The Community Assistant AI Work Map Report

The Method

Here is the machinery, because the map is only as trustworthy as the method underneath it, and because you should be able to argue with it. That's not a courtesy. That's the design.

Jobs to Be Done as the backbone. We built on the Jobs to Be Done framework that Clayton Christensen carried into the mainstream: the discipline of describing work by what it accomplishes rather than by the tool that does it. Theodore Levitt's old line, which Christensen loved, is the whole idea in one sentence. People don't want a quarter-inch drill, they want a quarter-inch hole. Communities don't want a welcome email. They want to feel welcomed.

So every job in the Work Map carries three dimensions, and this is where the method starts doing real work:

• Functional: what the job accomplishes. Deliver a welcome, introduce the norms, connect the new member to relevant spaces.
• Emotional: what it makes a person feel, or protects them from. Make the new member feel seen; reduce first-post anxiety.
• Social: what it does to status, belonging, identity. Signal that this is a community that notices each person.

Hold those three dimensions up against any job and something clarifying happens: you can suddenly see which part of the work is logistics and which part is meaning. A machine can execute the functional dimension of a welcome flawlessly. Whether the emotional and social dimensions survive automation is a different question entirely, and it's the question, job by job, 298 times so far.

Community types as the organizing axis. We organized the map by organizational archetype (B2B SaaS customer communities, professional learning, trade associations, developer and open-source communities, cause and advocacy, brand and consumer) because that's how practitioners actually self-identify. Nobody introduces themselves at a conference as "a lifecycle-stage specialist." They say "I run a developer community" and everyone in earshot immediately knows a dozen true things about their week.

The layered model. Early in the research we made a mistake worth admitting, because correcting it produced one of the map's most useful properties. Our first inventories repeated the common jobs in every archetype (welcome the newcomer, moderate the gray areas, watch the community's health) restated six times with cosmetic differences. When we extracted the genuinely universal work into a shared Base Map, our original B2B inventory shrank by more than a fifth. Most of what looked specific to a type of community was simply the job.

So the map works like this: the Base Map holds the invariant form of every universal job. Each archetype holds only what's genuinely its own, the jobs unique to its world plus its flavored versions of common jobs, because moderating for beginner safety in a learning community is truly not the same job as moderating regulated claims in a brand community. The Base Map plus your archetype equals your complete job. Nothing restated, everything inherited, every flavored job linked to the parent it specializes.

The classification. Every job is typed one of three ways, and each type has a test you can run yourself:

• Agent: the whole job is pattern-matching, retrieval, scheduling, or monitoring. The test: if this ran unattended for a month, would anything break that a weekly review wouldn't catch?
• Human: the job's value is the human doing it. The test: would automating this, even perfectly, destroy the thing it produces? An automated thank-you is not a smaller thank-you. It's a different object.
• Hybrid: the job splits cleanly. The agent carries volume, vigilance, and logistics; the human carries judgment, relationship, and authority. And here's the rule that keeps Hybrid from becoming a shrug: every Hybrid job in the map must specify its full coordination protocol. What initiates it, exactly what the agent does, exactly what the human does, and the precise trigger that passes control from one to the other. A Hybrid without a crisp handoff trigger is a misclassified job. We treated that as law.

Alongside the journey stages, every job also carries one of eight work categories (strategy, content and programs, community management, tools, governance, measurement, culture, leadership), a second, independent axis for filtering the map by the kind of work rather than the moment in the journey. Those categories will look familiar to anyone who knows The Community Roundtable's competency work, and that's deliberate lineage, gratefully acknowledged: ideas live inside relationships, and this map stands on prior maps.

Cadence, not calendar. One more mechanic: every job carries a structured cadence. Daily, Weekly, Monthly, Quarterly, Annually, Event-driven, or Continuous. Cadence is an attribute, not a stage; it's what lets the map answer "what does Tuesday look like" as readily as "what does the profession look like."

And the part that matters most: this is where the work begins, not where it ends. What you're reading is our initial research, the first public draft of our thinking, not a finished or certified taxonomy. It was made with Claude, not by it: our direction, our inputs, our decades of pattern and scar tissue, our curation and editing, with an AI doing what we argue AI should do, carrying the volume so the judgment could stay human. We hold ourselves to the same disclosure rule we're about to argue your community deserves. And now the validation begins, in public, with you. Every job has a permanent ID so it can be cited and challenged individually. Every classification carries its written rationale, the argument, not the assertion. The dataset is versioned with nothing ever silently overwritten, and where our own confidence is provisional, the flags are visible in the data on purpose. We've checked the structure so you can argue the substance. A map you can't argue with is a poster. This one is asking for the argument. That's how it becomes everyone's.

The Findings

Here is the distribution, walked through slowly, because every number in it breaks somebody's favorite story.

Thirteen percent of the work is fully Agent. Forty jobs. Eligibility checks, access provisioning, sign-up flows, integration monitoring, metrics collection, routing, the dashboards. This is the busywork in its purest form, and here's the finding hiding inside the small number: the fully automatable slice of community management is the smallest category on the map. If your mental model, or your executive's, is that this profession is mostly clerical work awaiting a script, the inventory simply does not agree. The busywork is real. It is also bounded, and now it is named.

