AI automation2026In use

Digital Office: 15 AI agents that feed each other, one human approver

An AI back office I built for the marketing of the Glass Railing Store, a Shopify store for frameless glass railing in Canada. Fifteen agents in six rooms read the store’s data on a schedule, hand their work from desk to desk and check each other, so I can also double-check my team’s work and compare their reports with what the agents find. A General Manager agent reviews every room and checks every agent each morning, and every move that matters waits for my yes.

GRS Digital Office
Isometric illustration of the Glass Railing Store digital office on a Monday at 08:00: most agents at their desks with blue spinners above their names, a few on a break in the kitchen, and a side panel counting down to Christmas

Opens in a full-screen window on this page. You stay on this site the whole time.

An office instead of another dashboard

A small online store runs on a dozen dashboards: ads, analytics, SEO, email and social. Checking them every day takes hours, and the signals that matter are easy to miss: an ad quietly burning money, a page slipping in search, a post that people actually save and share.

So the store got an office instead of another dashboard. Every desk is an AI agent with one job, a fixed schedule and clear limits. The agents read the store’s data, write what they find into shared documents, hand their work to the next desk and check each other. I get a short list of decisions, and nothing reaches customers, ad accounts or the live website without my yes.

It is also a second pair of eyes on my team: the agents read the same data as the people who do the work, so I can double-check the team’s work and compare their reports with what the agents find before I decide.

Fifteen agents in six rooms

Every instruction is written as a job description: the role, the data to read, the output format, the thresholds and the hard limits. The times are Toronto time.

General Manager
General Manager. Reads every room’s latest work, fixes small mistakes and sends weak work back, checks that the Ad Analyst and the Content Editor feed each other, writes the weekly report and tracks approvals. Every morning it checks that each agent ran and each connection works, and passes the agents’ requests to me. Mon and Wed 07:53, daily 08:51
Ads Marketing, 4 agents
Google Ads Specialist. A daily spend check, and a weekly campaign, region, device and search-term analysis with budget-shift proposals. Daily 06:19, deep dive MonMeta Ads Specialist. Creatives, ad sets, audiences and placements; flags ad fatigue. Daily 06:19, deep dive MonAd Analyst. Turns winning ads, search terms, trends and real engagement into a weekly brief with five scored ideas, and checks the drafts against the data on Thursday. Mon and Thu 06:47Amazon Seller Specialist. Sales, orders, stock, returns, the Buy Box and Sponsored Products spend. Waiting for access
SEO, 4 agents
Search Console Specialist. Clicks, impressions and average position, Canada and the US apart, and the pages that climb or slip. Mon, in the GM’s runSemrush Specialist. Organic keywords and estimated traffic, competitors and backlinks; hands keyword ideas to the Content Editor. Mon, in the GM’s runGA4 Analyst. Sessions by channel, top organic landing pages and conversions, week over week. Mon, in the GM’s runSEO Project Tracker. Follows the outside SEO specialist’s eight-week plan on a shared task board: checks finished tasks on the live site, writes feedback on each one, measures site health and ten goals every Monday and reviews the specialist’s weekly report. Mon, Wed, Thu, Fri 07:57
Content Studio, 3 agents
Trend Researcher. Reels, TikTok, Shorts and Pinterest trends and what competitors post, adapted to glass railing as video and photo ideas the store can film. Mon 06:07Content Editor. Scores the trend brief, plans the week’s content, briefs the creators and answers the Ad Analyst; drafts the approved pieces. Mon and Wed 07:07Blog Creator. Two posts a week from the editor’s brief, with photos from Google Drive and the store’s Shopify files: light, colour and shadows fixed without changing the product, under 150 KB, clear file names and alt text. Waiting for access
Social and Email, 2 agents
Social Media. One joint campaign with Email Marketing for every special day, and trend videos from the store’s own footage. Daily 08:23Email Marketing. A complete Shopify Messaging draft from the editor’s email brief: layout, copy, images from the store’s file library, tagged products and build steps. Mon 07:29, daily 08:23
Web and Tech
Web Developer. Site health every other day: products and SKUs, redirects, broken pages, the site audit and Search Console. Tells me right away when something breaks, and fixes only on a copy of the theme. Every other day 05:41

Feeding each other

The most important link in the office runs between the Ad Analyst and the Content Editor. Together they decide what content gets made, and real engagement data decides it, not guesswork.

  • Monday 06:07. The Trend Researcher’s report goes to the Content Editor: what is moving on Reels, TikTok, Shorts and Pinterest, turned into ideas the store can film.
  • Monday 06:47. The Ad Analyst’s trend brief opens with an engagement scoreboard (the posts people liked, shared and saved, the blog pages that hold readers longest, the ad angles that get clicked and bought), then five ideas, each with a hook, a target keyword, a score from 1 to 5 and its evidence.
  • Monday 07:07. The Content Editor accepts or rejects every idea with a reason, plans up to three pieces, briefs the creators and leaves feedback and questions for the Ad Analyst.
  • Wednesday 07:07. The editor drafts the approved pieces, each with an engagement goal and a note asking the Ad Analyst for data to check.
  • Thursday 06:47. The Ad Analyst checks every draft against the data, adds two alternative hooks for A/B testing and a predicted score, and answers the editor’s questions.
  • Next Monday. The new brief starts with the editor’s questions answered and last week’s results measured, so the loop learns every week.

