In 2024 every demo claimed an agent could run your entire marketing operation. By 2026 most of those companies are either pivoting or pretending they always meant something narrower. The reality is in between: agents do real work, but only on a specific set of tasks, and only with real scaffolding.
Here's what's working for us, and what isn't.
What "agent" means in practice
For our purposes, an agent is a model that can call tools, run multiple steps, and decide what to do next without a human approving each step. Claude Code, the OpenAI Agent SDK, and a handful of frameworks (Mastra, LangGraph, n8n with AI nodes) all qualify.
The bar to call something an agent isn't autonomy in the abstract — it's whether it strings together more than one decision before reporting back. A single LLM call that returns text is not an agent. A loop that calls a search tool, reads the results, calls a write tool, checks the output, and either continues or stops is.
The four places we get value
1. Reporting
Agents are good at reading dashboards, pulling numbers, and writing the same paragraph a human would write — only faster and at the same hour every Monday.
Our weekly client reports are now generated by an agent that:
- pulls last week's spend, conversions, and cost-per-conversion from Google Ads
- compares to the prior week and the same week last year
- reads the recent change history (campaign edits, budget changes, creative swaps)
- writes a 5–8 sentence summary in our voice
- flags anything anomalous for human review
A human still reviews and sends. The agent saves 30–45 minutes per account per week. Across 25 accounts, that's a half-time hire we don't need.
2. Audit and triage
Agents are good at running through a checklist. Account audits — the kind where you check 40 settings, 20 campaign-level decisions, and 10 ad group hygiene items — are exactly the kind of work where an agent saves hours.
Our intake audit is a checklist with 73 items. The agent runs through all of them, produces a draft report, and surfaces the 5–10 items that need a human judgment call. The human then writes the summary and recommendations. An audit that used to take a senior PPC strategist 4 hours now takes 1.
3. Drafting
This one is well-known but worth saying clearly: agents are excellent first-draft engines for any structured artifact — ad copy, landing page copy, briefs, follow-up emails, ad-grant applications. They're bad at finishing. They're great at starting.
We use them to get to a draft 80% of the way. Then a human pushes it to 100%. Total time per artifact drops 60%.
4. Data entry and migration
Moving a list of 200 keywords from a spreadsheet into Google Ads with the right match types, campaigns, and ad groups is the kind of task that puts a junior strategist to sleep. An agent with the Google Ads API does it in 8 minutes without typos.
Where agents still fail
Three categories of work where we've stopped trying to automate:
Strategy
The decision of what to bid on, which services to advertise, which cities deserve a budget tier — these are judgment calls that depend on conversations with the owner, knowledge of the local market, and judgment we can't easily put into a prompt. Agents that "do strategy" produce confident-sounding output that's wrong in ways nobody on the team would catch in a hurry.
We let humans decide strategy. We use agents to execute it.
Anything client-facing without a human in the loop
A client in 2026 expects that the human they hired is reading their messages. Even if an agent can write a perfect Slack reply, it shouldn't send it without a human glancing at it. The first time an agent confidently misroutes an issue or invents a number, the trust that took two years to build is gone.
We use agents to draft replies. A human always presses send.
Long-running creative tasks
Generating 15 headlines? Great. Maintaining a year-long content calendar with brand consistency, internal links, and seasonal pacing? Not yet. Agents lose the plot over long-running creative work in ways that are obvious to a reader but invisible to the agent.
What it costs
Worth being honest: this stuff isn't free.
Our agent infrastructure runs on Claude API plus a small set of MCP tools we wrote in-house. Total monthly cost for a 25-account agency is around $300 in API spend. The development cost is real — we put roughly 80 hours into building and refining the audit and reporting agents over six months.
That's the bar. You're paying with engineering time up front to save service-delivery time forever after. If you don't have someone on the team who can write or maintain the prompts and tools, you'll either pay an outside builder for the same thing or end up with a half-working pile of automations.
The point
Agents in 2026 are real, but they're a junior employee, not a chief of staff. They handle the structured, repeatable, high-volume work. They free up humans to do the strategy, the relationships, and the judgment calls. Treat them that way and they'll quietly compound. Treat them like a manager and you'll spend a year explaining to clients why their report has wrong numbers in it.