Window cleaning round management with AI: rain days and unpaid cleans.

A window cleaning business sells the same modest job hundreds of times over. Each address has a frequency agreed with the customer, a place in a street-by-street order and a price that matters only when it is added up across the whole round. Two things knock that arrangement over again and again. Rain takes away a day, and the streets left over need a new order. And much of the work can be done while the householder is out, so payment comes later, by whatever arrangement the customer has, or does not come.

On both sides of that, supervised AI assistants can carry the paperwork: a proposed sequence for the rest of the week after a wet day, a reconciled list of cleans still unpaid with reminder drafts for the owner to send, and a file of the documents that commercial height work depends on. The weather call, the access method a building needs, whether an anchor or a ladder is fit to use, and whether a customer who never pays stays on the round are settled by the owner and the people doing the work, never by a Bot.

This article says round, the word British and Australian window cleaners use; in American usage the same list is usually a route. Either way it lives somewhere: in round-management software, a spreadsheet or a paper book. Whatever holds it is the record, and nothing an assistant remembers replaces it.

What sets the trade apart is the arithmetic. A single lost afternoon moves the next-due date of every house on the streets that were not reached, and those houses were already promised a rough week. A single unpaid clean is too small to chase by hand, and a few hundred of them, left unchased across a year, add up to money the round has already earned.

The fields an assistant needs from each entry, all of them kept by the owner:

• FREQUENCY AND LAST CLEAN

  The interval the customer agreed and the date of the last visit, from which the next due date follows.

• POSITION IN THE ROUND

  The street and the order within it, which is what a re-sequence moves.

• ACCESS NOTES

  Side gates, a dog in the garden, a conservatory roof, parking for the van. The owner writes these; the assistant only carries them forward.

• METHOD

  Pole from the ground, traditional work, or an address flagged as needing a different access method, which is the owner's decision recorded once.

• PAYMENT ARRANGEMENT

  How this customer pays and how long after a clean the owner expects it, so a late payment can be told from a normal one.

Whether a day is lost to weather is decided at the van. Where work at height cannot be avoided, HSE's window cleaning page, read on 16 October 2026, asks whether there are procedures to stop it in adverse weather, giving high winds as its example. An assistant reading a forecast is not that procedure, and it is never asked to be.

What an assistant needs is not a forecast but the owner's rules for catching up, in writing. In Aevrion's pattern the owner creates a Round Planner Bot in Grok Bot and gives it those rules in chat to remember, which is how the documentation says rules that should stay true are set. That makes them guidance the Bot works to, not a lock on what it can do.

With the rules written, an evening routine turns the owner's note of finished streets into a proposal: which houses move, by how many days, and whose clean will now be later than promised. It changes nothing in the round software until the owner allows it. Reading that software is a custom integration or supervised browser use where the vendor permits it, subject to testing. The solo round guide follows a wet week of this in practice, from the overnight plan to approving the new order from a phone between houses.

Customers whose clean moved can be told by email, drafted for the owner to edit and send or throw away. The documentation describes no telephone calls or text messages, so a customer who rings about a missed week is the owner's to answer.

The catch-up rules a Round Planner cannot work without, each one the owner's to set:

• FIXED DAYS

  Shops and offices that expect the cleaner on a set day keep that day, and domestic streets move around them.

• SLIP LIMIT

  How many days past its promised week a house may move before the customer is told.

• CATCH-UP ORDER

  Whether unfinished streets go to the front of the next working day, to the end of the week, or are spread across both.

• CYCLE RESET

  Whether a late clean pushes the next due date back with it, or the address returns to its original rhythm.

• SAME AREA, SAME DAY

  Which streets belong together, so a re-plan does not send the van across town twice.

Based on the Grok Bot documentation at docs.x.ai as checked on 15 October 2026. Its statements on Bot rules, Bot-to-Bot messaging, routine triggers, connected apps, approval expiry and payments were rechecked against docs.x.ai and Cursor's help pages on 26 October 2026. The trade examples are Aevrion's illustrations, not documented Grok Bot use cases. Aevrion Ops is an independent consultancy. It is not affiliated with, endorsed by or a partner of xAI, SpaceXSI or Cursor. Grok Bot is named here only to describe it.

Unlike a trade that works inside the house, a window cleaner can finish the whole job with nobody home. The clean can happen on a weekday morning with the house empty, a note may go through the door, and payment follows on whatever terms the customer has with the owner. When one of those payments does not arrive, nothing visible happens: the next clean comes round, and the balance doubles.

Before any chasing there is a matching problem. A transfer arrives under a partner's name or with no reference; one payment covers two cleans; cash handed over on a doorstep is written down later; card payments land in a provider's batch. A Collections Bot, built to Aevrion's design, matches the cleans the owner marked done against an export of payments the owner downloads and saves to the shared workspace, and anything it cannot match by the owner's rules goes on a query list instead of being guessed.

