What does an AI ads audit engine actually check?

Mine runs 74 checks on Google Ads across conversion tracking, wasted spend, account structure, keywords, Quality Score, assets, Performance Max, bidding and settings, and 46 on Meta across Pixel and Conversions API health, event match quality, creative diversity and fatigue, learning phase, audience targeting and Advantage+. The checks are code. The model only writes up what the code found.

Why build a platform change-log agent?

Google Ads and Meta ship product and policy changes every week and nobody reads the release notes. Teams find out when a feature is auto-upgraded, a campaign breaks or a client asks. An agent that reads the notes for you, decides which ones matter to paid search and paid social, and posts a digest to Slack turns a surprise into a Tuesday morning message.

How should a small agency start with AI agents?

Pick one job you do more than 3 times a month and write down how you do it, step by step. Make every check you can deterministic. Put a person at the approval point before anything spends or sends. Cap the tokens. Count the hours before you start and again a month in. Reporting is the usual first build, and the one with the most published results.

Playbook · AI, and what it’s actually doing

The two agents I built, and the gaps nobody has filled

An audit engine that runs 120 checks on Google and Meta accounts, a Slack agent that reads release notes, and 5 agents no agency has published. Part 3 of 3.

By: Dom Carter
Published: 24 October 2026
Read: 6 min read

Part 3 of 3. The two builds are mine, the audit engine is my IP, and both run for the paid media team I work with. Neither has a measured result yet, which I'd rather say than pretend.

The short version

I built a self-serve audit engine: upload Google Ads and Meta exports, get an A to F health score, an executive summary and a 90-day plan. 74 Google checks and 46 Meta checks run as code before the model writes a word.
I built a Slack agent that reads Google Ads and Meta release notes on a schedule, classifies each change by relevance, and posts a plain-English digest with an action-needed flag. Nobody reads release notes until something breaks.
Neither has a measured result yet. The metrics to capture are listed below, because an unmeasured agent is an announcement, and part 1 was about how little those are worth.
Five agents have no published deployment with results anywhere in the trade press: platform change-log, pitch and RFP, meeting-notes-to-tasks, resourcing and profitability, and social response. Two of mine sit in those gaps.
How to start: pick a job you do more than 3 times a month, write down how you do it, make the checks deterministic, put a person at the approval point, meter the tokens, and measure hours before and after.

What I mean by

Audit engine: A tool that takes raw exports from an ad account and returns a structured review: what is set up wrong, what is wasting money, what to fix first. Mine runs 120 fixed checks across Google Ads and Meta by code, scores the account A to F, then asks the model to write the summary and the 90-day plan from those findings.
Change-log agent: A scheduled agent that watches the pages where ad platforms publish product and policy changes, removes duplicates, classifies each change by which team it affects, and posts a digest where the team already looks. It exists because nobody reads release notes until a campaign breaks or a client asks.
Unmeasured agent: An agent that runs but whose effect nobody has counted: hours saved, errors caught, work produced. Most agency agent announcements are unmeasured. I have two of them myself. The difference is saying so, and listing what the measurement will be.
Whitespace: An agent job that agencies plainly need, that is technically straightforward, and that no named agency has published with results. Five of them, as of September 2026: change-log, pitch and RFP, meeting-notes-to-tasks, resourcing and profitability, and social response. Each is a case study waiting to be written.

Part 1 argued that most of what agencies say about AI agents can’t be proved. Part 2 looked at the ones that can. This part is about the two I built, what they do, what they don’t do yet, and the 5 agents I think the market is missing.

I’m writing these up with the same rules I applied to everyone else. Neither of my builds has a measured result yet. I’ll say what the measurement will be rather than invent a number.

What does the audit engine do?

Every paid media audit I’ve ever been involved in went through one senior person. That person reads the account, applies a mental checklist built up over years, writes it up, and the new-business or account team waits. It is slow, it doesn’t scale, and the checklist lives in a head.

So I wrote the checklist down and turned it into code.

The engine takes Google Ads and Meta Ads CSV exports and returns a structured audit. The Google side runs 74 checks: conversion tracking, wasted spend, account structure, keywords, Quality Score, ad assets, Performance Max, bidding and settings. The Meta side runs 46: Pixel and Conversions API health, event match quality, creative diversity and fatigue, campaign structure, learning phase, audience targeting and Advantage+. Each check is deterministic. A missing conversion action is a missing conversion action, whatever the model thinks.

Only then does the model get involved. It takes the findings and writes the executive summary, the A to F health score rationale and a 90-day strategy, prioritised. The output exports to Word, PowerPoint or PDF, because that is what ends up in front of a client.

Stage	Who does it	Why
Parse the exports	Code	Fixed format, no judgement needed
Run the 120 checks	Code	A finding has to be reproducible, and a model can’t be relied on to find the same thing twice
Score the account A to F	Code, weighted	So two audits of the same account agree
Write the summary, rationale and 90-day plan	The model	This is the part that needs language and prioritisation
Review before it reaches a client	A person	The rule in every credible case in part 2

It is the deterministic-first pattern from part 2, the same one that cut Apollo Partners’ quality assurance from 8 hours to 2 minutes. The nearest published equivalent is Anicca’s 85-check Google Ads audit, which runs on Claude Code skills rather than as a standalone tool.

