Giant preface: I don’t think I’m an AI expert. There are probably better ways to do some of what I’m about to describe. But I run an agency where we use AI every day on real client work, and we’ve learned a lot about what holds up and what falls apart.
Here’s the short version. I was recently at a conference with other agency owners, and a lot of them are moving fast to take people out of their agencies. There’s a business model there. You cram a bunch of stuff into AI, it gives you a mediocre output, and mediocre for cheap is better than nothing. I just believe in a lot more than mediocre.
Where AI actually saves time
If I can never write another page title or code another line of schema again, I’d be thrilled. AI is great at the tedious stuff around the work. Here’s where it’s earning its keep for us right now:
- Reporting. I point AI at last month’s report and this month’s data export. It copies the deck, updates the dates and numbers, drafts the callouts, and leaves notes on every slide that still needs me.
- Technical SEO audits. An audit used to take one of our specialists about 8 hours. With a set of AI skills we built, the draft is about 80% done before a person ever touches it.
- Account management. A weekly AI review tells me if we’re on track for hours, if anything in our project management tool is waiting on me, and what I might be dropping.
- Sales follow-up. I sketched out a follow-up agent for our sales lead. She runs it once a day, it looks at her open conversations across call notes, email, and our CRM, and it suggests follow-ups. She approves what she wants.
- Websites. I built a website with AI. I’d never built one before, and I’m terrible at development. More on the guardrails below.
Aim for 80%, then let a person finish
The goal for almost every AI workflow we build is the same: get the draft about 80% of the way there, then have an experienced person finish it. A good audit, report, or page review needs someone who knows the client to look at it and say, “Oh no, this is weird,” or “This doesn’t feel right.”
Be honest about what the 80% is, too. When people first told me AI cut our audits from 8 hours to 20 minutes, that’s not really what happened. The first version mostly moved data from one tool into our template. Getting to a draft that actually thinks took weeks of back and forth, interviews with the people who do the audits, and a lot of testing.
Build gates, not just prompts
A prompt is a request. A gate is a step AI can’t skip. The workflows that hold up for us have gates built in:
- Required inputs. Before our audit workflow runs, someone spends about an hour gathering what it needs: Search Console exports, crawl data, and the pages to score, all in one folder. No folder, no audit.
- Questions before answers. The workflow reviews everything, then asks the specialist four or five questions. Is this issue a big deal for this client? Should we downplay it? You can’t just say “looks good, go.”
- A QC layer. Our audit workflow checks its own work for the mistakes we used to see most, like pulling URLs from the wrong account or language that reads like a bad translation.
- A human sign-off. Nothing goes to a client without a person reviewing it for accuracy, voice, and whether it actually makes sense for that business.
A good AI skill takes weeks
You’ll see people on LinkedIn sharing “25 AI SEO skills, go get them.” It takes two seconds to make a new AI skill. It takes weeks to make a good one.
What made ours good was boring work: interviewing the people who do the job, writing down what “great” looks like, adding examples, and correcting the instructions every time the output missed. Corrections go back into the instructions so the next draft is better than the last one.
Check AI’s numbers with a second AI, then a person
Large language models are built to predict the next likely word. They aren’t built to be factually accurate or do math. So for anything with numbers in it, we use two tools. One drafts the analysis, and a different one double-checks that it pulled the data accurately. Then a person spot-checks the high-level numbers against the source.
Here’s something I didn’t expect: data entry isn’t the slow part. Pulling numbers from analytics takes five minutes. The part that takes time, and the part people are often weakest at, is seeing something weird in the data and figuring out why. Use AI to speed up the manual part, and use it as a thinking partner on the “why.” Just don’t let it decide the why for you.
Where AI makes things worse
When nobody checks its work. The failures I see most:
- Content that sounds like everyone else’s. When I built my website with AI, I wrote the core content myself first. If I’d let AI write it, it would have been bad. Every time.
- Confident, wrong facts. I’ve had my own AI get an email wrong. It happens. A review step catches it.
- Audience backlash. On social media, people are actively calling out brands that use AI art and AI-written posts. Our own community management test saw negative reactions to AI content.
- Bigger risks. Content getting de-indexed, an AI agent handling client communication, or confidential data going into a tool that shouldn’t have it. We draw hard lines on all three.
The hard part is discipline
AI makes it so easy to skip the thinking. Your metadata is done, you submit it, you ship it, you’re on to the next thing. There’s early research suggesting heavy AI use dulls critical thinking unless people pause, read the output, and think about it.
So the real job for a marketing leader is building the pause into the process. That’s what the gates are for. It also helps if your team knows how the work is done by hand. When the AI asks a question, they need enough context to know the answer.
Track three things
If AI is saving time but quality is sliding, you don’t have a win yet. Track:
| What to track | What it looks like |
|---|---|
| Time saved | Hours per deliverable, before and after |
| Quality | Revision rounds, errors caught before they ship, client feedback |
| Business results | The traffic, leads, and sales you already care about |
For executives, I like a simple CEO-level scorecard: a handful of numbers per channel, checked monthly. If it takes more than a few minutes to read, it’s too long.
A 30-day starting plan
- Week 1: List every repetitive task your team does more than once a month. Note how long each one takes.
- Week 2: Pick the one that’s most tedious and lowest risk. Reporting drafts and meeting recaps are good first picks.
- Week 3: Build the workflow with real examples, a list of dos and don’ts, and one gate a person has to clear.
- Week 4: Run it side by side with the old way. Compare time, quality, and errors. Keep it, fix it, or kill it.
FAQ
Are we behind if we’re not using AI everywhere? No. Using AI everywhere isn’t the goal. Using it on purpose is.
Can AI write our marketing content? It can write a first draft. A person who knows your brand needs to shape it, check the facts, and make it sound like you. Fully AI-written content tends to sound like everyone else’s.
What should we automate first? The tedious, low-risk work with a clear right answer: report drafts, meeting notes, data pulls, metadata, and schema.
How do we keep AI on brand? Feed it real samples of how your brand writes and talks, plus a list of what it should never do. Then keep a person on final review.
Want this as a session for your team or event? See the AI in Marketing talk on the Topics page.