My AI minions got organizational charts this quarter. Full job descriptions, team assignments, reporting structures, and measurable output. World domination is still on the table… mostly… first, we’ve got to train the army and it might take a while.
Most IT leaders I talk to are using AI the same way they used Google in 2005. You go in, you ask a question, you get an answer, you close the tab. The idea that AI agents for IT leaders could actually run a meaningful chunk of your workday is still theoretical for most people. I get it. A year ago I was there too. Then I started building the thing I wanted to exist, and I have not stopped since.
What I have now is not a chatbot habit. It is a 47-agent workforce organized into functional teams, backed by a structured Airtable operating system I call NEST, and running on two AI tools with very different jobs. This morning, before lunch, my agents generated my daily briefing, drafted and sent four client emails, loaded 48 new contacts into NEST, wrote and structured this blog post (which I then rewrote, I mean they aren’t perfect yet), ran a software license audit for a nonprofit client, and built a project plan for a new AI assessment engagement. I drank my coffee. I reviewed and approved (and also approved a stack of documents from Accounting that were printed on paper and stapled together. That’s ancient technology for those of you who are wondering). This is what my day looks like.
Here is how I built it, what it actually looks like, and why the three-tier framework of Skills, Agents, and Agent Teams is the architecture that finally made it stick.
The Two-AI Setup That Makes This Possible
Before the agents make sense, the foundation has to make sense. I run two AI tools with completely different functions.
Claude is my thinking partner. When I am staring at a vendor contract, working through a governance decision, or trying to figure out why a migration is failing, that is Claude. Conversational, exploratory, no structure required. I call her Bestie or Brain for Exploration, Strategy, Thinking, Ideas, and Everything
Claude Code is my execution engine. It has file system access, it runs scripts, it reads my calendar, it calls APIs, it writes to databases. Claude Code is where the agents live and where work actually gets done. I call her Bestie #2 or Build Engine for Systems, Tools, Infrastructure, and Execution. The distinction matters enormously. If you are trying to use one AI tool to do both things, you are going to frustrate yourself. Pick the right tool for the job before you build anything on top of it.
I wrote more about building the right AI habits before you automate anything in this post on AI prompt habits. The short version: get disciplined about how you ask before you start delegating to machines.
Tier One: Skills Are Your SOPs for AI
The first thing I built was not agents. It was Skills.
A Skill in Claude Code is a markdown file that lives in a .claude/skills/ directory. It contains instructions for how to handle a specific task type. Frontmatter controls whether you invoke it with a trigger word or whether Claude loads it automatically when it detects the right context. Think of it as the difference between briefing a temp contractor every single morning versus having a trained employee who already knows the procedure.
My most used Skills right now:
/briefing— Pulls my ConnectWise tickets, Airtable task board, calendar, and flagged emails into a formatted morning report. Delivered at 6am./commit— Handles git commit formatting, changelog updates, and version notes for any repo I am working in./connectwise— Manages the full ticket lifecycle: status, time entry, notes, agreement assignment. All the rules I kept having to re-explain are now baked in./meeting— Takes a Plaud transcript, extracts structured notes and action items, formats them for the client, and populates Airtable.
Skills are not agents. They do not have personalities or independent judgment. They are reusable instruction sets that any agent can call. This Skills layer is where AI agents for IT leaders become genuinely repeatable rather than one-off parlor tricks. Building it first was the right call, because it forced me to actually document my own workflows before I tried to automate them. If you cannot write down exactly what you do and in what order, you are not ready to delegate it to an AI.
The Anthropic Skills documentation is a solid starting point if you want to see the technical structure. The mental model is simple: Skills are the training manual. Agents are the employees who read it.
Tier Two: AI Agents for IT Leaders Need Real Job Descriptions
This is where things get interesting and, I will be honest, a little absurd in the best possible way.
I have 47 named agents. Each one has a defined role, a personality, a set of tools it is permitted to use, and a clear scope. They run as background workers, dispatched in parallel when needed, and they return findings to me rather than demanding my immediate attention. That last part is critical and I will come back to it.
A few of my favorites to give you a feel for how this works:
April is my Inbox Lieutenant. She monitors email every 45 minutes, triages by urgency and client, drafts responses in my voice, and cross-references my calendar before suggesting any commitments. She has saved me from responding to things that did not need a response, which turns out to be a significant percentage of my inbox. She is based off my real life former admin April who my boss stole because she’s amazing. I am not bitter about this at all. Not at all…
Stuart is my blog writer. He is writing this post right now, which is either a meta moment or a philosophical crisis depending on your perspective. Stuart knows my voice, knows Yoast requirements, knows the minions joke is non-negotiable, and knows never to use an em dash. He was given a job description and he does the job.
