Ask a question. Get your real numbers, with the rows behind them.

Owners ask the same handful of questions every week. Are we busy next week? Why was Tuesday slow? Which business unit is carrying us? Is the new CSR booking calls? Getting those answers usually means opening several reports, setting the same filters each time, and hoping both screens count things the same way.

Ask the business lets you type the question. The part that matters is what happens next: the AI never answers it directly. It turns your question into lookups against the same metric dictionary your reports use, and TradesWork computes the numbers.

Why the AI isn't allowed to write numbers

General-purpose AI assistants are good with words and unreliable with arithmetic over data they half-see. ServiceTitan's own help article for Atlas says responses may be inaccurate, incomplete or outdated, and that you should verify details before acting¹. That's honest, but an owner can't verify every answer, or there's no point asking.

So TradesWork splits the work:

1. 1

   Plan

   The model receives your question, the list of metrics in your dictionary and the labels of your business units, technicians, job types and campaigns. It returns up to four lookups: metric, period and filters. It's told to treat the question as data, not instructions.

2. 2

   Validate

   TradesWork drops any metric that doesn't exist or that you aren't allowed to see. Costing and pay metrics need their own permissions. Filters a metric can't use are reported back, not silently applied.

3. 3

   Compute

   Each lookup runs on the metric engine, the same code that fills the Reports page and dashboards.

4. 4

   Write

   TradesWork writes the answer sentence from the computed values. If a value is missing, the card shows a dash, never a zero.

The model used for planning is a fast, inexpensive one, and answers come back quickly. If the dictionary can't answer your question, you get one sentence saying what it can answer instead of a confident guess.

What an answer card shows

Each card carries the metric name, its plain definition, formula and date basis; the value for the period; the change against the previous period; how many records it's based on; the first five rows; and a link to the full drill-down. A tag says whether the metrics were picked by AI or matched by keyword.

The rows above are illustrations of the flow, not real results.

Demand vs capacity: the next four weeks

The second half of the page answers the question every owner asks on a slow Monday, a month too late: are we going to be busy?

Typical late bookings are the hours that usually get booked into a week after this point, averaged over your last 8 completed weeks. The seasonal factor nudges that up or down for the time of year, kept between 0.6 and 1.6 so one odd week can't swing it wildly. The forecast is compared with your capacity goal, 85% of available technician hours by default.

• Thin: forecast at least 8 hours short of goal.
• Full: at or above your full point, 100% by default.
• Confidence: low, medium or high, with a plain-language explanation of why.

All of these are adjustable: goal from 40% to 100%, full point from 80% to 150%, the thin threshold from 1 to 200 hours, and the horizon from two to four weeks. The forecast runs for the company and for each business unit, since a slow week for plumbing can hide inside a busy week for HVAC.

Thin weeks suggest audiences to work: unsold estimates, maintenance due and replacement candidates. Act on them through automations, the membership due list, or a call list for the CSRs.

Needs doing, by money first

The needs-doing list is short on purpose and sorted by dollars at stake:

1. Estimates untouched for 7 days or more, with the value at risk.
2. Overdue invoices, with the amount owed.
3. Failed membership charges, which turn into cancellations if nobody calls.
4. Thin weeks, with approximate revenue you'd expect to fill them.
5. Full weeks, so you can protect capacity for emergencies.

Where there's no honest dollar figure, the item shows none rather than a made-up one.

Watchlist and alerts

Each person can watch up to 25 metrics with one of four rules: goes above, drops below, changes by more than, or looks unusual. "Looks unusual" compares the value with earlier periods and needs at least four of them; sensitivity runs from 1 to 5. Periods are today, this week or this month.

Watches are checked hourly as their owner, with that person's permissions, and alert at most once per watch per period. Delivery is in-app, email or text. Email and text wait out quiet hours, 9 pm to 7 am by default, so a revenue alert doesn't wake anyone up.

The morning brief

Turn on the brief daily or weekly (Mondays), at the hour you pick, 7 am by default. It carries yesterday's numbers, today's appointment count, the top five needs-doing items and any watches that fired. It's the two-minute version of the page, for the owner who's on a job site by 7:30.

Questions worth asking in week one

A few that tend to change what owners do, all answered from the dictionary:

• "Average ticket by technician last month." The spread between your best and your newest tech is usually larger than you think, and the rows show why.
• "A/R over 30 days." Then open the rows and call the top five.
• "Callbacks this quarter by business unit." Return visits within 30 days, the cheapest quality signal you have.
• "Speed to lead this week." Minutes from a lead arriving to the first reply.
• "Gross margin by job type this year." Only if costs and burdened rates are set; otherwise the card shows a dash and tells you what fills it in.

Who this is for, and who it isn't

This page earns its keep for an owner or GM of a shop with roughly 5 to 50 techs, where the numbers move fast enough that weekly reports are too slow and the office is too busy to pull them on request. For a two-truck shop, the owner already knows the numbers; the forecast is still useful, the Q&A less so.

It's not a general chatbot. It won't write your marketing copy or explain a code requirement, and it won't answer questions about individual customers' equipment. For those, the right tools are elsewhere in TradesWork or outside it.