The Self-Driving Comms Operation
The tools are ready. Are communications teams?
Sam Altman and other AI leaders like to talk about a phenomenon called the capability overhang. It’s a shorthand for the gap between what these frontier LLM models can accomplish today and what people are actually doing with them. I’ve spent the last year consulting with comms teams on AI tooling alongside the usual work of narrative, positioning, and crisis, and I can confirm: in communications, the overhang is enormous.
Here’s what the state of the art looks like at most organizations right now. The company’s engineering or IT team has deployed a platform. ChatGPT, Claude, Gemini, a custom harness that lets people pick models by task. The comms team has access. They use it periodically. And what they use it for is, almost universally, some version of “smarter Google”. Paste in a draft, ask for revisions. Provide some background materials and ask for a first pass. Summarize this. Rephrase that. Conversational search with better prose on the back end.
The engineering teams knew immediately what to do with them. The design teams probably had ideas. The comms team got a login and a pat on the back.
Nobody has shown them what’s possible.
Users and Builders
I’ve written before about the difference between being a user of AI and being a builder with it. That distinction matters even more here than it did when I was talking about shipping silly interactive projects on a weekend.1
Every comms team I talk to is trying to become better users of AI. Better prompts, better workflows, a template library, maybe a series of lunch-and-learns. That’s the wrong frame.
The point is building systems: wiring together your knowledge, your workflows, your standards, and your tools into something that operates continuously, not something you just open in a browser tab when you need a first draft.
I know this because I built one.
I run a one-person strategic communications consultancy. No more than five concurrent clients, spanning from growth-stage private companies to Fortune 500 and everything in between. Over the past year, I’ve built what I call a self-driving comms operation: roughly 60 integrated components running on a backbone of Claude/Claude Code, Obsidian, Todoist, Granola, and a pile of connectors and scripts that would probably cause a normal person to back slowly out of the room.2
I use the word “operation” deliberately here. Not self-driving strategy. Not self-driving judgment. Not self-driving client relationships. The operations: prep, research, scheduling, monitoring, formatting, task routing, the thousand small acts of administrative hell that eat your days if you let them.
Let’s go through a normal day. At 5:30 a.m., a script scans my calendar and creates transit buffers with real drive times. At 6 a.m., an inbox sweep classifies overnight email by urgency and turns anything actionable into tasks. A morning briefing, assembled from dozens of feeds and tailored searches, lands on my phone before I wake up.
When I sit down, two commands set up the day. A morning sweep triages every open task, classifies each one as something AI can handle, something it can prep for me, or something that needs my brain, and dispatches parallel agents to start executing the approved items. Then a time-blocker pulls the remaining tasks, maps them against my calendar with real-time drive distances, and builds a schedule down to the five-minute buffer, complete with errand batching, gym slots, and rollover recommendations for anything that doesn’t fit.
Throughout the day, meeting transcripts get scanned for commitments I made, and tasks appear automatically with full context. At night, a process walks my entire knowledge base and weaves backlinks between related documents.
Then there’s the queue. Whenever I hit something I know Claude can handle (draft this follow-up, research this company, update this note, clean up this task list), I tag it #dothisclaude. Every hour, a job picks up tagged tasks, executes them against my full set of tools and context, and reports results to me via Telegram.
The work just... appears. I went from pulling work through a conversation to pushing tasks into a queue that resolves itself. That is a very different relationship to AI than opening a fresh chat window and, once again, explaining the entire backstory of a task you already explained three sessions ago.
All-in monthly cost of all this: a middling dinner for two in San Francisco.3
Does it actually help get things done? I had Claude audit my daily work throughput for the last six months, measured by tasks marked completed in Todoist, normalized for complexity and scope. The analysis showed a 2.5 to 3x increase from when I didn’t have this system in place to when I did.
And that’s just throughput. It doesn’t capture the four hours a day back for work that requires my brain. There are projects I’ve taken on since building this (outlining a book, building open-source tools, being much more diligent about getting Person Familiar out every week) that I simply would not have had the bandwidth to attempt before.
