We sell a SaaS platform called Marketing OS that runs a digital marketing agency end-to-end with 60+ Claude-powered AI agents supervised by humans. We also use it ourselves. SourceForge runs marketing for nine product brands — SourceForge ERP, Gold & Jewellery ERP, Hotel Management ERP, Restaurant POS, OCR Payable Agent, AI Assistant for BC, Language Lab, Marketing OS itself, CognitiveIQ — through a single Marketing OS workspace. One operator runs it. This is the operator's-eye view.
The shape of the work
Each product gets a brand workspace inside Marketing OS. Brand voice is captured separately for each (Gold ERP's voice is not Marketing OS's voice). Channel mix differs — Business Central products lean heavily on LinkedIn and Microsoft partner channels; CognitiveIQ leans on Google search and parent communities; Language Lab is school-by-school direct outreach.
Across nine brands the daily volume of marketing work is:
- 6-12 blog posts authored across categories (most generated, all human-reviewed)
- 30-50 social posts across LinkedIn, X, Instagram per day
- 15-25 paid-media creatives per week
- 200-400 outbound prospecting emails per day
- 6 monthly client reports compiled by AGT-083
A single operator approves everything. The operator's role has shifted from "writer who reviews AI" to "editor and strategist who approves agent output". The bottleneck is no longer production volume; it's judgment.
What the agents are good at
Volume work. Anything that requires producing 50 variants of a Google Ads RSA, 30 angles for an Instagram caption, 12 cuts of an email subject line. The agents shine when you'd previously have asked a junior writer for a list and graded the list.
Research. AGT-003 (competitor teardown) and AGT-010 (prospect research) consistently surface points we'd have missed manually. AGT-024 (SERP tracker) catches ranking shifts faster than we used to.
Structured generation. SEO content briefs, ad copy variants, multi-platform adaptations of one core post. The Claude prompts are tightly scoped and the output is reliably useful.
Continuity. Brand voice across nine products is more consistent now than it was when we had three external writers. The Brand Voice Enforcer scores every piece against the voice fingerprint and flags drift.
What the agents are not good at
Original strategic insight. The agents will produce excellent material on a topic; they will rarely tell you that the topic itself is wrong. Strategy still happens in a room with humans.
Customer empathy. An agent can write to a buyer persona; it cannot replace the felt sense of what a Bangalore café owner actually cares about. The operator role exists in part to provide this layer.
Long-horizon judgment. Knowing that a brand voice should shift over six months as the company matures — agents don't see this. They optimise for the prompt as written.
Tone for sensitive moments. Customer complaints, post-incident communications, crisis comms — these stay with a human. We don't even try to automate them.
How hiring changed
A year ago, the SourceForge growth team would have been three people — a content writer, a paid-media specialist, an analyst. Today it's one operator running Marketing OS.
The role profile we hire for is different from "marketing manager". We look for:
- Editorial judgment — can read a generated piece and know what's good
- Prompt engineering instinct — can refine an agent's output by adjusting its inputs
- Comfort with AI as a colleague, not a tool
- Strategic literacy — can think two quarters ahead about a brand
We pay this role at the senior end of the market because the leverage is enormous. One operator at this level produces more brand impact than the three-person team they replace, at a fraction of the cost. We do not call this "AI-augmented marketing"; we call it "marketing".
What we measure
The numbers we look at across the portfolio:
- Cache hit rate on Claude prompts (target 80%+, currently 84% average)
- Cost per published deliverable (blog post, ad creative, email, report)
- Approval rate per agent (what fraction of generated drafts the operator approves with no edits)
- Time from brief to publish (was days, now hours)
- Pipeline contribution per brand per week
- CSAT on automated customer responses (when we let agents reply)
The metric that moved most: time from brief to publish. A blog post used to be a five-day cycle through a freelance writer. It's now a same-day cycle. The implications cascade — we can respond to news within the news cycle, which fundamentally changes what marketing can do.
The edge cases that take all the time
Eighty percent of the work runs smoothly. The remaining twenty percent is where the operator earns their salary.
- An agent confidently produces a wrong statistic. (Resolution: tighten the system prompt, add a fact-check tool.)
- A brand voice drifts subtly over weeks; the Enforcer scores everything pass but the cumulative effect feels off. (Resolution: weekly human re-calibration.)
- A customer complaint thread that an auto-reply agent handled poorly. (Resolution: pull this conversation type back to human-only, log the pattern.)
- A regulatory change that requires every ad to add a disclaimer. (Resolution: update the compliance agent's prompt, re-run last week's drafts.)
These edge cases are why the operator role can't be automated away — yet. Possibly ever.
What this means for an agency considering this
If you're a marketing agency considering running on Marketing OS — or any equivalent multi-agent platform — three things to expect.
First, your headcount will compress. Probably by 60-70% for the production roles. Some of that headcount can redeploy to strategy and client success; some won't fit and won't stay.
Second, your gross margin will widen substantially. Production cost per deliverable falls by an order of magnitude. Your pricing doesn't have to drop the same amount, so the difference is margin.
Third, your strategic skills become the differentiator. The thing you used to charge for — production volume — is now table stakes. The thing you can charge premium for — judgment, taste, narrative — is what's left.
Our own experience suggests this transition is hard but worth it. We're a happier company on the other side, and a more profitable one. Whether your agency makes the same transition depends on how willing leadership is to redesign the org chart from scratch. Bolting AI onto an existing structure doesn't work; we tried it for six months in 2024. Rebuilding the org around AI does. That's the choice.