Anthropic on July 8, 2026 published an account of how its own marketing operations team rebuilt two of the most manual jobs in the function - the weekly metrics report and the campaign infrastructure behind every event - around its Claude Cowork agent, compressing work that once took days into hours.

The post, credited to Ian Chan and Annabel Custer of the marketing operations team, describes workflows already running inside the company rather than a product launch. It arrives roughly six months after Anthropic first opened the automation capabilities of Claude Code to non-developers, and it offers an unusually detailed look at what those capabilities produce when a technical marketing team applies them to its own recurring workload.

The framing is candid about a contradiction familiar to anyone in the discipline. Marketing operations teams own automation as their mandate, yet a large share of the day-to-day remains stubbornly hands-on. According to Anthropic, martech tools do not integrate cleanly with one another, reports get consolidated by hand, and landing pages are spun up one at a time. The company put the tension plainly: while automation sits firmly in the team's purview, a lot of the work is anything but automated.

The weekly report that ate two days

Before the change, Ian Chan spent one to two days a week assembling the weekly marketing metrics review for marketing and leadership. The reason was structural rather than a matter of effort. In an ideal setup, every metric would already live in a dashboard and the analyst's job would reduce to writing the narrative. In practice, the data is scattered: some metrics sit in the dashboard, some have not yet moved from the data warehouse into it, some have not been piped into the warehouse at all, and new ones might exist only in a Slack message or a call transcript.

That fragmentation is the specific problem the agent was pointed at. According to Anthropic, the business moves faster than a traditional reporting pipeline can keep up with, and the manual reconciliation of these sources is what consumed the analyst's time. Claude Cowork now handles most of that data hunt.

What actually filled the two days

The number worth interrogating is the two days itself, because on its face a week is a long time to build a metrics report in 2026, an era with no shortage of dashboards and business-intelligence tools that assemble one automatically. The source account makes clear the time was not spent drawing charts. It was spent getting numbers into a trustworthy state before any writing could begin, and the metrics arrived at four different levels of readiness.

Some sat in the dashboard already, defined and rendered, and cost almost nothing. Others lived in the data warehouse but had no dashboard view built for them, so the figure had to be pulled by direct query and formatted by hand. Others had not been piped into the warehouse at all and had to be retrieved from whatever upstream system held them. And some existed only in a Slack message or a call transcript, unstructured in any database, recoverable only by reading the thread or the transcript, locating the number, and understanding the surrounding context well enough to trust it.

Layered on top of retrieval was validation, which is where the hours compounded. Because the figures came from systems that do not define things identically, they did not always agree, and reconciling a disagreement is not a lookup. It means working out why two sources diverge and which definition is correct. Anthropic's own example is a sales-team reorganization that left marketing's reporting no longer matching the sales team's figures. Multiply that kind of chase across every metric that failed to line up, add the framing work of deciding the week's narrative and producing a separate leadership slide from the same data, and the two days are accounted for. None of it is slow report assembly. It is the cost of data that was not yet clean, defined, or centralized.

That distinction matters because it explains why the abundance of reporting tools did not help. Automated reporting solves the last mile: once every number is defined, validated, and sitting in one governed place, drawing the report is trivial. What it cannot solve is a metric that lives only in a transcript, a definition two teams dispute, or a reorganization that silently breaks the join between two systems. The bottleneck sat upstream of anything a dashboard touches.

A workaround, or a fix?

There is a skeptical reading the vendor-authored account does not volunteer, and it deserves stating plainly. If a company routinely generates decision-grade metrics that surface only in Slack threads and call transcripts, the more fundamental problem may be an underbuilt data pipeline rather than a uniquely hard reporting task. An agent that hunts down scattered numbers every Sunday evening automates around that gap; it does not obviously close it. The disputed definitions still exist. The metrics still originate outside the warehouse. The agent simply retrieves and reconciles them faster than a person could, week after week.

The account gestures at this without pressing on it. Anthropic notes that Chan now has bandwidth to go deeper into the data layer and ensure Claude interprets numbers, definitions, and regional structures the same way the warehouse does, which is close to conceding that those definitions were not fully settled beforehand. Whether an agent papering over an immature data foundation counts as a genuine fix or a durable workaround is a real question, and it is not one the company has an incentive to dwell on. What can be said is narrower and still useful: the reported gain is real within its own terms, and the mechanism is retrieval and reconciliation at speed, not the elimination of the underlying fragmentation.