Twenty-seven percent is fully Human. Eighty jobs where the classification test came back unambiguous: automating this, even perfectly, destroys what it produces. Resolving conflict between members. Holding the community together in a crisis. Recognizing the contribution that actually mattered, because automated gratitude spends trust and human gratitude compounds it. Personally reaching the drifting member who matters. Coaching the volunteer moderator through the hard call at eleven at night. Governance. The stakeholder relationships where the relationship is the deliverable (a rule that became a design principle with its own number), and any research layer underneath it is a tool, not a lane.

Read that list again and notice what it is. It isn't a list of tasks. It's a job description for a more senior role than most community managers hold.

And sixty percent (178 jobs) is Hybrid. This is the finding. The majority of community management is neither safely human nor cleanly automatable. It's work a human does with AI, work that splits, and the value lives at the seam. Which is why we forced every one of those 178 jobs to declare its handoff, and why the handoffs turned out to be where the map teaches the most.

Because when you line up the Hybrid protocols, one pattern repeats until it stops looking like a finding and starts looking like a law:

Read the protocols closely and you see it's never just an alarm. The agent tracks behavioral decline against each member's own baseline, weights it by what's at stake, and hands the human a prepared moment: the who, the history, the why-now, the suggested next move. The human doesn't arrive at the situation cold; the human arrives briefed. The agent filters the moderation queue at volume and clears the obvious, then delivers the gray-area case with the member's history attached, because the gray area is where context, mercy, and precedent live. The agent detects the sentiment shift at two in the morning, assembles what changed and who's affected, and the human decides what it means by nine. Preparation is the quiet half of the pattern, and it's the half that turns automation from an alert into productive engagement.

Here's what that looks like as an actual entry, one detect-and-respond pair, straight from the map:

That pair is the hybrid model in miniature. Formally cross-referenced, one per community type across the whole map, seven pairs in all: the machine holds the vigilance and sets the table; the human owns the moment that matters. And here's the built-in-public confession: one of our own inventories had merged that pair into a single blurry "prevention" job. The data around it argued for the split. The data won. It usually did.

One honest limitation, before the next finding, because a skeptical reader has earned it. A count of jobs is not a count of hours. When we say 13% of the work is fully Agent, we mean 13% of the jobs. The map does not claim that 13% of your week is automatable, because the map doesn't yet weight jobs by time. A single Human job like resolving a member conflict can eat a day; a single Agent job like access provisioning runs in milliseconds forever. Time-weighting is exactly the kind of finding this map should earn with real usage data rather than assert in version one, so we haven't. The classification tells you what kind of work each job is. What share of your Tuesday each kind consumes is a question we intend to answer with you, not for you.

What surprised us most wasn't in the member journey at all. Late in the build, reviewing stage distributions, we found a hole in our own method, and it deserves owning plainly, because we turned out to be exactly like the profession we were describing. We had mapped the member's journey exhaustively. We had mapped the program's journey. We had mapped everyone's work but our own.

Community managers are professionally excellent at making other people's journeys visible and constitutionally terrible at making their own count. So the map now has an eighth stage, Operate, holding the twenty jobs of running the practice itself, across seven disciplines: planning and prioritization, situational awareness, reporting and stakeholders, reflection and iteration, the personal intelligence system, learning and practice, resourcing. The stage axis of the whole map resolves to one question: whose journey does this job serve? Six stages follow the member's. Grow follows the program's. And Operate follows the practitioner's: the running of the program itself, and of the community manager's own career alongside it. Put simply, Grow increases the network effect; Operate runs the conversations.

Two Operate jobs deserve a sentence each here, because they are the easiest to skim past, and the costliest to overlook. Keep score, the weekly discipline of capturing your accomplishments and impact into a durable record, earns its place in the Base Map because community managers' impact is chronically invisible in standard metrics, and career capital you never wrote down is career capital you don't have. And guard against cognitive outsourcing, the continuous, irreducibly human discipline of verifying rather than deferring, of noticing when the tool is telling you what you want to hear, is, this report will argue, the distinguishing professional skill of the next five years. The practitioners who keep their judgment sharp while everyone else's quietly dulls will not need a defense of their value. They'll be busy.

One more finding, and this one is a known property of the dataset. When we tagged every job's operating rhythm, the calendar-driven work (the weeklies, monthlies, quarterlies) turned out to be the minority. 230 of the 298 jobs, 77%, run event-driven or continuous: triggered by a member joining, a thread turning, a signal crossing a threshold, or simply never stopping at all.

Sit with what that means. Community management is not a calendar job. It is a vigilance job, and vigilance is precisely what humans are worst at sustaining and agents are built for. That single finding, more than any thesis statement, is the practical case for the hybrid model: the agent holds the watch so the human can hold the relationship. Nobody should be proud of answering the 3 a.m. signal personally. Someone should be proud of what they did with it at nine, arriving prepared.

