An assistant that knows what needs to be done
Wipjet’s assistant knows what needs to be done. It reads your live jobs, schedule, estimates and money, flags what needs attention in a brief written in plain sentences, rewrites it as the day changes, and links every notice to the thing that fixes it. Ask about your own operation and it answers from the same records, showing every lookup. Ask it for an estimate or an invoice and it drafts one for you to check.
A brief that rewrites itself as the day changes
You land on a short list of what needs you right now, not a wall of charts: one sentence each, leads first, then the work and the schedule, then the money. Every note links straight to the place you can act: the inquiry that has waited longest, the job that is finished and not yet billed.
The leads notes watch for inquiries nobody has answered, estimates that have sat out too long, and accepted work with no date yet. The work notes watch for a day with nobody assigned, parts still open, and a technician’s raised hand: a job flagged not complete. The money notes watch for jobs ready to bill and invoices aging.
This is not a report that goes stale by ten o’clock. When your data changes, the brief rewrites. Accept two leads and book one job, and the list now says so. And it stays short on purpose: a note leaves the list when the thing is fixed, so the brief reads as a briefing, never a backlog.
What’s urgent is decided by rules, not by a model’s mood
The part worth being skeptical about is the part we made boring on purpose. Which items surface, and in what order, is decided by deterministic rules — plain, testable code, not a model’s judgment — rules that put an unanswered inquiry above pipeline housekeeping and a declined booking above nearly everything. The AI’s only job is to write the sentence. It cannot invent an emergency, and it cannot quietly drop one.
Ask it anything about your own operation
The same assistant sits in the header of every office page. Ask what’s still unscheduled this week, what a job was billed at, which estimates are sitting unanswered, what parts the techs have asked for — it answers from your live customers, schedule, estimates, slips and parts. And it shows its work: every lookup it runs appears as a visible step, ticked as it lands, so an answer is never a black box. Links in its answers go to the real page inside the product, where you can act on what it found.
Ask it how do I (how do I take a deposit, how do I reopen a closed job) and it answers by quoting the same help pages a person would read, rather than improvising from memory. One assistant, both kinds of question: what is going on, and how do I do this.
Ask it for the estimate
Ask for an estimate on a lead and it starts one the way you would: from the template that fits, with the questions about the house answered from what the customer wrote, every part priced from your rate sheet or marked for a price. Once the work is done, ask it to invoice the job and it drafts the invoice from the signed estimate, the change orders and the deposit, and sends it only after you say yes. It works in Wipjet, and inside Claude or ChatGPT if that is where you already are. It only ever writes drafts: nothing reaches a customer until you say so, and it never touches an estimate or invoice that has already gone out.
What does the AI assistant actually do?
Three things. It writes a brief of where the business stands (leads, then the work and the schedule, then the money, one plain sentence each with a link to the place you can act) and rewrites it as your data changes. It answers questions about your own operation from live data: customers, schedule, estimates, slips, parts and money. And it drafts: ask for an estimate on a lead and it starts one the way you would, from your templates and your rate sheet; ask it to invoice a finished job and it drafts the invoice from the signed estimate, change orders and deposit, and sends it only after you say yes. It never makes up a price, and it never touches a document that has gone out.
How do I know what it tells me is right?
Because the judgment isn’t left to the model. What surfaces in the brief, and in what order, is decided by deterministic rules — plain, testable code — and the AI only writes the sentences. When you ask a question, it looks the answer up in your live data and shows each lookup as a visible, ticked step, so you can see where the answer came from.
What can I ask it?
Anything you could work out yourself from your own records: what’s in the inbox, what’s still unscheduled this week, what a job was billed at, which parts requests are open. And “how do I…” questions — those it answers by quoting the same help pages a person would read, rather than improvising from memory.
Who can ask my assistant for answers?
Everyone in the office: owners, admins and office staff. Anyone who signs in to the office side can ask, and they get answers from the same live data they could look up themselves. The field app is a different surface and doesn’t carry the assistant; techs get their job, their forms and their paperwork there instead.
Is this the same assistant my customers talk to?
No. The assistant in this chapter works for you, in the office: the brief, the answers and the drafts. The assistant on your website talks to prospects and captures leads, and it answers from the knowledge base you wrote, not from your books, so it can’t promise something you never put in writing. They are kept separate on purpose.