Bid Season Tripled the AI Bill. The July 2 Claude Enterprise Update Is How a Janitorial Owner Caps It.

You have heard the objection in your own voice: "I am not paying for something I cannot see." Fair. But in most cleaning companies the problem is not that the AI is a toy. The problem is the opposite — it is working, your two account managers are quietly using it for everything, and you have no idea what it costs, who is using it, or what got typed into it about the Service Contract Act wage determination on the federal building.

That is a bookkeeping problem, not a technology problem, and as of 2 July 2026 it has a bookkeeping answer.

September to November is when the meter runs hottest

Every building service contractor knows the autumn shape. Property managers set next year's operating budgets, the re-bids land, and the RFPs from school districts, county facilities departments and the big brokerage-managed portfolios all arrive inside the same eight weeks. Mandatory pre-bid walk-throughs on Tuesday and Thursday mornings. Square footage takeoffs off floor plans that are four years out of date. Production rate assumptions to defend. Staffing plans, transition plans, references, insurance certificates, a W-9, and for the public work a wage determination you have to price correctly or you lose money for three years.

This is the season when an estimator discovers what these tools can actually do, and it is the season when your bill quietly triples. Not because anyone did anything wrong — because a 60-page RFP response is genuinely a lot of work, and the machine will happily do as much of it as you let it.

Here is the sentence to keep: a spend limit is a hard monthly ceiling you set per person and for the company as a whole, and when spending crosses 75% and again 90% you get an email, and at the ceiling the work stops instead of the invoice growing.

What the July 2 governance update actually put on the desk

The Claude Enterprise governance update on 2 July 2026 added four things an owner cares about, and no things a developer cares about. First, a usage and cost dashboard: who used what, when, and what it cost. Second, spending limits at both the company level and the individual level. Third, alerts at 75% and 90% of whatever limit you set. Fourth, model defaults and entitlements — meaning you decide which people are allowed to use the expensive model, and which people get the cheap one that is perfectly adequate for rewriting a scope of work.

There are also reporting connections so your bookkeeper can pull usage into the same spreadsheet where the rest of the overhead lives, rather than logging into one more portal. The practical upshot: AI stopped being a surprise on the credit card statement and became a line item you budget, cap, and review — the same way you already treat consumables, uniforms and equipment rental.

flowchart LR
  A["Monthly AI budget set by role"] --> B["Estimator drafts the RFP response"]
  B --> C["Spend crosses 75% — alert to owner"]
  C --> D{"Is it bid season?"}
  D -->|Yes| E["Raise the estimator's cap, write down why"]
  D -->|No| F["Hold the cap, pull the usage report"]
  E --> G["Monday cost review with the ops manager"]
  F --> G
  G --> A

Who should be allowed near the wage determination

Entitlements are the half of this update that owners skip past, and in janitorial they matter more than the spending caps. Think about what actually sits in your files. Service Contract Act wage determinations and fringe rates on federal work. The 32BJ or Local 1 collective bargaining scale if you are union. I-9 documents. Your experience modification factor and the open workers' compensation claims underneath it. In California, your registration with the Department of Industrial Relations under the Property Service Workers Protection Act and the harassment-prevention training records that go with it.

None of that belongs in a general-purpose chat that every account manager can open. With entitlements you can put the payroll clerk and the controller in one group that is allowed to work with payroll exports, and put the account managers in another group that can draft tenant communication and inspection summaries and nothing else. It is the same instinct that made you lock the key cabinet: not distrust, just a smaller blast radius.

A budget that survives a forty-bid autumn

Assumptions, all illustrative: you chase 34 bids between 1 September and 30 November. Two account managers and one estimator do the writing. Your estimator currently spends about nine hours per response on the parts that are pure assembly — reformatting the scope of work into the customer's template, building the staffing grid, restating the transition plan, chasing the certificate of insurance language.

Now set the cap from that table, not from a feeling. Thirty-four bids at $11 is $374 across three months, so a $250 monthly ceiling for the estimator holds with room, and the 75% alert lands before the ceiling does. If the estimator hits it in week two of October, that is not a violation — that is information. Either the bid volume is higher than you planned or somebody is using it for something you did not budget, and the usage dashboard will tell you which within about four minutes.

The real return is not the $6,970. It is that a shop with one estimator can respond to 34 RFPs instead of 22, and in a trade where every contract carries a 30-day cancellation-for-convenience clause, the only defense against churn is having more at-bats than you lose.

The cap you should not set

Do not put a hard ceiling on the thing that is winning work. Owners with a cost-control reflex tend to cap the estimator first, because that is where the biggest number is. That is backwards. Cap the places where usage is casual and unmeasured — general chat for people whose job does not require it — and leave headroom where the spend converts into bids submitted.

Also, do not set one company-wide number and call it governance. Different roles have different shapes. Your recruiter posting in three languages every August will spend more in August than in March, and that is correct. Set the limit per role, per month, and review it on the same Monday you review labor hours against budget.

What a spending cap will not save you from

A cap controls cost. It does not control judgment. It will not stop an account manager from sending a tenant a polished email that promises quarterly carpet extraction which is not in the specification, and that promise will follow you for the life of the contract. It will not catch a bid where the machine assumed 4,000 cleanable square feet per hour on a building that realistically runs 2,600 because of the security screening at the door.

Keep a human on three things regardless of budget: any number that goes into a bid, any sentence that changes what you owe under a contract, and anything touching an employee's pay or immigration paperwork. And keep your own eyes on state rules — Texas TRAIGA and California SB 53 both took effect on 1 January 2026, and federal preemption is still unsettled as of this July, so whichever states your buildings sit in, that state's law binds you.

Frequently asked questions

What is a realistic monthly AI budget for a company doing $5M in janitorial revenue?

Build it from roles, not from revenue. Estimator and proposal writing is the big one; account managers writing tenant and property manager communication is second; recruiting posts and translation is third. Most owners land in the low hundreds per month outside bid season and roughly double that inside it. If you are being quoted thousands, ask precisely which of those three jobs is consuming it.

My controller wants this in the general ledger. Where does it go?

Put it in overhead alongside software and telephone, and if you can, split the estimator's portion into selling expense. That split matters at renewal time, because when you can show what proposal support costs per bid, you can also show your win rate against it — which is the argument for keeping the budget when the year gets tight.

Can I stop people from using their personal accounts instead?

Not entirely, and pretending otherwise is how shadow usage starts. The reliable move is to make the company account better than the personal one: it has your specification book, your dilution chart, your bid template. People route around controls that make their job harder and accept controls that make it easier. Write a one-page rule about wage data and employee files, have everyone sign it, and enforce that part seriously.

Does the 75% alert go to me or to the person spending?

Send it to both. The owner-only version turns you into the police. The user-only version means you find out at the ceiling. Both means your estimator can walk into your office in the middle of October and say "I am at 78% and I have eleven bids left" — which is the conversation you actually want.

One line item worth measuring the same way is the phone. In this trade the inbound calls that matter — a property manager with an emergency, a tenant complaint that starts a cure clock, a prospect who found you at 8:20 PM — arrive when the office is dark. CallSphere builds AI voice and chat agents that answer the line and the web chat 24/7, book the walk-through, and capture the lead with the building details attached. Like everything else in this post, price it as a monthly line, review it against bids booked, and cap it if it is not earning its keep.