The team had Claude. Their marketing data was connected to it. They used it in chat to answer ad-hoc questions about campaign performance and pipeline. They got good answers. And every Monday morning, an operator still spent the better part of an hour pulling metrics together for the team standup. Every Friday afternoon, someone had to do the same thing with the exec update. They still scrambled to assemble the presentation for their monthly channel deep dive.
We’ve seen this chat bloat across clients. More data connected to Claude means more exploration means more disconnected chats. There’s still internal updates that need to happen. Those updates get marginally better but the team gets no time back and quickly gets overwhelmed.
The path to scheduling analytics agents
Carving out analytics tasks that a dedicated agent can run on a schedule not only reduces your team’s chat bloat, but it improves your operating discipline and intelligence. Here are the phases we went through with this B2B tech startup to build their analytics agents.
Connect your data.
Connecting marketing data to Claude seems simple. If you’re working inside a large company with an infosecurity team, chances are it is not.
We connected through MCPs with user-level OAuth so each agent's access matched the operator's. Read-only. We started with the team's most active channel before broadening to the rest of the stack. Single-pipe tools like Supermetrics skip the per-source MCP plumbing if you have the budget. For this team, Supermetrics plus a handful of MCPs covered the full marketing stack.
Layer in business context.
The agents read the same set of markdown files at the start of every session: the company's products, ICPs, active campaigns, channel KPIs. Without that context, an LLM looking at marketing metrics produces generic commentary.
This is the prerequisite the whole stack depends on. We've written separately about what goes into the context layer and how to structure it: Context engineering for marketing agents.
Dial in prompts using a Claude Project.
A Claude Project is the prompt-tuning surface. It’s where everyone on the team runs their ‘data chats’.
The Project Instructions give every chat rules for ‘how to be a good data analyst’. Your business and product context live as Files.
- Don’t make assumptions.
- Lead with the observation.
- Tie claims to numbers.
- Less is more.
- If you don’t know, say it.
We used the Project for ad-hoc queries and manually-generated weekly reports for weeks before scheduling anything. That's where the shape of Routines emerge.
Routines are the graduation, not the start.
The Project is where you dial in the prompt; the schedule is where you ship it.
Structure scheduled prompts as sequenced tasks
The discipline rules say what good looks like. The task sequence is how you get the model there. Every Routine we ship breaks the work into eight steps — same skeleton, different specifics:
- Persona setting. "You are a data analyst." Sets tone for everything downstream.
- Gather context. Read the business-context markdown. The agent's first move is orienting itself, not reaching for the data.
- Gather data. Pull from Supermetrics and the MCPs. Terminate and alert the operator if the fetch fails — don't let the agent improvise around missing inputs.
- Stage and treat the data. Exclusions, bucketing, time-window normalization. State the prep rules explicitly so the analysis runs on a known dataset.
- Produce the core dataset. A single named artifact — for example, a weekly view of clean marketing metrics, this week versus last. Every downstream task references it.
- Perform the analytical task. Now identify the top 3 movements, score the channel, fill the KPI table. Not before.
- Produce the output. Apply the template. Follow the copy instructions verbatim. This is where credibility is won or lost.
- Take action. Post to Slack. Drop the doc in Drive. Ping the operator. The Routine isn't done until something visible has happened.
Two things make this skeleton worth copying. Separating "produce the core dataset" from "perform the analysis" forces the model to commit to a clean view of the data before reasoning over it. Rules to terminate without a complete dataset help reduce hallucinations.
The three agents we shipped
Same skeleton. Three different cadences and outputs.
What Moved Monday. Posts the top three most meaningful metric movements from the prior week to Slack at the start of the team standup. The team spends 5 minutes reviewing, then moves on. Prompt lesson learned the hard way: identify the observation, don't diagnose why. The model nailed the top 3 immediately and produced sloppy explanatory copy until we tightened the prompt. The agent has routinely caught things the team would have otherwise missed.
Friday Exec Update Generator. Pulls the predefined KPIs by channel against targets and writes a 1–2 sentence insight per channel for a senior non-marketing reader. The agent produces a Google Doc; an AppScript listens for the output and formats the exec-facing version the leadership team actually opens. The insight slots have strict instructions that hit these principles: every word does a job, don’t guess, blank is okay.
Weekly Channel Deep Dive. Rotates Email → Ads → Content → Website on a four-week cycle. Three blocks per channel: a metrics scorecard against the channel's KPIs, a list of what the team shipped in the last month with associated metrics, and reasoning plus recommendations on what to do next. The cadence the team already had. The agent's job is to write the draft.
What we'd tell another team
Three positions, defended briefly.
Chat is the first milestone, and doesn't get replaced. Every marketing team should have a Claude Project connected to their data, loaded with business context, and instructed to ‘act like a good analyst’. It handles the messy "wait, why did this happen" investigations that don't fit any schedule. That capability earns its keep before any Routine exists. Don't skip it on the way to the schedule.
Nail the landing or no one trusts the agent. Even when the upstream data is right and the analysis is sound, slop output tanks credibility. Treat output format as a core part of the system, not a finishing touch.
People should still analyze data. Agents don't replace analytics talent. They let analytics talent analyze instead of aggregate.
This is an anonymized account of work performed by Lobo Growth for a client engagement.