andrei achim.
How I built the automation layer behind a global snacks brand's paid media, on contract at a top-three agency network. AI agents, Chrome extensions and flows, all on the tools the team already had.
A day of the build log, replayed. The system runs; the team checks.
Illustrations use invented names and data. The figures are from my build records.
// 00 — Take the templates
The reactive tracker works out every post's status, SLA and next step from its dates. The media plan prices every fee from the market you pick and won't let a plan ship until 19 checks pass. Both come with invented example data, so you can watch the formulas work before you make them yours.
// Reactive social tracker
8 statuses · SLA by working day · next step written for you
// Media plan template
14 markets · fee engine · benchmarks by KPI · 19 checks
Excel files. Built around my workflows, so expect to adapt them to your rate cards, markets and sign-off rules.
// 01 — The problem
One global brand, many markets, paid social and programmatic on Meta, TikTok, Pinterest and Snapchat. The plan lived in Excel. Briefs arrived in a client tracker. Tickets sat in Jira. Questions came in by email, Teams and calendar invites.
Every campaign was built by hand, platform by platform, market by market. So the same four things kept going wrong:
Every ad set typed in by hand, every week.
One wrong name and the ad drops out of reporting.
What's live slowly stops matching what was signed off. One wrong budget setting at set-up can spend money for days before anyone looks.
A brief lands on a Friday afternoon, a cut-off slips past, the client chases.
A big account run on spreadsheets and memory produces these errors every week, however good the people are.
And the rules were tight. Locked-down enterprise Microsoft 365, standard licences, strict data-loss policies, and no approved connector into any ad platform.
// 02 — What I built
10 are live, 2 are in test, 4 are drafts or paused.
System map. Inputs: the media plan in Excel, the client tracker, Jira, and the inbox, Teams and calendar. They feed five build groups: agents in Microsoft 365, Chrome extensions, intake and SLA flows, system connections, and media finance and QA. Outputs: the Meta, TikTok, Pinterest and Snapchat ad accounts, the team digest and alerts, regional sheets, and SharePoint.
// 01
Copilot Studio and M365 Agent Builder. A client-comms agent sweeps mail, chats and the calendar, logs every client ask, checks for proof it's done and drafts the reply. A plan-builder agent (co-built) builds media plans through chat. A daily digest lists the set-up steps for lines going live in the next 14 days, and a launch-day QA skill posts 5 to 9 findings to the team.
// 02
Two Manifest V3 extensions. The first checks every Meta, TikTok, Pinterest and Snapchat build against the plan before it saves: budget, dates, naming and landing URLs, with preview links pulled for sign-off. The second reads an ad account against the plan, the naming taxonomy and the creative matrix, then renames, fixes URLs and builds the ads.
// 03
Power Automate on standard licences, plus Office Scripts. The client's tracker export becomes tracked campaigns and a daily digest with an SLA countdown. A reactive social tracker runs the SLA rules (cut-offs, the Friday stop) and alerts at each handoff. A readiness flow builds a 14-step checklist per campaign line, counted back from go-live.
// 04
Apps Script pulls self-hosted Jira tickets into 8 regional tabs twice a day, parses the taxonomy, follows renamed tickets and alerts the owner. SharePoint holds the shared data.
// 05
A media plan template with a fee engine (DSP, DST, ad serving, levies), monthly FX, KPI benchmarks and 19 automated checks. In test: a campaign QA pipeline that runs intake agent, plan build, fixed-rule checks, compliance agent, then approval.
10 live · 2 in test · 4 drafts or paused
// 03 — The numbers
// Jira sync
| What it does | Result |
|---|---|
| Chrome extension that checks ad builds against the plan before they save | 58/58 live ads and 58/58 URLs matched the taxonomy |
| Chrome extension that builds and fixes ads from the plan and creative matrix | 13 tested releases in about 8 hours |
| Client-comms agent (Copilot Studio) | 7 users, 95 conversations in 6 days |
| Plan-builder agent (Copilot Studio, co-built) | 185 conversations in 2 weeks |
| Media plan template with fee, FX and benchmark engine | 1,314 formulas, 0 errors across 14 markets; 19 automated checks |
| Jira to Sheets sync (Apps Script) | About 980 tickets scanned per run, twice a day, 0 failed runs |
| Reactive social tracker with SLA rules | 22 unit tests; alerts routed to the right person, 0 duplicates on re-run |
Client, agency and tool names are removed. Figures come from my build records.
// 04 — How I work
Nothing touches a live ad account or a shared file until it has run clean on a copy.
Unit tests on the rules, end-to-end runs on real data, and read-only test runs for anything that writes.
Every change ships as a numbered version the team can roll back. One extension went through 13 versions in a working day.
What it does, what it touches, how to run it, how to switch it off. The team can run it without me.
Standard licences, data-loss policies, no unapproved connectors. When three routes for the set-up digest were blocked, I found a fourth the licence allowed.
Outside Microsoft I build on n8n, Supabase, Next.js and Claude.
// 05 — For your team
Agencies, in-house media teams, any business where people copy data from one system into another and check it by eye.
I come in on contract, sit with the people doing the work, find the tasks costing the most hours, and build the fix on the tools you already pay for. Usually that's Microsoft 365. Sometimes it's Google Workspace, or n8n when you're free to choose.
What you get: agents your team actually opens (7 people in the first week here), checks that catch an error before money is spent, and a build record for everything, so it keeps running after I leave.
01 · The call
20 minutes, free. You show me the plan, the tracker and where briefs land. I'll tell you which workflow I'd automate first.
02 · The diagnostic
One week, £1,950. I sit with the people doing the work, time the manual steps and write up what to automate first, what each fix saves and what it costs to build. Credited in full if you go ahead.
03 · The build
Fixed price, quoted from the diagnostic. Built on your tools, tested on copies, with a build record so it runs after I leave.
Smaller business, no media team? Same method, smaller scale. One workflow, one session, built on your own tools. The build session · £150 →
AI agents, automation and AI architecture for founders, teams and companies.
Book the free call or get the templates first ↑Prefer email? andrei@andreiachim.com · CV on request.