How the scheduled run works

A scheduled task runs every Sunday evening. It prompts Claude to read the previous week's review and the latest meeting transcript, check Slack for what the sales team is focused on, query the data warehouse, and leave a folder containing the numbers alongside a few suggested focus areas. On Monday morning, Chan opens Claude Cowork and pulls an initial report that already contains the metrics tables and suggested headlines.

The human decision comes next. Chan reviews the suggested areas of focus, confirms or redirects where the narrative should sit, and then instructs Claude to expand the chosen threads with supporting details and examples. The emphasis shifts week to week: some weeks respond to a sales priority, others to a product launch. At the quarter turn, Chan tells the agent to lead with quarterly plans and feeds in the quarterly review document. From the same data and narrative, Claude generates the leadership slide covering what changed, why, and what teams are doing about it. Follow-ups become Asana tasks.

One design detail stands out for reliability. When the numbers do not line up, the agent flags the mismatch rather than guessing. Anthropic gives a concrete instance: after a reorganization on the sales team, marketing's reporting no longer matched the sales team's figures, and Claude surfaced the gap and asked how to handle it instead of silently reconciling the two.

Three skills doing the work

The process runs on connectors to the team's marketing platforms plus three custom skills that Chan maintains and revises continually. A prep skill drives report assembly, covering focus selection, headlines, and expansion with supporting detail. A proofreading skill checks every number in the draft against a verified source. An action-items skill converts follow-ups into Asana tasks.

The maintenance loop is itself part of the method. At the end of each weekly session, Chan asks Claude to summarize what came up that should feed back into the skills: a new sales reorganization structure, corrections he made, or a different way he wanted headlines framed. The net effect, according to Anthropic, is that a process that once took up to two days now takes up to two hours.

The reclaimed time did not vanish into idleness. A meaningful share of it moved toward helping other marketers frame their questions, refine their prompts, and interpret the numbers they pull themselves from Claude. Chan also gained room to work deeper in the data layer, ensuring the agent interprets numbers, definitions, and regional structures the same way the data warehouse does. Human validation, the company notes, has become integral to the workflow rather than incidental to it.

Event builds across four vendors, handled by agents

The second workflow tackles what Anthropic calls one of the most manual processes in marketing: standing up the infrastructure behind campaigns. Every event, webinar, or integrated campaign has to be configured in the CRM, in the marketing automation platform that runs email sequences, and in the event management platform that hosts the registration and landing pages. Each of these is typically a different vendor, and the integrations between them are rarely complete.

Annabel Custer, who focuses on campaign operations, previously picked up every request from a dedicated Slack channel and worked through that sequence by hand. Her rebuilt setup is now almost entirely handled by Claude, and it begins with an intake form on which requesters specify what they need: an event build, a data import, apply-to-attend support, or approval support.

A dispatcher that only routes

Once an hour, a dispatcher skill reads the channel, selects the most urgent request, stamps the ticket so the work is not duplicated, and hands it to one of five specialist skills. The dispatcher performs no setup itself. Its sole job is deciding what runs next, and keeping routing separate from execution lets Custer refine each specialist skill in isolation without disturbing the logic that assigns work.

For an event build, the most complex request type, an event-build skill handles the full sequence end to end: CRM campaign creation, the marketing automation campaign with its workflows and lists, event platform setup, email drafting, landing page generation, and every integration between them. When the build finishes, it hands off to a separate agent for audit.

A fresh agent that checks the work

The audit step is where the architecture gets deliberate about trust. The audit skill runs on a separate, fresh Claude instance that starts with no prior context. It submits a test registration on the live landing page, opens the confirmation email in Gmail, and marks the Asana task complete only if everything looks right. Custer reviews each result before it ships.

The full specialist set numbers among the skills Custer has built and updates as she encounters new edge cases. Alongside the dispatcher and event-build skills, a webinar-landing-page skill spins up landing pages for webinars, an apply-to-attend skill handles in-flight changes to the registration flow, an approval-support skill manages event approvals and sends the appropriate emails on a scheduled cadence, and a data-import skill scrubs lists and processes attendee data. Custer also keeps a separate manager agent open; when a run misfires, she asks it to examine what happened and propose an adjustment, and anything worth keeping goes back into the relevant skill.

Notably, the stated motivation was not primarily speed. According to Anthropic, while the automated workflows will become significant time savers, Custer's main reason for building them was quality. As a marketing team scales, marketers cloning event pages from whatever template is nearest can introduce bugs, such as confirmation emails surfacing the wrong city name or broken landing pages. The agent-driven approach yields consistency across builds at scale, and it frees Custer to concentrate on enablement and on optimizing campaign architecture for better insights.