That export matters for another reason. Every Bot on an account works on one shared computer with the same signed-in sessions, and the documentation says separate Bots are not a security boundary, so a bank login does not belong there. Where the owner wants a login kept apart, the documentation's answer is a separate user, which brings its own computer.

Reminder drafts, and the decision about a customer who still will not pay, are shown week by week in the solo round guide. The point here is narrower: a reminder sent to someone whose transfer simply went unmatched costs more goodwill than the balance was worth, so the query list is cleared first. The Bot never takes a payment and never removes a customer.

1. 1

   Cleaned

   Owner

   The owner, who marks the visit done in the round software

   Advances when

   The payment window the owner set for that customer passes

2. 2

   Awaiting a match

   Owner

   The Collections Bot, reading the latest payment export

   Advances when

   A possible payment is found that does not fit the owner's matching rules, or nothing is found

   Ends when

   A payment matches by the owner's rules and is marked paid

3. 3

   Query

   Owner

   The owner, shown the candidate payment and the customer's history

   Advances when

   The owner confirms that nothing came for this clean

   Ends when

   The owner assigns the payment to the clean

4. 4

   Unpaid

   Owner

   The owner alone

   Advances when

   Nothing automatic: the owner decides whether a reminder is drafted, and later whether the address stays

   Ends when

   Payment, or the owner's decision written into the round record

In Great Britain, HSE's window cleaning page (as checked on 16 October 2026) puts avoiding height first, with water-fed poles from the ground or cleaning from inside, and treats ladders and fall arrest as later steps, not defaults. In the United States, OSHA's rope descent rule, 29 CFR 1910.27, requires written anchorage information from the building owner before any descent; where a state runs an approved plan, as California does through Cal/OSHA, the plan's rules apply.

The companion guide on height-work site files goes through both sets of rules and the file a firm keeps for each commercial site. The assistant's part is narrow on purpose: it holds the dates and documents the owner enters and says when one is missing or out of date. Whether a building, an anchor or a ladder is safe to use is a competent person's decision, and the access method is the owner's.

The documentation advises starting with the smallest roster that works and adding a Bot only when a job has its own goal, tools, approval boundary or schedule. A window round has three such jobs, and they are best added one at a time, in whichever order hurts most. The names and roles below are Aevrion's pattern; Grok Bot provides named Bots, routines and drafts, not these particular assistants.

A routine that did not run leaves no message. The documentation describes a history of recent runs, with failures shown and a reason, but no automatic second attempt and no alert; the app keeps the twenty most recent records for each routine. On a wet morning the risk is plain: the owner drives to yesterday's order because no proposal arrived and nobody noticed it was missing.

The planner ends in a file rather than a change, and the owner applies or edits it in the morning. So the design gives the owner one habit rather than another tool: before loading the van, look for the Round Planner's proposal. If it is not there, open the routine's history, read the reason, and plan from the round software directly. A file that names the date of the round data it used makes a stale proposal easy to spot.

The solo round guide is where the re-plan and the chasers are shown in motion: an illustrative one-person water-fed pole business whose Monday plan is undone by rain, with approvals given from the van and the non-payer decision kept with the cleaner. It also shows handoffs between Bots, which a one-person round can use but does not need on day one.

The height-work site files guide is for firms taking on commercial buildings. It sets out what the HSE page and OSHA's rope descent rule ask to be in place before work starts, and how a Site File Bot chases a missing document without ever judging an anchor.

This site's readiness assessment asks for no payment and no sign-up. It takes one workflow, rates it on twelve fixed dimensions, among them data availability, manual re-entry, exception frequency and ownership clarity, and answers on the spot with one of four verdicts (from not ready yet to a high-leverage candidate), the blockers behind it and a suggested next step.

Matching payments to cleans is an honest test for a window round. Exception frequency will rate poorly if many transfers arrive under the wrong name, and manual re-entry if cash goes on paper first. If the plan assumed a bank login on the Bots' computer, security and compliance readiness will score low, and that single score holds the whole verdict down, as will a low mark for rule clarity or ownership clarity. With a usable verdict in hand, Aevrion Ops can sit down with the owner over the round software's export, the payment export and the written matching rules, and check whether the vendor's terms allow a tested integration.

1. A window round fails in two familiar ways: rain leaves streets unreached, and customers who were out when the clean was done do not always pay.
2. A Round Planner Bot can propose a new street order each evening; the owner decides whether a day is lost and whether the new order is applied.
3. A Collections Bot can match payments to cleans and list the ones it cannot place, but it never takes payment and never drops a customer.
4. HSE's order for height work starts with avoiding it, and OSHA's rope descent rule needs written anchorage information from the building owner before use.
5. The assistants keep dates, drafts and lists; every safety call stays with the owner and the competent people doing the work.