The report that took a day now builds itself in 90 minutes

At most agencies, the deliverables that win and keep clients are handmade. A competitive analysis for a prospect, a priced proposal, a quarterly strategy review: each one is a skilled person’s full day of gathering, pasting, formatting, and second-guessing. The work is valuable. The assembly is not.

I built an AI agent system for a marketing agency that automates its three core deliverables end to end. Work that took an analyst most of a day now runs in about 90 minutes of largely unattended build time, and any team member can kick it off.

Deliverables are pipelines wearing a trench coat

Watch someone build a competitive analysis report and you will see a pipeline: pull the prospect’s rankings, pull the competitors’, run the audits, screenshot the evidence, interpret the gaps, write the narrative, assemble the deck. Every step has defined inputs and outputs. Humans just happen to be executing it by hand, badly, with tab-switching in between.

Once you see the pipeline, the automation writes itself:

• Deterministic steps stay deterministic. Data pulls, audits, math, and formatting are scripts. No model involved, no model needed.
• Judgment steps get an agent. Interpreting what the data means, choosing what to lead with, and writing the narrative are LLM work, fed by the deterministic steps and constrained by templates.
• A human signs off at the end. The system produces a finished draft, not a published deliverable. Review is minutes, not hours.

The unattended part is the point

Ninety minutes of build time sounds slow next to a chatbot answering in seconds. But nobody is sitting there. The analyst starts the run, goes back to real work, and comes back to a complete draft. The metric that matters is not latency. It is unattended time: how much of the build happens while no human is paying attention.

The goal of an agent system is not to make work instant. It is to make work happen while nobody is watching it.

That framing also changes who can operate it. When the process is a button instead of a craft, the most junior person on the team can produce the agency’s best-structured deliverable. The senior people move up the stack to reviewing and selling.

What made it reliable

Three unglamorous decisions did more for reliability than any prompt engineering:

1. Structured handoffs. Every stage writes typed, validated output for the next stage. An agent never receives a vague blob of context.
2. Retry with evidence. When a step fails validation, the system retries with the error attached. Most failures self-heal on the second pass.
3. Everything on file. Every run leaves its inputs, outputs, and drafts on disk. When something looks off, you read the trail instead of guessing.

None of this is exotic. It is the same discipline as any good data pipeline, applied to work that agencies still treat as artisanal.

If your team has a deliverable that eats a skilled person’s day on a schedule, that is the first candidate. Get in touch and I will tell you whether it is pipeline-shaped.