The playbook Anthropic put on record

The post closes with four specific recommendations for marketing operations teams, framed as what worked internally rather than as prescriptions. Turn repeated corrections into skills: when the same correction recurs, that feedback belongs in a skill, and Claude can build the skill rather than requiring the person to. Build a proofreading skill first, so that every number traces back to a verified source. Ask Claude to reflect after the first runs of a new workflow, on the theory that the model reads instructions differently than a human writes them, and feed what surfaces back into the skill. Lean on scheduled tasks, because work that runs on its own every Sunday night or every hour is work no one has to remember to do.

Why this matters for marketing operations

The account lands in a period when PPC Land has documented a steady procession of marketing platforms wiring their data directly into Claude and rival assistants. The connective tissue in most of those stories is the Model Context Protocol, the standard Anthropic introduced to link AI models to external data and tools. Where earlier coverage tracked vendors opening their systems to agents, this post documents the demand side: a team consuming those connections at volume to run its own operations.

The timeline behind it is instructive. Anthropic opened Claude Code's automation power to non-developers through Cowork on January 12, 2026, initially as a research preview for Claude Max subscribers on macOS. In the months since, the surrounding ecosystem filled in rapidly. Meta opened its advertising infrastructure to Claude and ChatGPT via MCP connectors in late April, marking the first time the platform allowed third-party AI systems direct access to live advertiser accounts. B2B measurement vendor Channel99 connected its cross-channel performance data to Claude, ChatGPT, and Copilot through an MCP server in February. Contentsquare later plugged behavioral data into third-party agents on the same architectural logic that data should flow to the agent rather than the user to the dashboard.

That logic runs directly through Anthropic's own workflows. Both the reporting build and the event build depend on connectors reaching into CRM, marketing automation, and event platforms, precisely the category of live integrations the wider MCP buildout has been enabling across the sector.

The reliability questions the post addresses are the same ones practitioners have raised elsewhere. When PPC Land first covered Cowork's debut for non-developers, the analysis flagged that brand voice consistency and factual accuracy remain human responsibilities regardless of agent capability, and that AI systems can produce plausible-sounding content containing errors. Anthropic's answer, as described in this account, is procedural: a dedicated proofreading skill that verifies every figure against a source, an audit agent that starts fresh and tests the live output, and human sign-off before anything ships. Hallucination in reporting was among the first concerns raised in public discussion of the post.

One shift is worth naming as the throughline. As the agent absorbs the reconciliation and the click-through setup, the human work concentrates on defining data, validating outputs, building processes, and explaining why two systems disagree. That reframes marketing operations from execution toward governance and enablement, and it is being modeled by the company that builds the underlying agent, which is both the account's value and its limit as evidence.

Timeline

  • January 12, 2026: Anthropic launches Claude Cowork as a research preview for Claude Max subscribers, extending Claude Code's file automation to non-developers on macOS.
  • February 24, 2026: Channel99 launches an MCP server connecting its cross-channel B2B marketing data to ChatGPT, Microsoft Copilot, and Claude Cowork.
  • April 29, 2026: Meta opens its advertising infrastructure to Claude and ChatGPT via MCP connectors and a companion CLI.
  • July 8, 2026: Anthropic publishes the account by Ian Chan and Annabel Custer detailing how its marketing operations team automated weekly reporting and event builds with Claude Cowork.

Summary

Who: Anthropic's marketing operations team, specifically Ian Chan, who owns the weekly metrics report, and Annabel Custer, who focuses on campaign operations.

What: A published account of two internal workflows rebuilt around the Claude Cowork agent. The first compresses the weekly marketing metrics review from up to two days to up to two hours using three custom skills and a Sunday-night scheduled run. The second automates event and campaign infrastructure across CRM, marketing automation, and event platforms using a dispatcher skill, five specialist skills, and a separate fresh-instance audit agent.

When: The account was published on July 8, 2026, roughly six months after Claude Cowork opened to non-developers on January 12, 2026.

Where: Inside Anthropic's own marketing function, with the workflows running on connectors to the CRM, marketing automation, and event management platforms the team uses, plus Slack, Gmail, the data warehouse, and Asana.

Why: Marketing operations teams own automation yet spend heavily on manual reconciliation and repetitive campaign setup. Anthropic frames the effort as recovering time and, in the event-build case, improving quality and consistency at scale, while shifting the human role toward data definition, validation, and enablement.