The first version runs in the browser and asks each user to paste their own model key, which is friction nobody wants. The next version moves the key server-side behind a small function so anyone on the team just opens a link. The IP is mine; the paid media team I work with runs it.

What I’ll measure: audits produced per month against the number before, hours per audit, and the win rate on pitches that included one.

What does the change-log agent do?

Google Ads and Meta change something every week. A match type behaves differently, a campaign type is auto-upgraded, a policy tightens, a reporting column disappears. The release notes exist. Nobody reads them. Teams find out when a campaign breaks or a client asks why the number moved.

The agent reads them instead. On a schedule, it visits the Google Ads and Meta release-note sources, removes the duplicates, classifies each change by relevance to paid search and paid social, summarises it in plain English with an action-needed flag, and posts the digest into a Slack channel the team already lives in. No human trigger. Someone reads the post with their coffee and raises anything that needs a client conversation.

I’m not the only person who has built this. Ivan Potekhin, a senior paid search specialist at the agency AI Digital, open-sourced a weekly Claude routine that visits 13 source pages across Google Ads, Microsoft, ChatGPT Ads, Meta, LinkedIn and TikTok, de-duplicates into a store, and posts separate search-team and social-team summaries to Slack. First digest in August 2026, still active. His is more thorough than mine on sources. Neither of us has published a result, and no named agency has either, which makes it the first of the gaps.

What I’ll measure: minutes per week saved against reading the notes by hand, the number of changes flagged action-needed per month, and one story of a change caught before it reached a client.

Which agents has nobody built in public?

Going through 20 published cases and the trade press behind them, 5 jobs came up again and again as things agencies need, and not one has a named agency deployment with results.

The job	Why agencies need it	Closest thing published
Platform change-log	Weekly platform changes, nobody reads the notes	My build and Potekhin’s open-source routine, both unmeasured
Pitch and RFP	Every pitch starts by trawling past pitches for content	Mekanism runs one, self-reported, no numbers
Meeting notes to tasks	Actions agreed in client calls get lost between the call and the tracker	Brainlabs writes notes to Notion tasks, no result published
Resourcing and profitability	The 13% average margin in part 1 is a resourcing problem, and most agencies find out after the month ends	Nothing
Social response	Replies to comments and messages at brand scale	A vendor product with one brand’s self-reported 175% response lift

Two of mine sit in the first row. The other four are open. If you run an agency and want a case study nobody else has, pick one of those.

How should an agency start?

Everything I learned from the 20 cases and my own two builds comes down to 6 steps, in this order.

Pick a job you do more than 3 times a month. Brainlabs’ rule. Reporting, audits, onboarding, pacing checks, release notes. Not the clever thing, the repeated thing.
Write down how you do it. Step by step, as if briefing a capable new starter. This document is the skill, and it is worth more than the agent, because it is the first time the job has existed outside someone’s head.
Make every check you can deterministic. If a rule can be code, make it code. Save the model for language, prioritisation and judgement.
Put a person at the approval point. Before anything spends money, changes an account or reaches a client. Every credible case in part 2 does this. The ones that don’t, don’t publish results.
Meter the tokens before you scale. An unattended agent runs up a bill. Dept learned that at 1.5 million tokens in a day. Set a cap and a dashboard on day one.
Count the hours before and after. Otherwise you have an announcement, and part 1 was 2,000 words on what those are worth.

The stack is in part 2 and it is cheap: a model, a scheduler, a data connector, a chat channel. None of the small agencies that got results hired a developer. All of them had someone senior who understood the job.

The so-what

I’ve spent 15 years in agencies and most of that time watching the same knowledge live in the same few heads. The agent isn’t the valuable bit. The written-down job is. The agent is just the first thing that makes writing it down pay back in weeks rather than years.

The cost of not doing it is the one in part 1. Clients have stopped paying for time, the average small agency has 13 points of margin, and somewhere an agency with 30 people and no developers is already running 80 skills. Pick a job. Write it down. Measure it. Then write up what happened, because the market is short of people who do.

Questions people ask

What does an AI ads audit engine actually check?: Mine runs 74 checks on Google Ads across conversion tracking, wasted spend, account structure, keywords, Quality Score, assets, Performance Max, bidding and settings, and 46 on Meta across Pixel and Conversions API health, event match quality, creative diversity and fatigue, learning phase, audience targeting and Advantage+. The checks are code. The model only writes up what the code found.
Why build a platform change-log agent?: Google Ads and Meta ship product and policy changes every week and nobody reads the release notes. Teams find out when a feature is auto-upgraded, a campaign breaks or a client asks. An agent that reads the notes for you, decides which ones matter to paid search and paid social, and posts a digest to Slack turns a surprise into a Tuesday morning message.
How should a small agency start with AI agents?: Pick one job you do more than 3 times a month and write down how you do it, step by step. Make every check you can deterministic. Put a person at the approval point before anything spends or sends. Cap the tokens. Count the hours before you start and again a month in. Reporting is the usual first build, and the one with the most published results.