A lot of posts about AI-run agencies lead with a revenue screenshot and a claim that agents do 80% of the work. Nobody actually measures that 80%, so I'm not going to use it.
I lead SEO at Bonsai Media Group, and a large share of the analysis that crosses my desk is now run by AI agents I direct: Kai, my own agent on OpenClaw, plus Codex and Claude. I design the workflows, decide what goes out, and own the result. The agents do the collection, the first pass, and a surprising amount of the checking.
This is a map of where they carry the load, where they stop, and what keeps them honest.
Where the agents carry the load
1. Keyword research and content planning
The agent:
- Pulls keyword and SERP data (DataForSEO, Semrush, Search Console)
- Identifies gaps (terms competitors rank for that the client doesn't)
- Groups keywords into clusters
- Drafts a prioritized list by effort vs. potential
I decide which of those opportunities are worth a client's money. The agent is fast at collection and grouping. It has no idea which topics the client's sales team actually cares about.
Cost is low enough to stop being a factor. Verification scans on the rank tracker I built (it combines Search Console, Semrush and DataForSEO, with cost caps) run between $0.37 and $1.67 each.
2. Content operations
This is where I put the most design work, because it's where AI goes wrong most quietly.
- Every factual claim in a draft is traced to a source in a claim ledger before a human ever sees it
- An independent reviewer checks the claims against those sources
- Each review packet carries planted defects, so a reviewer who skims gets caught
On one measured batch, first-pass accuracy was 97.78%: 14 errors in 633 claims. I treat that as an upper bound. The point of the system is that the 14 get found before a client sees them.
3. Client reporting
The agent pulls Search Console and analytics data, compares periods, and drafts the read. That part is commodity work.
The valuable part is the checks I built around it. One month, a burst of bot impressions would have made an RV brand's +8.4% impressions look like −15.1%. That's a 24-point swing in the wrong direction, and it was caught before the deck shipped. An agent that writes a confident narrative on top of bad data is worse than no agent at all.
4. Measurement that wasn't affordable before
Some work only exists because agents made it cheap. In a September 2026 pilot, I ran 60 buyer questions, 4 times each, across 6 AI engines: 1,440 answers for about $8 in API cost. Nobody was going to hand-collect that.
The work that can't be automated
- Strategy and client judgment. Knowing what the client actually needs this quarter, and saying no to the wrong work.
- Writing that has to sound like a person. Agents can draft, but a person with taste decides what's worth publishing.
- Link decisions. Our link program runs through Loganix with an anchor framework and vetting gates. I've cut a $1,900/month link subscription that shipped without review, and stopped a vendor reorder after auditing its campaign at 20% link retention.
- Diagnosing why something moved. When a client asks why traffic dropped, an agent can pull every chart. Deciding which chart matters is still the job.
Those four areas need a person. Most of the rest is data processing.
How the economics actually shift
An illustrative example (not a measurement): say a monthly report takes a strategist six hours to assemble and one hour to sanity-check. Hand the assembly to an agent and you don't get six hours back. You get maybe four, because review gets harder: you're now checking work you didn't do yourself.
That's the part most "agents do 80%" math leaves out. The savings are real, but they move into review and verification. If you don't reinvest some of them there, the errors scale as fast as the output.
What could go wrong
Tool dependency. If one model provider changes pricing or quality, a single-vendor workflow breaks. I route work between two model families on purpose, and I published the routing as an open-source skill (codex-claude-tandem).
Confident wrong output. Agents rarely fail loudly; they usually fail by producing something plausible. That's why the checks above exist, and why I audit my own methods too. When my MarineBench test showed frontier models answering 20 to 44% of catalog questions cold and 100% with the catalog in the prompt, I went back and found circular questions and grader bugs in my own setup.
Client trust. Clients mind being surprised by AI more than they mind AI itself. The answer is being able to show where every claim came from.
Scaling limits. Agents scale output but not attention, so at some point the bottleneck is how much a human can actually review.
Summary
The agency business has always been labor-intensive. If you can move commodity work (data pulls, research, first drafts, reporting assembly) onto software, the economics shift.
An agent isn't better at strategy or writing. It is cheap per run and endlessly repeatable, and its mistakes repeat too. Pair it with human judgment and real verification and you get a team that covers far more ground. Skip the verification and you get more mistakes, faster.