Hugo is my plan validator. Before any project plan moves forward, Hugo reviews it against a structured rubric and issues a pass, conditional pass, or fail with specific notes. He has caught scope gaps that would have caused real problems. I stopped skipping project reviews once I had Hugo doing them automatically. Also he’s a thorn in my side. He rejected my 3AM idea. Told me to go get coffee and come back.
Segrid is my project manager. She builds the plans that Hugo then validates. She creates project folders, sets milestones, assigns task owners, and maintains the project structure in NEST. Segrid and Hugo have a working relationship I did not plan for but am grateful exists. Seg is based off real life Seg who is the most amazing Project Coordinator I know.
Archie is my knowledge collector. His job is to scan meeting transcripts and capture tribal knowledge: who prefers which communication style, which vendors have burned us before, what a client said six months ago that turned out to matter. Archie feeds the NEST Knowledge Base so that information stops living only in my ADHD brain.
Warren is my business strategist. When I am evaluating a new initiative, Warren looks at it across the three tracks I care about: client delivery, internal operations, and AdaptoIT product development. He asks the questions I forget to ask when I am excited about something new.
Radar is my AI scout. He watches for developments in AI tooling and connects them to specific work I am doing. When a new capability ships that affects my clients, Radar surfaces it before I find out from a vendor.
The complete roster runs to 47. Here is the OrgChart. But the point is not the number. The point is that each agent has one job, and they are good at it because that is all they do. That is the same principle behind a good org chart at any company. Every person on your team needs a job description. So does every agent.
Tier Three: Agent Teams Change the Calculus
Individual agents are useful. Agent Teams are where AI agents for IT leaders go from helpful to genuinely transformative.
Teams are functional groupings of agents that work together toward shared outcomes. Each team has a lead agent who synthesizes input from the specialists under them and surfaces a single coherent picture to me. I am not managing 47 conversations. I am managing a handful of team leads.
Here is how the teams break down:
Client Directors — I have a dedicated client director for each major client base. Each director agent synthesizes input from three specialists: a security agent, a helpdesk agent, and an infrastructure agent. When something happens with a healthcare organization client, the helpdesk agent flags the ticket pattern, the security agent checks for risk indicators, the infrastructure agent notes the environment context, and the client director gives me one consolidated picture. This is what a good account team does. Mine runs in the background without a weekly sync.
Content Team — Dave brainstorms angles and post ideas. Stuart, Otto, Bob, and Kevin each write in specific formats (long form, quick takes, case studies, technical how-tos). Phil audits every post for Yoast compliance before it leaves my hands. The pipeline is: idea to draft to audit to my desk for final review. I stopped treating blog writing as a block of time I could never find.
Dev Team — Stella handles frontend, Bruno handles backend, Margo runs QA, Nigel manages DevOps pipelines, and Russ handles Azure architecture. This team built and maintains the AdaptoIT product infrastructure. They are not replacing developers. They are doing the tedious structured work that gets in the way of developer thinking.
Operations — April handles inbox, Hector manages ConnectWise ticket creation and routing, Dewey handles time entry backfill, Chip tracks task completion in Airtable, and Cron runs the scheduled briefings. Operations is the team that keeps the lights on without my intervention.
The Claude subagents architecture supports teams coordinating directly rather than routing everything through me. That is the feature that made team-level delegation possible. Before that, I was the bottleneck in my own AI workforce, which is a special kind of irony.
The Agent Queue: How I Stay in Control Without Drowning
Here is the part nobody talks about when they pitch AI automation to IT leaders: if every agent talks to you immediately when it finds something, you have just replaced your inbox with a different inbox. You have not solved anything.
The Agent Queue is how I avoided that trap. It is an Airtable table inside NEST where agents surface questions, findings, and requests for approval. I work through it on my own schedule, usually twice a day. Agents do not interrupt me. They post to the queue and continue whatever else they are doing.
This morning the queue had seven items: two draft emails from April waiting for approval, a flag from Radar about a new Claude Code feature, Hugo’s conditional pass on a project plan with two items to address, a contact enrichment batch from Tim ready to load, a content brief from Dave for next week, and a vendor risk note from one of the client director chains. I cleared all seven in about twenty minutes.
The queue is the single most important architectural decision I made. Without it, the agents would have created more cognitive load than they relieved. With it, I am the decision-maker, not the dispatcher. That is the correct role for a CIO using AI agents for IT leadership work.
NEST: The Operational Backbone
All of this runs on top of Airtable, which I have built into what I call NEST. The tables that matter most:
- Tasks — Everything in flight, owner assigned, connected to ConnectWise ticket where applicable.
