# Jobber "Marketing Agent" — concept prototype **Built by Corey Brown · July 2026** **For:** Jobber, Director, Product Management (Marketing Agent) application **Live prototype:** https://coreywbrown.com/lab/jobber-marketing-agent/ **Status:** Unlisted concept. Not affiliated with or endorsed by Jobber. The business, jobs, reviews, campaigns, and results are invented sample data. One thing is real: the "why now" weather reading is a live pull from Open-Meteo for Toronto. The interface is built on Jobber's own public Atlantis design tokens. --- ## Why this exists The role is Jobber's first agentic AI product, a Marketing Agent for home-service pros. Rather than describe how I would approach it, I built a working version of the thing the job is about. It is the same move I made for HelloFresh, Owner, and HouseSigma: take the real problem, ship a small honest fix, and let the thinking be visible instead of described. When the product itself is an AI agent, a prototype is not a side-demo, it is the audition. ## The problem A service pro finishes great work in a neighbourhood full of potential customers, and nobody nearby knows the company exists. They understand marketing matters. They have no time, no marketing hire, and no patience for tools that add work. Jobber already ships a lot of marketing surface: reviews, referrals, campaigns, job showcases, a website builder, Google Business Profile management, an AI marketing plan. The trouble is that it is a grab-bag of separate signals and channels, and the AI layer sits on top as recommendations the owner still has to assemble, judge, and approve one by one. Their own product copy is honest about it: "nothing goes out without your sign-off." That promise is correct, but the way it is delivered hands more work to the person whose entire constraint is that they have none. The landscaper who could not find twenty minutes to write a post now has a queue of drafts to review. ## The insight The automation everyone reaches for first is automating the send. That is the wrong target, and it fights Jobber's own promise that the owner stays in control. The work a service pro actually cannot do is the thinking. Which of this month's jobs is even worth marketing. Who to reach, and when the moment is right. What to lead with. Whether an offer is worth the margin. Whether last month's campaign paid for itself. That judgment is a marketing manager's job, and it is exactly what a small operator has neither the time nor the training for. So the agent's job is to combine every signal Jobber already holds, reviews, job history, showcase photos, online presence, seasonality, local weather, neighbourhood activity, and the results of past campaigns, into a small number of educated, cost-justified campaigns. It does the marketing manager's thinking. The owner keeps doing the work, and keeps the send button. The automation is in identifying the optimized campaign, not in pushing it out the door. This is also where the "intelligence central to the value, not a layer on top" test is actually met. Any language model can write a Facebook post. Only Jobber's agent can say: of your ten jobs this month, this is the one worth a campaign, to these homes, for this cost, because these signals line up and this is what worked for you last time. ## The solution **Marketing Agent**, a surface that turns a month of finished work into a single, reasoned recommendation. - **A suggested campaign, not a queue.** The agent watches jobs close and conditions change, and surfaces the one campaign worth running now: a neighbourhood waterproofing push, triggered by a real job on Ellis Ave and the wettest week in a decade, aimed at the twelve hundred homes most likely to have a wet basement. - **The reasoning, shown.** It lays out how it built the campaign, what it is known for from its own reviews and job record, why this week, and why these homes, so the recommendation is legible, not a black box. - **Cost against return, up front.** Every recommendation carries its own math: what the campaign costs to run across mailers and ads, against the jobs and revenue it is projected to book, with an honest sub-one-percent booking rate labelled as normal for local outreach. The pro is never asked to spend without seeing the trade. - **The jobs it held back.** Of ten jobs closed, three became campaigns and seven did not, each with a reason: a warranty call-back is not a win to show off, a client whose invoice is in dispute is not someone to feature, and the homes that actually flooded are not an audience to advertise to. Knowing which jobs are not worth a campaign is the same intelligence as knowing which are. - **It gets sharper.** The agent shows what it has learned about this specific business, that photos beat text, that Tuesday evenings land, that discounts do not move this owner's conversion, and folds those lessons into the next recommendation. Sending a campaign teaches it; holding one back teaches it too. - **Control without a countdown.** The campaign is scheduled, reviewable, editable, and reversible before and after it sends. There is no pressure timer, because real control should not cost the owner their vigilance. - **See it before it sends.** One click previews the exact Facebook ad, postcard, and email, each rendered natively, so the owner validates the look and feel before anything goes out. ## Where it fits in Jobber's AI Jobber's shipped AI already runs a clear gradient of autonomy: Recommendations that only advise, Rewrite and Voice that act when asked, Automations that run rules the owner switches on, and the AI Receptionist, the one feature that acts on its own, answering and booking inbound calls. The Marketing Agent is the same trusted-autonomy pattern crossing to outbound, where a mistake is public, permanent, and reaches many people at once. That is why the design keeps the send decision with the owner and puts the autonomy where it belongs, in the analysis. It extends a pattern Jobber has already proven rather than importing a new one. ## Why it maps to the role The role owns the Marketing Agent end to end and is measured on commercial outcomes, activation on marketing features, and customer growth. This is built as an activation play. Jobber's own data says customers who use the marketing tools grow faster, so the bottleneck is not whether the tools work, it is whether a time-poor owner ever assembles them into something worth sending. An agent that does that assembly, and proves the cost-benefit each time, is a direct lever on activation. It is also the growth engine I ran at theScore, pointed at a new audience: take a large base with real but hard-to-read intent, and convert it by doing the judgment work for them. And it answers the hardest thing the job asks for, defining quality when a product acts on a customer's behalf, with a specific position: at the outbound frontier, quality is not the content the agent produces, it is the accuracy with which it decides what is worth doing and what to leave alone. ## How it was built - A single self-contained HTML, CSS, and JavaScript file, no framework and no backend, built on Jobber's public Atlantis design tokens so the surface reads as a native Jobber screen: the fourteen-pixel Inter type, the warm taupe canvas, the workflow color system, the real component dimensions. - The "why now" weather is a live call to Open-Meteo for Toronto, so at least one signal is genuinely real. Everything else is illustrative sample data. - The numbers are modelled to be realistic rather than flattering: twelve hundred homes reached, a click-through just over one percent, a booking rate under half a percent, and a return around seven times spend, with the win framed as low cost per booked job, not a fantasy conversion rate. - It is deliberately honest about its limits, and says plainly that it runs on simulated data. For a company whose voice guide values sincerity over enthusiasm, overclaiming would have been the wrong build. ## What I would do next with real data - Wire the signals to the real thing: derive "what you are known for" from the owner's actual jobs, invoices, and reviews rather than a sample, and pull neighbourhood demand from Jobber's own cross-account data, which no standalone marketing tool can see. - Make the learning loop real. Measure each campaign's booked revenue against its cost, attribute it back to the owner, and let the agent's recommendations genuinely improve per business over time. - Make "campaign accepted" and "campaign held" both first-class training signals, so the agent learns each owner's taste, not just what converts in aggregate. - Treat activation on the agent's recommendations as the leading metric, and test the honest cost-benefit framing against acceptance. Telling an owner when not to spend builds the trust the product runs on, and the tension with short-term feature usage is real and worth measuring.