On August 27, 2026, Cisco switched on MyAgent for its roughly 90,000 employees. Every one of them got a personal agent. It runs on Circuit, Cisco's internal multi-model platform, and it works across Outlook, Webex, Jira and SharePoint, taking an objective from the employee and coordinating the steps to reach it. It remembers preferences and past interactions over time. Thimaya Subaiya, Cisco's executive vice president of operations, wrote in the launch post that agentic interactions on Circuit had grown "nearly 350% quarter over quarter" before the rollout. He called MyAgent "the blueprint that helps our customers and partners implement Enterprise AI at scale."
Cisco did not publish a new org chart that day. It did not need to. The org chart changed anyway.
This essay is about that gap, and it is an argument about AI and organizational design. For a hundred years, designing an organization meant drawing boxes and lines. The boxes said what you did. The lines said who you could ask, who could ask you, and whose summary of the world reached the person above. In the last thirty days, three public events showed that those lines have moved into places no org designer looks: the scope of what an agent may read, the rooms it sits in, and what it keeps from one month to the next. Executives still redesign companies by moving boxes. The decisions that shape how their companies think are being made in admin consoles, by people who would never call themselves designers.
The chart was always a routing table
Jay Galbraith gave the cleanest account of why org charts exist in his 1974 paper "Organization Design: An Information Processing View." His claim was simple and a little cold. An organization is a machine for processing information under uncertainty. Nobody can process everything, so you split the work into units, and you build a hierarchy to handle the exceptions the units cannot resolve alone. The more uncertainty, the more information has to move, and the more structure you need to move it.
Read an org chart with Galbraith's eyes and it stops looking like a map of power. It looks like a network diagram. Managers are routers. Span of control is bandwidth. A reporting line is a pipe that carries status up and priorities down, and it is narrow on purpose, because the person at the top of it can only read so much.
Melvin Conway made the matching observation in 1968: the systems an organization builds copy the communication structure of the organization that built them. Put three teams on a compiler and you get a three-pass compiler. The chart does more than describe how people talk. It decides what the company can produce.
Email did not kill the chart, and neither did Slack. Both made transmission nearly free, and the chart survived because the scarce thing was attention. Somebody still had to read the forty status updates, decide which six mattered, and tell the right people. That somebody was usually a manager, and the chart said which manager.
Agents go after that exact job. They read, summarize, and route. With MyAgent, an engineer who wants to know whether a Jira epic in another division will slip does not have to ask her manager to ask the other manager. She asks her agent. The agent goes and looks.
If it has permission.
That conditional clause is where the org chart now lives.
Where the lines went
The old chart had one kind of line. The new one has at least three, and none of them is drawn by the people who drew the first.
Scope: what the agent can read and write
A reporting line used to decide which information reached you. An agent's permission set does the same job, faster and without asking anyone's feelings. If a sales rep's agent can read the product team's roadmap in SharePoint, that rep now has a line into product that bypasses two layers of management. If it cannot, the rep is as siloed as she was in 2015, no matter how flat the chart says the company is.
Most permission sets are inherited. The agent gets what the user has, or what the platform vendor ships as a default, or what an IT administrator set up in a hurry during the pilot. Each of those is an organizational design decision. None of them went through the process a company uses to approve a reorg.
Membership: which rooms the agent sits in
On September 24, 2026, a startup called Ando came out of stealth with $20 million in seed funding from Accel, Index Ventures and Emergence Capital. Ando is a team messaging app, a rival to Slack and Teams, built on one premise: an agent is a first-class member of the workspace. Agents get channels, threads, direct messages, a persistent identity, permissions, and memory. They follow the conversations they are part of and chime in when they judge it useful. Ando works with agents from several providers, Claude and Codex among them.
"Starting from scratch, unencumbered by legacy scaffolding, lets us design around what agents are actually capable of," said Sara Du, Ando's founder and CEO.
Human channel membership was always governed by social friction. You did not join the pricing channel unless you had a reason, and people noticed if you lurked. Agent membership has no friction. An agent added to a channel hears every word and can carry it anywhere else it is allowed to speak. The channel roster used to be a list of colleagues. Now it is a wiring decision about which parts of the company can see each other, and on most teams the person making it is whoever clicked "add."
Memory: what survives a reorg
Cisco's launch materials made persistent memory a headline feature. MyAgent keeps preferences, past interactions and context. That is useful, and it has a side effect nobody put in the press release.
When you move a person from one team to another, the chart updates in an afternoon. The agent's memory moves with the person. It still knows the old team's projects, the old team's customers, the old team's unresolved arguments. The reorg moved the box. The memory kept the old line alive, quietly, in a place HR cannot see and will not think to audit.
Multiply that by 90,000 employees and a few reorgs a year, and the company's real communication structure becomes a sediment of every structure it has ever had.
