Google’s developer blog posted a reminder on 2 September 2026: Google Ads API v22 stops accepting requests on 7 October 2026. Every automated report, dashboard and alert still pinned to that version will start failing on a Wednesday morning. Most of the people running them don’t know yet.
Short answer first, so you don’t have to scroll. AI automation in a marketing agency reliably handles four layers: pulling data out of platforms, reshaping it, drafting first versions of text, and watching for thresholds you define. It doesn’t handle judgment about whether the number it just pulled is real. Get that order wrong and you don’t save time. You manufacture confident wrong answers faster.
What AI automation in a marketing agency actually replaces
Strip the word “agent” off it and almost every working agency build is one of four things.
- Retrieval. Pulling numbers out of Google Ads, GA4, a CRM or a call tracker on a schedule, without anyone logging in.
- Reshaping. Joining those pulls together, deduplicating, mapping a campaign name to a client, writing the result somewhere readable.
- First drafts. Ad copy variants from a search terms report, a commentary paragraph on a monthly pack, a follow-up email written from a call transcript.
- Watching. A threshold you defined firing a message when spend, cost per lead or conversion volume crosses it.
Notice what is missing. Nothing on that list decides anything. Retrieval, reshaping, drafting and watching all sit downstream of a human decision about what matters and what good looks like. Left alone, every one of them will run happily on broken input for months without complaining once.
The drafting layer is the one that gets oversold. A model producing thirty headline variants from your search terms report is useful. It’s also the easiest part of the job to do badly at volume. You still have to know which of the thirty are safe to run under Australian consumer law, and which are just the landing page H1 restated.
The adoption data says the tool stopped being the bottleneck
McKinsey’s global State of AI survey ran from 4 May to 8 June 2026. It found that nearly nine in ten respondents now use AI regularly in at least one business function. 44% say AI is scaling across the enterprise, up from 38% a year earlier. Among large organisations, 40% report scaling AI agents specifically, up from 27%. Then comes the number that matters: thirty-seven percent attribute at least some EBIT impact to AI, roughly the same share as the year before. Adoption moved a long way. The financial result stayed where it was.
The same survey found nearly three-quarters of its high performers had redesigned workflows outright because of their AI use, up from 55% the previous year. An agency that bolts AI onto the process it already had just gets a faster version of that process, errors included. If your cost per lead has been drifting upward while the reports stayed green, the cause is usually in the measurement, not the media.
Which is a polite way of saying something blunt: most AI automation in a marketing agency is a monthly reporting pack that used to take six hours and now takes forty minutes. It’s still built from the same unaudited conversion data as before.
The audit that has to happen before you automate a single report
A roofing account in Australia came to us reporting 318 conversions at roughly $10 a lead. On paper it was the strongest performer in the portfolio. We traced each conversion action back to an outcome. 96% of those 318 turned out to be taps on the phone number button, counted the instant a thumb landed on it, whether or not a call ever connected. Real traceable leads, meaning form fills plus calls that actually connected through the call extension, came to 12 across 16 months. True cost per lead was closer to $263.
The part that matters here: Smart Bidding had been training on that fake signal for over a year. It had learned to buy the cheapest available taps, and it was pouring budget into junk Display placements that produced them reliably.
Now put a beautiful automated reporting pipeline on top of that account. Sixteen months of correctly formatted, on-time, AI-summarised PDFs, each one faithfully reporting a number that was wrong by a factor of twenty-six. An automation has no opinion about whether a metric is real. It checks whether the API returned a value. The API returned 318 every month, because 318 was what the conversion action said.
So the sequence is audit first, fix whatever fires on intent instead of outcome, then automate. In Google Ads that check lives at Tools > Conversions > select the action > Settings, where Count is set to either Every or One. A phone number tap should almost never be your primary conversion action. We published the full teardown of that account in the 318-lead mirage. The broader version of the same problem is in where wasted spend actually hides.
Four n8n settings that decide whether you can trust the output
Most AI automation in a marketing agency that I get handed runs on n8n, currently stable at 2.38.5. A fresh workflow’s default behaviour is close to the opposite of what you want for client-facing data. Four node settings fix most of it.
