Live
Ada
Lead Counter
Counts every lead at the door: website forms, tracked calls, listing inquiries and landing pages. Tests are removed and repeat contacts are merged.
- Reads
- Lead email, landing page forms
- Never
- Replies to a lead
AI marketing team
Nine agents. One scorecard.
Most "AI marketing teams" are a list of prompts. This one has job descriptions, a number it answers to, and a person who signs off on everything. I built it for RTL Equipment, a heavy equipment dealership with five branches and an online parts store.
9
members live, running every morning
Counting, targeting, auditing, making
0
can publish, spend or send without a person saying yes
The rule
60/40
how the team is judged: machine leads, then parts revenue
The scorecard
How the work moves
Six steps, one of them human
Every week runs the same loop. The agents do the counting, the ranking and the drafting. Nothing reaches a customer until a person approves it.
01 · Sense
Count what happened
Leads at the door and parts sales, from the systems that already record them.
Ada, Tilly
02 · Decide
Pick the week's targets
The buyer questions the dealer is closest to winning, in AI answers and on Google.
Hunter, Brie
03 · Make
Draft the fix
A page, an ad change, a landing page, a graphic or a post, for one target at a time.
Paige, Penny, Dale, Hugh, Buzz
04 · Approve
A person says yes
Drafts wait in a queue. No agent can publish, spend or send.
The marketing lead
05 · Ship
Publish what passed
Done in the open, by the person who approved it.
The marketing lead
06 · Measure
Grade last week
Each target is checked the next Monday: moved, no change, or lost ground.
Brie
The roster
Each one has a job, and a thing it never does
A narrow job is easier to test and easier to trust. The line every member will not cross matters as much as the work it does.
Live
Tilly
Parts Counter
Reports parts store sales and orders by product and source. Her weekly totals match the store's own analytics to the cent.
- Reads
- The parts store
- Never
- Changes a product, price or stock level
Live
Hunter
Near-Miss Scout
Scores 109 buyer questions for how close the dealer is to being the answer: cited by some AI engines but not all, named without a link, or ranking just off the first page.
- Reads
- AI answer tracking, Search Console
- Never
- Writes or edits content
Live
Brie
Briefer
Builds the Monday brief, picks five targets split 60/40 between machines and parts, and grades last week's.
- Reads
- Only what the other three produce
- Never
- Pulls her own numbers
Live
Paige
Page Fixer
Drafts the fix for a target: why the page that wins the question wins it, what to change, and paste-ready copy. Every draft ends with a list of what still needs confirming.
- Reads
- The week's targets, the page being fixed
- Never
- Publishes
Live
Penny
Ads Watchdog
Audits the ad account each week against 83 checks, compares the result with the last audit, and proposes the top fixes. A fix that was not rechecked is never reported as resolved.
- Reads
- The ad account
- Never
- Changes an ad, a keyword, a budget or a setting
Live
Dale
Deals Builder
Watches the inventory for newly listed machines and drafts a deal kit for the best one each week: landing page copy, search ad copy and posts, built only from the listing.
- Reads
- The public inventory
- Never
- Builds or deploys a page, creates an ad, or posts
Live
Hugh
Designer
Makes three branded graphics for each deal kit, sized for feeds, portrait and link posts. He uses the listing's own photo and facts and no AI at all, so there is nothing for him to make up.
- Reads
- Brand files, listing photos
- Never
- Invents a machine, a spec or a price
Live
Buzz
Social Writer
Drafts a week's posts for Facebook, Instagram and LinkedIn, built only from fix drafts a person already approved, and shaped by what the page's own past posts earned. Engagement is his lane, and it is reported apart from the lead score.
- Reads
- Approved fix drafts, past post results
- Never
- Posts, or counts toward the lead score
A member is listed as Live only after it has run on real work. All nine have.
The rules
What keeps it honest
The interesting part of an agent team is what it cannot do.
Guardrails
Enforced by access where possible, and by code where not
263automated tests, plus an independent code review that found 15 problems before launch. All 15 were fixed.
- Counted at the door
- A lead counts when someone submits a form or calls. The score does not depend on anyone logging it in a CRM afterwards.
- Counting is code
- Lead counts and revenue are computed by plain code, so they come out the same every time. No number in the brief comes from an AI model.
- Read-only and draft-only
- The counters hold read-only access, and Paige has no access to any site. Penny is the exception: the ads sign-in on file is able to make changes, so her limit is in code. She can call four read commands, a test fails the build if a fifth is added, and applying a fix takes a person in the session.
- No guessing
- If the data behind a number is stale or missing, that section reads "Not available this week" and says why. It never shows an estimate.
- Private by default
- Customer details stay out of the code repository. The public sample below is a fictional dealer, behind a check that blocks the build if anything real appears.
- A writer with no tools
- The model that drafts gets text in and gives text back. It cannot browse, open a file or run anything, so a hostile web page can at worst produce a bad draft that a person rejects.
- Drafts checked in code
- Every draft is scanned for numbers and claims that are not on the source page. Ad copy is also checked for length limits and for phone numbers or email addresses. The flags travel with the draft into the approval queue.
- Graded every Monday
- Each target is checked a week later against the same score that picked it. A target nobody starts in two weeks is retired.
See it working
The Monday brief, with a made-up dealer
These are the real pages the team produces, filled with Northgate Equipment, a dealer that does not exist. Select a lead source to open its leads, or read the drafts in the approval queue.