A Monday morning

  1. 06:07 The Trend Researcher writes the trend report and walks it to the Content Editor.
  2. 06:19 Ads Marketing compares yesterday’s spend with the seven-day average and writes the weekly ads analysis, with the winning search terms and creatives the Ad Analyst needs.
  3. 06:47 The Ad Analyst builds the trend brief and the engagement scoreboard and walks them to the Content Editor.
  4. 07:07 The Content Editor writes the content plan, with an email brief for Email Marketing, and leaves questions for the Ad Analyst.
  5. 07:29 Email Marketing writes this week’s email draft from the brief.
  6. 07:53 The General Manager reads every room’s work and writes the weekly report: the numbers, one line per room, the content plan, the decisions to approve, the agents’ requests and next week’s plan. The SEO desks report in.
  7. 07:57 The SEO Project Tracker measures site health and the plan’s ten goals and checks the SEO specialist’s week.
  8. 08:23 The Social and Email desk checks the special-day calendar and moves any campaign that is due.
  9. 08:51 The General Manager checks that every agent ran and every connection works.
Monday 07:06: the Ad Analyst has walked the trend brief to the Content Studio and talks it through at the Content Editor’s desk, while most of the team works at their desks
Monday 07:06: the trend brief changes hands at the Content Editor’s desk.

House rules

Green and red

The General Manager can approve the green work: analysis, reports and plans; drafts of content, email, product metadata, designs and social posts; small fixes inside department documents; test plans on a copy of the theme. The red work waits for me: publishing on social, sending email, ad budgets, bids and campaigns, prices and discounts, live site changes, new access or connections, deleting anything, and orders, payments and customer records. The one exception: the Blog Creator may publish a post once the General Manager has reviewed it. I answer each decision in the weekly report with approve or reject; green items with no answer after 48 hours are approved by the General Manager, and red items always wait.

The General Manager’s review

  • On Monday and Wednesday it reads every room’s latest work: sources and date ranges, whether conclusions follow from the data, links, that every image really exists in the store’s file library, brand fit, and claims the website doesn’t make.
  • It checks the loop: does the brief answer the editor’s questions, does the plan pick the high-scoring ideas, do the drafts carry the analyst’s notes?
  • A small mistake is fixed in place with a note at the top of the document; a bigger problem is sent back with the reason and never reaches me.
  • Every morning at 08:51 it checks that each agent ran on time and each connection works, and I get a phone alert at once if one did not.
Wednesday 08:03: the General Manager reviews the drafts from his office with a blue spinner above his name, most desks are at work, and the two agents waiting for access sit in the kitchen
Wednesday 08:03: the General Manager reviews the drafts from his office.

One owner for every question

Whatever I send or ask goes to the agent whose job it is, and only that agent handles it; an agent that gets something outside its job names the right desk instead. Agents never give each other work directly, and they never connect tools or sign in anywhere themselves: every document ends with “what would help me do better”, and the General Manager brings the requests to me.

Special days and data

  • Every special day gets one joint campaign from Email Marketing and Social Media, started weeks ahead and reviewed by the General Manager; any offer is my call.
  • Store data is read-only, and no report carries a customer’s personal data, only totals.
  • No guessing: every number carries its source and date range, and no verdict on a region, ad or keyword with less than 75 CAD of spend or 40 clicks behind it.
  • Social posts are scheduled with auto-publish off and I publish them; I build and send the emails myself.

Under the hood

Each desk with its own schedule is a scheduled Claude task: a fresh session starts at a set time, does its job and finishes, and the SEO desks work inside the General Manager’s Monday run. The data comes from Shopify (read-only), Google Analytics 4, Google Search Console, Google Ads and Meta Ads through Windsor.ai, Semrush and Metricool, and the photos and footage from Google Drive. Work travels between desks as shared documents (the ads analysis, the trend brief, the content plan, the drafts, the email draft and the weekly report), and agents only add notes to their own documents. The SEO Project Tracker writes to the SEO specialist’s task board through its own connector, and the board can start it the moment the specialist saves an update. It is the kind of system I build in AI automation, beside the Image Library the same team works in.

The office view

The interactive office is a hand-drawn isometric scene on one HTML canvas, with no libraries. It plays a Monday, a Wednesday, a Thursday and a Friday morning in a loop. An agent on a run shows a blue spinner and moves at its desk; when work changes hands, the agent carries the paper to the colleague’s desk and the two talk it through. On this site most desks stay at work between their runs, and two agents at a time take a break in the kitchen; the General Manager stays in his office. A panel counts down to the next special day, and All ready stops the loop. Hovering over an agent shows its job, tools and schedule, and the floor can be zoomed. It is a simulation of the real schedule, not live data.

The digital office on a phone, playing a Monday morning: most agents at work with blue spinners, and the special days panel below the office
On a phone, tap an agent to see its job.

More work

All work

Want a second pair of eyes on every report?

Tell me about your business and what you want to change. A few sentences are enough, and I reply by email.