- Contacts — Client contacts, vendor contacts, prospects. Enriched and maintained by Tim and Archie.
- Agent Queue — The inbox described above. Agents write here, I review here.
- Blog Pipeline — Every post from idea through draft through publish. Content Team writes to this table.
- Plans — Project plans reviewed by Hugo, built by Segrid, tracked to completion.
- Knowledge Base — Tribal knowledge captured by Archie from every meeting, call, and transcript.
Airtable is my executive function tool. I have ADHD, and the combination of ADHD and a complex multi-client CIO role is a recipe for things falling through cracks. NEST is what closes the cracks. The agents write to it. I read from it. Nothing lives only in my head anymore.
I wrote about AI and executive function tools in more depth in the AdaptoInbox post if the inbox management piece is where you are feeling the most pain right now.
What This Morning Actually Looked Like
I want to be specific because vague AI success stories are useless. Here is today, April 8, 2026, by noon:
6:00am: Briefing generated by Cron. ConnectWise tickets sorted by client and urgency, Airtable tasks filtered to due today and overdue, today’s schedule with prep notes for each meeting, overnight email summary with three flagged items, and one risk note from the healthcare client director chain about a pattern in helpdesk ticket volume. Formatted, readable, done. I did not open ConnectWise until 9am.
7:30am: April’s morning triage complete. Four draft emails in the queue, two of which I approved as-is, one I rewrote the closing paragraph on, one I killed because the situation had already resolved. Total time touching email: eleven minutes.
8:15am: Tim completed a 48-contact enrichment batch from a recent event list. He pulled LinkedIn titles, mapped each contact to the correct NEST client record, flagged six as potential prospects with brief context notes, and loaded the full batch. I spot-checked five records. All accurate.
9:00am: Segrid delivered a project plan for a new AI assessment engagement. Folder structure created, milestones set, task owners assigned, ConnectWise project ticket populated. Hugo reviewed it within ten minutes, issued a conditional pass, and flagged two scope questions: one about deliverable format and one about stakeholder approval authority. I answered both in under five minutes. Plan approved and in motion.
10:30am: The infrastructure agent on the nonprofit client director chain completed a full software license audit. Clean summary in the queue: 47 licenses active, 3 unused seats flagged for reclamation, no compliance gaps. No action required from me today, but I have the data when the renewal conversation comes up next month.
11:00am: Stuart delivered this blog post draft. Phil reviewed it for Yoast compliance and flagged eight items. I am fixing them now, which feels appropriately recursive.
None of this required me to hold context across tasks, context-switch at someone else’s schedule, or remember to do things. The agents managed the sequencing. I made the decisions. That is the correct division of labor. Also during this time, I am attending meetings of my own.
Getting Started with AI Agents for IT Leaders
Do not start with 47 agents. Start with one Skill.
Pick the workflow that costs you the most cognitive overhead every single day. For most IT leaders that is either the morning briefing, the inbox, or meeting follow-up. Write the Skill first. Document the steps, write the markdown, test it until it does exactly what you would do. That exercise alone is worth doing even if you never build an agent on top of it.
When the Skill is solid, build one agent whose entire job is to run that Skill and report back. Give it a name. Give it a personality that fits the role. Hold it to the same standard you would hold a new hire: clear scope, defined outputs, feedback when it misses. If the agent is vague about what it does, that is a you problem, not an AI problem. Tighten the job description.
Then add the queue before you add more agents. The queue is the discipline that keeps you from building a system that runs you instead of the other way around. This sequencing matters: Skill first, then one agent, then the queue, then more agents. Skip any of those steps and you will hit a wall and blame the tools.
One thing worth reading before you go too deep: I wrote about where low-code and AI automation hit their limits in this post on when low-code stops helping. The same traps apply here. Knowing where the guardrails are before you build saves a lot of backtracking.
Expect the first month to feel slower, not faster. You are building infrastructure. The payoff is not in week one. I spent about three weeks just writing Skills and testing them before I deployed a single agent in a real workflow. That investment is why April does not send embarrassing emails on my behalf.
AI agents for IT leaders are not magic and they are not hype. They are infrastructure. You build them deliberately, you maintain them, and they compound over time. The 47th agent is useful because the first 46 already know how the operation runs.
The org chart for my full agent roster is live here. If you want to see exactly how the teams are structured and what each agent’s job description looks like, start there.
And if you want AI working on your inbox before you have time to build any of this yourself, AdaptoInbox is where I started, it will be up soon for testing.
World domination is still on the roadmap. My agents are handling the scheduling.
Minion Army Org Chart