Open any node, go to the Settings tab, and switch on Retry On Fail for every node that talks to an external API. Google Ads and GA4 will throw transient errors on a schedule nobody can predict. A workflow that gives up on the first one silently skips a client.
In that same tab, On Error defaults to Stop Workflow. For a loop running across a client list, that default is usually wrong: one expired token kills the run for everybody else. Switch it to Continue (using error output), and wire the error branch somewhere you will actually see it. Plain Continue passes the last valid data forward instead. That’s how an agency ends up sending a client last month’s numbers with this month’s date on them.
Third, set an error workflow. It lives in Workflow Settings under Error workflow. Whatever workflow you point it at has to start with the Error Trigger node. (It is buried about three clicks from anywhere you would think to look for it.) Without it, a failed execution just sits in the executions list, waiting for someone to remember it exists.
Fourth, use the Stop and Error node on purpose. A zero that flows quietly through into a report is worse than a crash. A crash gets fixed on Tuesday; a quiet zero gets presented to the client. If your Google Ads pull returns zero rows for an account that definitely spent money yesterday, that is a failure. n8n will treat it as a success unless you tell it otherwise. Put an If node on the row count, and send the false branch into Stop and Error.
What I would not automate, and why
I don’t automate anything I haven’t run by hand for at least a month. You can’t specify the edge cases of a process you have only watched someone else do. Every agency automation is roughly 20% happy path and 80% edge case. The month of doing it manually is where you find out that two clients use the same campaign naming convention, that one of them renames campaigns mid-month, and that the call tracker drops anything under eight seconds.
Now the part plenty of competent people will disagree with. Most of what gets sold as an “AI agent” in agency land is a scheduled if-statement with a language model bolted on the end. The agent framing makes the build worse, not better, because it encourages you to hand the thing a vague goal instead of a specification. Vague goals are how you get a workflow that decides, unprompted, that a brand campaign is underperforming, and pauses it. Deterministic steps, with a model doing one bounded job inside them, are the ones still running in month six. I would change my mind for open-ended research work, where there is no correct sequence to specify. Monthly client reporting is not that.
Vendors demo the happy path, obviously.
Where these builds break
Four failure modes account for nearly everything I have had to go back and repair.
Version sunsets. Google Ads API v22 stops working on 7 October 2026; v25 is the current version. If your automation was built by a contractor eighteen months ago, and nobody has opened it since, find out today which version it pins. The failure arrives as an empty report, not an email from Google.
Token expiry. OAuth refresh tokens for Google APIs expire. In n8n, the symptom is a credential that worked for months and then returns an auth error at 6am on a reporting day. Nothing warns you in advance. Nothing. Build the error workflow before you need it.
Silent schema drift. A client renames a Google Sheet tab, or inserts a column. Your mapping node keeps running anyway, and writes nulls into every row. The execution is marked successful.
Compounding on bad data. Covered above – and it’s the expensive one, because unlike the other three, it produces output that looks completely normal for as long as nobody checks.
The part I still have not solved
Alerting thresholds for small accounts. I do not have a clean answer yet. On a $20,000-a-month account, the rules work: flag a 40% day-on-day cost-per-lead move, flag pacing more than 20% off target. Below about $2,000 a month, those same rules are useless, because three conversions instead of one is already a 200% swing, and the alert fires most mornings. I have tried longer lookback windows and minimum-volume gates. Both mostly just delay the alert until the money is already spent. For now, I check those accounts by hand on Tuesdays. That’s not automation – it’s a calendar reminder wearing a costume.
Audit what your conversion actions genuinely count. Automate the retrieval and the watching. Leave the judgment where it is. If you want another pair of eyes on what your account is really measuring before you build anything on top of it, book a free 30-minute audit. Our Google Ads management is senior-led, month to month, no lock-in, no reporting theatre. For the narrower version, the walkthrough on automating Google Ads reporting with n8n and AI covers one pipeline end to end. What AI actually handles in website creation asks the same question about build work. For the content side of the same problem, the AI SEO content workflow that does not get penalised covers the review gates Google asked for on 1 October 2026.
Researched and drafted with AI assistance, then reviewed, fact-checked and edited by Rajat Verma before publishing. All account examples are from real campaigns managed by Turquoise Digital and are anonymised.