A family member knew the exact car she wanted. She didn't want to spend her weekends driving lot to lot, fielding callbacks from sales managers, or haggling under fluorescent lights.
So I built her an agent.
Two weeks later she drove home in the exact spec she wanted, $4,500 under what the first dealer quoted her, without ever picking up the phone.
The car wasn't the interesting part. The workflow was.
Step 1: Research the entire local dealer pool
The first job was inventory discovery. The agent pulled every matching listing from every dealer within driving range, normalized the data, and flagged the ones that actually matched her spec versus the ones that were "close enough" bait listings.
This sounds simple. It isn't. Dealer sites lie. Listings go stale. The same VIN shows up on three different sites at three different prices. The agent reconciled all of that into one clean shortlist with real prices, real options, real availability.
What used to be a Saturday of tab-juggling became a five-minute review of a single sheet.
Step 2: Anti-spam infrastructure before the outreach
Here's the part most people skip and regret.
The second you fill out a "request a quote" form on a dealer site, your phone melts. Sales reps share leads, BDCs auto-dial, and you get pulled into a week of voicemails from people you'll never buy from.
So before any outreach went out, I stood up dedicated infrastructure for her: a fresh email address routed into a filtered inbox, a forwarding number that funneled calls and texts away from her real phone, and rules that quarantined anything that smelled like a templated drip.
She kept her actual phone clean. The agent absorbed the noise.
This single layer is what makes the rest of the workflow livable. Without it, you're just outsourcing your own spam problem.
Step 3: Slack triage with one-tap reply buttons
Every dealer reply landed in a Slack channel as a card. Each card showed the dealer, the offer, the delta from the best price so far, and three buttons: Counter, Pass, Ask for OTD.
She didn't write emails. She didn't draft texts. She tapped a button. The agent wrote the reply in her voice, sent it from the dedicated email or number, and logged the exchange.
A negotiation that normally takes a week of phone tag took her about ten minutes of total attention spread across two days. Most of it during a coffee break.
The buttons matter more than they sound. Friction is what kills personal automations. The second a workflow asks you to context-switch into "compose mode," you stop using it. One tap keeps the human in the loop without making the human do the work.
Step 4: A live leaderboard in Google Sheets
Every offer, counter, and concession landed in a single Sheet, sorted by out-the-door price. As dealers moved, the leaderboard re-ranked in real time.
This did two things.
One, it gave her perfect visibility. At any moment she could see who the leader was, by how much, and what the next-best fallback looked like if the leader flaked.
Two, and this is the leverage part, it gave the agent a number to push against. When the third-place dealer asked what it would take, the agent knew the exact figure to name. No bluffing, no guessing.
The leaderboard turned a fragmented set of conversations into a single competitive auction the dealers didn't know they were in.
The outcome
She bought the exact car she wanted. $4,500 under the first quote. No showroom visit until pickup day. No haggling. No callbacks to dodge.
The savings are nice. The hours she didn't spend are nicer.
The bigger point
We talk about AI agents like they're a future enterprise category. They're already a present-day household one.
The car-buying agent isn't a product. It's a weekend project that paid for years of API credits in a single transaction. The same pattern works for contractor bids, insurance quotes, apartment hunts, medical billing disputes, anything where the work is "contact a bunch of people, normalize their answers, and pick the best one."
The interesting question isn't whether agents will handle this kind of work. It's how many of these workflows you're still doing by hand because nobody's bothered to wire one up for you yet.
What's the next one you'd build for someone you love?