The Argument

Now the thesis, in full, with the map underneath it instead of adjectives.

The story being sold to your executives is replacement. The story being whispered among practitioners is defense. The map says both stories are wrong, and it says so with numbers. The fully automatable work is 13% of the inventory, the irreducibly human work is 27%, and the decisive 60% is work that gets better when a human and an agent split it along a well-drawn seam: the agent watching and preparing, the human responding.

So here is the argument, made to a skeptical CFO, a curious journalist, and a tired practitioner in the same room:

AI takes the busywork so community managers can expand into the work only they can do. Not "protect." Expand. Look at the Human column of this map (trust, context, conflict, culture, recognition, crisis, governance, organizational partnership) and notice that most community managers today get to spend a fraction of their week there, because the Agent and Hybrid columns are eating their calendar one routine task at a time. The honest promise of the agent layer is not efficiency. It's reallocation: hours moved from the 13% and the agent-side of the 60% into the 27% where careers, and communities, are actually built.

One line from co-host Marius Ciortea has become one of our halcyon calls:

That's the stakes, stated precisely. The practitioners in danger are not the ones who adopt agents. They're the ones who cede the wrong ground, who let the automated welcome absorb the human welcome, who let the templated re-engagement replace the personal one, who hand the gray-area judgment to the filter because the filter was already open. Do that and the box builds itself. You become the supervisor of a system that does the legible parts of your old job, while the illegible parts, the parts that were the point, quietly stop happening, and stop being missed until the community is a support forum with fast responses and no humanity serving its greater purpose.

The map's answer is a pair of governance principles, and they're written into the data itself. The first we ended up calling one-way sacred: agents hand up to humans on ambiguity, emotion, stakes, or authority; humans may delegate down; agents never absorb a human job by drift. Drift is the enemy. Not the technology, the drift. The second is simpler, and it protects the members rather than the manager: disclosure, always. Any conversational answer an agent gives a member is labeled as machine-made, with sources where possible. Non-negotiable, in every community type. A community that can't tell who's talking to it isn't being served by AI; it's being experimented on. And disclosure is fast becoming infrastructure, not etiquette. The Content Credentials standard (C2PA) now embeds signed provenance, including AI involvement, directly into images and media, with the major AI labs adopting it this year and regulators beginning to require transparency labeling. We'll carry those credentials on our own published media as the tooling matures, and this report practices what it preaches: you already know exactly how it was made. Trust is the product of this profession, and both principles exist because trust, once spent, doesn't refund.

Which brings the argument to its most consequential point, and the reason the map is a management tool and not a manifesto. Deciding what to automate and what to protect is itself human work, maybe the most senior human work in the whole inventory. A community manager who walks into the budget meeting with this map isn't defending a headcount. They're presenting an operating model: here is everything my function does, here is what I'm delegating and the protocols governing it, here is where I'm irreplaceable and what I'll do with the reclaimed hours, and here is how we'll know it's working. That person has stopped answering do we still need you, a question the map reveals was always malformed, and started answering the real one: what does this role become when it's designed for augmentation — technology holding the watch and carrying the busywork, so you can finally do the work you got into community for?

This report is written by an optimist, the earned kind, seasoned across enough technology cycles to know that the fear is always sincere and the outcome is always a choice. We don't need to stop this. We need to shape it. The profession that names its work first gets to write the terms of its own amplification.

The Invitation

Everything in this report comes from a map you can open, filter, argue with, and extend, right now, in its first public version.

And because the map was built for your Tuesday, not our thesis, here's how to put it to work this week:

And argue with is meant literally. Every one of the 298 jobs carries its classification rationale in the open (the argument, not the assertion) and a visible flag wherever our own confidence is provisional. Every job also carries a feedback button, and we're asking you to use it for more than corrections: tell us how you apply a job in your world, how you adapted it, what you'd revise, what we've never named that you do every Tuesday. Application stories are validation data. They're how initial research becomes shared knowledge. Everything submitted is captured for review, nothing publishes without human eyes, and accepted contributions ship with your name on them. This is a living map built in public by a profession that has always been better together than alone; a rising tide is the only kind of tide worth raising. The operating ethos is simple: we are better off when you are better off. That's not a slogan here; it's the model, the model for better communities and a better world.

The map is also the spine of everything else we're building around it. The CommunityAssistant.ai show works the map out loud: practitioners, builders, and honest skeptics stress-testing the classifications against real deployments, platform-agnostic on principle, because these frameworks have to work on whatever stack your organization actually approves. The newsletter delivers the same discipline in inbox form: one idea, one workflow, one tangible, every issue. And a forthcoming section of this report will take the analysis one layer down into the tooling itself: the state of AI capability across the major community platforms. That piece is in scoping now, and it will hold to the same rule as everything else here. Evidence over hype, and we publish what the tools can't do with the same confidence as what they can.

This work exists because we believed, long before we could prove it, that communities are where people become more than customers, more than users, more than audiences. This map is the proof stage of that belief. And the difference now, for the first time, is that the watch can be handed to something tireless, and the humans given back to the human work.

The map is open. The work is named. Let's get to it.