Two AI Employees Doing Eight Jobs
How we gave a lean design marketplace two AI employees, one inside the company and one on the storefront, each with a name, a job description, tools, standing rules, and a manager, so the coordination, reporting, and follow-up that fell between roles has an owner.
8
Jobs the two employees hold
2
AI employees, each with a job and a manager
90 days
From kickoff to both employees live
CLIENT OVERVIEW
A curated marketplace for collectible design, art, and fashion from independent sellers, which also runs in-person collector events. The team is lean and flat: the CEO sets the goals, leads run marketing, listings, partnerships, and logistics, and ops and support staff keep orders moving. Pieces ship from sellers and warehouses, often across borders, so one order can involve a seller, a buyer, a carrier, and customs. The company wanted help that could hold a whole job, not one more tool to check.
THE PROBLEM
What was broken
Work fell between roles
Chasing an unanswered thread, building the weekly report, curating the newsletter, closing out a stuck order: this is the work a coordinator, an analyst, or a manager usually carries. On a lean, flat team, it fell between roles and got done when someone had time. The weekly report was built by hand, and the newsletter ran on hand-run scripts.
Single-purpose bots leave a person in the middle
A bot that answers one kind of question or writes one report solves one task. Someone still has to notice its output, connect it to the next step, and follow up. Each new bot adds one more thing for a person to watch.
Every automation came with its own interface and rules
Each new automation needed its own screen, its own instructions, and its own idea of what it was allowed to do. Nobody could see in one place what the automations were doing, or who they answered to.
THE GOALS
What success looks like
Fill whole jobs with two AI employees instead of a growing pile of single-task bots
Give each employee what a person on the team has: a name, a job description, tools, standing rules, and a manager
Reach people where they already work: the admin, email, the CEO's chat assistant, the seller portal, and the storefront
Keep a person in charge of what matters, with previews, approvals, caps, kill switches, and logs
OUR APPROACH & SOLUTION
Hiring AI employees instead of building more bots
We started small and grew one employee a role at a time. The project kicked off on July 16, 2026, with an assistant in the admin sidebar. Each new job after that became a role on the same employee, with its own tools and playbooks and one memory across all of them, instead of a new bot with a new screen. The admin assistant, the chief of staff, and the shopping companion in our other case studies are each part of these two employees. A new desk runs in shadow mode before it goes live, and any change to data goes through a preview, a confirm, and a log. Next on the plan is a customer service desk for the storefront employee that covers the support inbox and text messages.
Phase 1
A Read-Only Ops Assistant
The project kicked off on July 16, 2026, with an assistant in the admin sidebar. It answered ops and support questions from the live database and the code, and it went live in August 2026 with read-only access.
Phase 2
Chief of Staff
From late August, the same employee took on a second role: personal check-ins for every team member at 8am and 5pm on weekdays, goals read from managers' sent mail, and team recaps to the CEO. The CEO directs it from the chat assistant the CEO already uses, or by email.
Phase 3
The Storefront Employee
On October 7, 2026, a second employee went live on the storefront. It knows the page on screen, searches the full catalog, quotes shipping that matches checkout, and answers signed-in shoppers about their own orders.
Phase 4
Write Tools With Previews
On October 30, 2026, the operations employee gained write tools for listings, shipping, and seller work. Every write shows a preview, needs a confirm and a reason, and is logged, and caps limit what it spends and how often it messages anyone.
Phase 5
Eight Jobs
Both employees were live within 90 days of kickoff. With the write tools and the desks added since, the two hold 8 jobs. The operations employee holds 7 of them, 15 scheduled jobs run the recurring work, and 2,138 automated tests cover the agent code.
THE JOBS THEY HOLD
Whole jobs, not single tasks
Chief of staff
For: Every team member and the CEO
- Writes each person a personal check-in at 8am and 5pm local time on weekdays
- Keeps a "waiting on your reply" list of threads older than 4 business hours, with reply speed compared to the team median
- Reads goals from managers' sent mail
- Keeps a running list of asks that got no answer
- Sends the CEO a team recap every Monday and Thursday
- Takes goals and priorities from the CEO by chat or email
Guardrail
Every figure must trace to data, and a draft that names a colleague is rejected. A rejected check-in isn't sent.
Business analyst
For: The CEO and ops
- Builds the weekly business report every Friday at noon, covering orders, in-store and trade sales, traffic, email, and events, with a code check on every number
- Pulls report history for any week or the last 12
- Builds ad hoc reports from the database plus the email marketing, product analytics, web analytics, and support tools
- Attaches reports as CSV or Markdown downloads
- Sends up to 4 helper agents out in parallel for big questions
Guardrail
Database queries run read-only.
Marketing coordinator
For: The marketing lead and the CEO
- Builds two newsletter editions a week: a theme, a 32-piece collection, fact-checked copy, and a draft in the email platform
- Scores pieces for each collection, agreeing with hand scores 90.5% of the time
- Emails each edition for review and applies "approve", "reject", or "rebuild" replies
- Tracks each seller's assets and promotion for a live event and puts reminders in check-ins
Guardrail
The newsletter never sends on its own, and it goes out only after a 24-hour preview.
Listings coordinator
For: The listings team
- Creates and updates listings, variants, and images, and publishes or unpublishes them
- Builds a listing from a product link with up to 8 photos
- Proposes listing batches from spreadsheets and photos
- Sets shipping details per listing: parcel size, pricing method, and tariff code
- Runs a review desk, in shadow mode, that rewrites scraped candidates to the house standard and prices them
Guardrail
A person approves every batch, and every change can be reverted.
