
Every founder touch is logged with its cause, so the reasons cluster into a list rather than a feeling, which means the biggest one gets fixed and that whole category stops coming back. The number falls, and because the cause is attached you know exactly which part of the operation just got stronger. That is a business improving on evidence instead of on effort, and it is the only version of this you can repeat at eighty doors.
It is published honestly, which is the only reason it is worth anything. Solo handle rate as observed: 14 per cent, 56, 71, 83. The 14 was a genuinely hard mix that should not have gone to a bot at all, explained rather than excluded. The meter also caught itself over counting twice, and data from before that fix is flagged as inflated rather than quietly smoothed.
IF YOU ARE THE FOUNDER
A number that is true on the bad days is the only kind worth having on the good ones.
IF YOU ARE ON THE TEAM
Less of did anyone tell the founder about this yet, hanging over every task.
THE OPERATION REMEMBERS
BEFORE
An SOP lived on a whiteboard, in the founder's head, or nowhere. Decisions were made in meetings and survived only in whoever attended. People left and took the operation's memory with them.
NOW
File a procedure and it wires itself in as a standing instruction with no code and no prompt edit. Recorded meetings, calls and boards are mined into that same corpus automatically, so what gets decided becomes what gets done.
THE CHAIN
A decision is made in a meeting, so it is captured and becomes a standing rule rather than a memory, which means it is applied to every relevant situation from that day on without anyone remembering to apply it. Onboarding stops being months of absorbing what is in the founder's head. A new VA is useful in days because the company's judgement is a thing they can query, not a thing they have to be told, and that is what makes the next ten doors cost less to run than the last ten.
It checks its own work. A two month retro mined 119 recorded meetings into hundreds of findings, and the first pass produced two fabricated claims: a recommendation the founders had explicitly ruled out, and a statistic that appeared in no transcript. Both were caught against the raw recordings and cut before the report was handed over. Where two sources genuinely contradicted each other, it flagged the conflict for a ruling rather than picking one.
IF YOU ARE THE FOUNDER
The company stops being one head with a team attached.
IF YOU ARE ON THE TEAM
You stop depending on whoever remembers the meeting where it was decided.
BUILT TO BE TRUSTED
BEFORE
Most operations find out a tool was wrong when a guest tells them. There is no log, no cause, and no way to be sure it will not happen again next week.
NOW
It fails closed. When it is unsure it asks; when a check does not return cleanly it stays quiet rather than guessing. Every answer is traceable to the data it came from, and a high risk change goes through an adversarial review before it is deployed, not after.
THE CHAIN
A payment tool was reviewed before it could write anything, so a refund path with no idempotency key was caught in review rather than in production, which means the double refund it would have allowed never happened and never appeared on a statement. Do that before every risky release and the failure rate stops being luck. That is what lets you hand it more of the operation, and how much you can hand it is the whole ceiling on what this saves you.
Where it has been wrong, the cause was found and named. It once told a VA several properties were free when one was booked solid. She pushed back and she was right. The cause was not the data: it did not know the current date, so a subtly wrong question returned confident nonsense. The date is now injected on every single turn. That is the class of fault worth fearing, and it is why the reviews happen before the deploy.
IF YOU ARE THE FOUNDER
You are not trusting a black box. You are trusting something that shows its working and stops when it is not sure.
IF YOU ARE ON THE TEAM
It going quiet, or asking instead of guessing, is the system behaving correctly.
THE POINT
Every number here came out of running 60 doors with a founder, a business partner and a small remote team. The failures are on the page for the same reason the wins are: an operation you cannot see the mistakes in is one nobody has looked at properly.