The work that defines our profession isn’t coverage reports. It’s stubbornly, unavoidably human: creating narrative tension, reading a room, managing situations with real people and real emotions, finding ways to do things nobody’s done before. AI tools are literally derivative: their outputs are derived from what came before. We’re in the business of new because new captures attention and moves hearts, minds, and markets.
So the focus on operations is deliberate. Automate the machinery so the people doing the work have time and space for the thinking, the creativity, and the strategy that no system can replicate. Operations matter enormously, but they’re the layer you can actually put on rails.
The Hard Part Wasn’t the Tech
Of course, this was easier for a solo practice. The knowledge lives in one head, the voice is one voice. I was the architect, the domain expert, and the end user simultaneously. But the lessons still translate directly.
The hard part was never the AI. The hard part was articulating what I used to do by feel. Why does this draft sound wrong? Why does this meeting prep feel incomplete? Why do I keep tightening this same kind of opening paragraph? Why does one calendar invite feel polished and another look like it was assembled by raccoons?
Every time I answered one of those questions explicitly, the system got better. It developed standards, checks, and rules. We refined its editing patterns and made them context-specific. The discipline wasn’t technical. It was self-knowledge.
That discipline is exactly the same whether you’re one person or fifty or five hundred. The institutional knowledge that makes a comms team excellent already lives inside the team. Your senior VP who can brief the CEO before a hostile reporter call without breaking a sweat. The media lead who reads a reporter’s tone and knows before the story runs whether it’ll be fair. That internal comms manager whose editing instincts are so good that everyone routes drafts through her before anything goes out.
That knowledge is the asset. That’s the work I’ve been doing with comms teams over the past year: not building the system for them, but helping them see their own operation clearly enough to build the right one.
This is also where I’ve seen the most common failure mode. Someone decides to build the system for the team. They interview everyone, map the workflows, disappear into a room, and emerge with a framework, a Notion workspace called “Comms OS,” and a workshop nobody asked for. Six weeks later, the team is doing the work the old way. The system that got built reflected how an outsider understood the team’s work, not how the team actually does it. The people inside the operation can feel that difference immediately. Systems built for teams collect dust. Systems built with teams compound.
And compounding is the thing. You codify one workflow, and you realize six others could work the same way. You structure one piece of institutional knowledge, and the whole context layer starts getting richer. The first time you help someone on a team build something, not use a tool but build a thing, the whole frame shifts. I’ve watched it happen enough times to know it’s reliable. That first build is the unlock, and everything after it moves faster.
It’s Time to Start
None of this requires starting with sixty components. The first build that usually matters most (and the one I’d recommend to any team trying this for the first time) is automated meeting processing. Connect your transcript tool and your task manager. After every call, the system pulls out commitments, updates the relevant files, and drafts the recap. An afternoon to set up, maybe two. Thirty to forty-five minutes saved from the first week.
That’s it. That’s the first rung. Once you have it, something shifts. The project files get richer every time a transcript is processed, which means every future meeting prep improves, which means the system starts earning its own keep. People stop asking “does this AI stuff actually do anything?” and start asking “what else can we put on rails?” Every recurring frustration is a spec for the next build.
Every C-suite in technology is pushing organizations to show results with AI. CEOs and CFOs at public companies need to speak with specificity about what AI has actually changed in their operations. And right now, almost no comms team has a great answer.
That gap is an opportunity, and it won’t last. The teams that build this now, thoughtfully, with their people rather than for them, will have what amounts to an unfair structural advantage that grows every week.
Your comms operation, whether in-house or at an agency, already works a certain way. There’s a rhythm: people check things, prep things, route things, update things, chase things down. Some of that work requires the full weight of their professional judgment. A lot of it doesn’t. Look at the parts that don’t.
That’s where this starts. Not with a transformation strategy or a vendor evaluation or a slide deck about the future of communications. With one workflow at a time that makes next Monday morning better than last Monday.
Workplace Buzzword Slots remains, in my professional estimation, a masterwork.
This includes a photography alert system that checks astronomical calculations, fog predictions, tide tables, and atmospheric soundings, then texts me before dawn when conditions are worth shooting. I contain multitudes.
If you want to go deeper on the tech behind this system, you can check out this piece I wrote as well as this GitHub repo.