Partner manager
For: Sellers and the partnerships team
- Chats with sellers in the seller portal and can drive the portal on screen
- Works through the live-listings audit with the seller
- Answers seller email replies, up to 10 a day
- Gives staff 11 partner tools, including commission, re-costing, publishing a seller, and inviting a partner
Guardrail
Messages per order are capped at 3 a week to a seller and 2 to a buyer.
Logistics coordinator
For: The logistics lead and support
- Uses 15 logistics tools: labels, customs documents, pickups, ship-by dates, freight quotes, and messages to buyers or sellers
- Turns every label failure into an incident, where rules retry first and the employee then writes the escalation email and answers in the thread
- Sweeps for stuck orders, and one sweep found 33 stuck items
- Closes stuck orders from tracking so the seller gets paid
- Books carrier pickups
Guardrail
Label purchases are capped at $200 and pickups at $25 each.
Ops support lead
For: Ops and support staff
- Answers "where is this order" from the live database
- Answers "how does this work" from more than 2,000 source files
- Shows, opens, fills, and clicks non-destructive controls on any admin page
- Ends each answer with "Do it for me", "Show me where", or "No thanks"
Guardrail
It never clicks a destructive control or types into password, card, or bank fields. Every write needs a preview, a confirm, and a reason, and is logged.
Shopping advisor
For: Every storefront visitor
- Knows the page on screen and searches the full catalog with sort and filter
- Quotes shipping that matches checkout, with live carrier rates for any country's postcode
- Answers order status questions for signed-in shoppers
- Answers from press, decks, and company pages refreshed nightly
- Guides the shopper on the page by clicking only reversible controls
Guardrail
One switch turns it off everywhere.
WHAT WE BUILT
Features and capabilities
01
A Named Employee With One Memory
Each AI employee has a name, a job description, standing rules, and a manager, like anyone else on the team. The operations employee holds 7 jobs and keeps one memory across all of them, so staff work with one colleague instead of seven bots.
02
Role Profiles
The operations employee runs on 4 role profiles: admin operations, chief of staff, seller support, and listing reviewer. Each profile sets the tools and the rules for the people it serves. Adding a job means adding tools and playbooks to a profile, not building a new bot.
03
Typed Tools
Work happens through typed tools with defined inputs and results, not open-ended access: about 60 in the admin role, 34 in the seller role, and 18 for the storefront employee. Logistics alone has 15, from labels and customs documents to pickups and freight quotes.
04
Skills Written as Playbooks
How to do each job is written down as a skill: a plain-language playbook the employee follows, the way a new hire follows a runbook. A skill reads like the instructions you'd give a person, so the team can see how the work is meant to be done.
05
Scheduled Work
Fifteen scheduled jobs run the recurring work without anyone asking: check-ins at 8am and 5pm, the business report every Friday at noon, the team recap every Monday and Thursday, and two newsletter editions a week.
06
Channels Where People Already Work
The operations employee works from a sidebar on every admin page, its own email address, and the CEO's chat assistant through a secure connector. It sends check-in emails to staff, chats with sellers in the seller portal, answers seller email replies, and works shipping incident threads with the support inbox. Leadership can email it instructions, and it acts only on senders verified with DMARC and an allowlist.
07
The On-Screen Stage
On any admin page, the operations employee can show, open, fill, and click non-destructive controls, and each answer ends with "Do it for me", "Show me where", or "No thanks". It can drive the seller portal on screen during a seller chat, and the storefront employee guides shoppers by clicking only reversible controls.
08
Guardrails in Code
Database access runs through scoped database roles, and only one of them can read the sealed back-office data. Every desk has its own kill switch and runs in shadow mode before it goes live. Caps limit label purchases to $200, pickups to $25 each, seller email replies to 10 a day, and messages per order to 3 a week to a seller and 2 to a buyer. The agent code has 2,138 automated tests.
09
Human Approval
A person approves every listing batch and every newsletter edition, and listing changes can be reverted. Every write needs a preview, a confirm, and a reason, and is logged. Every reply to a leadership email is blind-copied to our engineer.
KEY TAKEAWAYS & IMPACT
What changed for this business
8
Jobs held by two AI employees
Chief of staff, business analyst, marketing coordinator, listings coordinator, partner manager, logistics coordinator, and ops support lead are all one employee. The shopping advisor is the other. A person approves the changes that matter.
272→91
Threads in the CEO's "waiting on you" list
A relevance filter cut the CEO's list of threads waiting on a reply from 272 to 91, so the CEO starts from the threads that need an answer.
90.5%
Agreement with hand scores on newsletter picks
Newsletter curation used to run on hand-run scripts. The operations employee builds two editions a week, and its scoring agreed with hand scores 90.5% of the time. Each edition still waits for a reviewer's reply and a 24-hour preview before it sends.
THE OUTCOME
The wins
-
The weekly business report a staff member built by hand arrives every Friday at noon, covering orders, in-store and trade sales, traffic, email, and events, with a code check on every number.
-
Every team member gets a personal check-in at 8am and 5pm on weekdays, and a list of threads that have waited more than 4 business hours for their reply.
-
A sweep for stuck orders found 33 stuck items, and the employee closes stuck orders from tracking so the seller gets paid.
-
Every shipping label failure becomes an incident: rules retry it first, then the employee writes the escalation email and answers in the thread.
-
The CEO manages the operations employee from the chat assistant the CEO already uses, with 11 management tools, 2 that read and 9 that make changes, or by email.
-
New work becomes a new role, playbook, or tool on an employee the team already knows, not another bot with its own screen and rules.