Half the chart has no author
The same day Ando launched, Dataiku unveiled a product called Agent Management at its Succeed conference. It scans agents across AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio, Salesforce Agentforce, Snowflake Cortex and Dataiku's own platform, builds a single inventory, and assigns each agent a risk tier. General availability is set for October.
The product matters less than the survey Dataiku released alongside it. The company polled 685 chief information officers at firms with more than $500 million in annual revenue, across eight countries, between July 9 and July 29, 2026. Nine in ten said they had complete tracking of their agents. Then 81 percent admitted they lacked complete oversight of agents built outside approved systems, and 84 percent said employees build agents faster than IT can govern them. Dataiku also cited IBM research finding that fewer than one in five organizations keep a complete, current inventory of their AI systems.
Florian Douetteau, Dataiku's CEO, put it plainly. A bank can tell you where every server sits. Ask it about its AI agents, he said, and the answer is "a shrug or a guess."
Read those numbers as an org designer would. Every agent an employee builds is a small reorg. It is a new node in the network with read rights, write rights, a place in some channel, and a memory. It changes who knows what. When 84 percent of CIOs say employees create these faster than anyone can govern them, they are saying that the communication structure of their company is now designed, bottom up, by its fastest builders. Conway's law still holds. The decisions, products and customer experiences that come out of those companies will mirror a structure no executive approved and no executive can describe.
This is the chart nobody drew. It is the one that actually runs the place.
Flattening by layoff pulls the wrong lever
Meanwhile, the most visible organizational design work of 2026 has been done with a severance budget.
On May 5, 2026, Coinbase CEO Brian Armstrong cut about 700 people, roughly 14 percent of the company, and announced a new structure meant to make Coinbase "lean, fast, and AI-native." The plan caps the company at no more than five layers below top leadership, replaces what Armstrong called "pure managers" with player-coaches who also do hands-on work, and sets up AI-native pods, some as small as a single person directing agents that cover engineering, design and product work.
That is a real, deliberate redesign, and Armstrong deserves some respect for saying out loud what it was. It is also a redesign of the human boxes. Layer counts say very little about how information moves once agents do the routing. A five-layer company whose agents are each scoped to one team is still a company of silos, only with fewer people to notice. A nine-layer company whose agents can read across functions, under clear rules, may route information faster than the flat one. The number of layers was a decent proxy for coordination cost when humans carried every message. Agents broke the proxy.
On August 28, 2026, Harvard Business Review published a piece by Faisal Hoque, Tom Davenport and Paul Scade arguing that AI transformation requires redesigning work rather than cutting roles. They called many AI layoff announcements "AI-washing," ordinary restructuring dressed in AI language for investors, and cited a Goldman Sachs estimate that AI had reduced U.S. monthly payroll growth by about 16,000 jobs over the past year, roughly 0.1 percentage points on the unemployment rate. The labor market has moved far less than the press releases suggest.
I agree with their prescription and would push it one step further. Most of the work redesign that AI demands is an access redesign. The question every CEO announcing a flattening should be able to answer is this: after the cuts, which agent can read what, on whose behalf, and who decided? If the answer is a shrug or a guess, the flattening changed the payroll and left the actual organization to chance.
AI and organizational design, done on purpose
There is a better way to do this, and it does not require buying anything. It requires deciding that the permission graph, the channel roster and the memory policy are organizational design, and then treating them with the seriousness a company gives a reorg.
Draw the access graph before you redraw the chart
Before the next restructuring, map every class of agent in the company. For each one, write down what it can read, what it can write, which human it acts for, and which human answers for it. Do it for sanctioned agents first and then go hunting for the rest, because Dataiku's survey says the rest exist.
The map will show you your real org chart. It will show which functions can already see each other through agents, which cannot, and where a single employee's homemade agent has become the only bridge between two divisions. That last pattern is common, and it is fragile. When that employee leaves, a line in your company disappears and nobody files the paperwork.
Treat scope changes like reorgs
Moving a team from one vice president to another takes approvals, a memo and an all-hands. Giving the whole sales organization's agents read access to the product backlog is a bigger change to how the company works, and today it takes an administrator and a ticket.
Reverse that. Any change that widens what a class of agents can read across a functional boundary should go through the same review a reorg does, with the same people in the room. Narrow changes inside a team can stay fast. The cross-boundary ones are structural, so govern them as structure.
Make channel membership a design decision
Ando's premise, that agents belong in the conversation as members, is probably right, and it will spread to the incumbents. Once it does, each team needs a rule for which agents sit in which rooms, and what they may carry out.
A legal channel with an agent that also sits in the sales channel is a new reporting line from legal to sales. Maybe you want that. Decide it on purpose. The default, where anyone can add any agent anywhere, produces a company where everything leaks sideways and nothing escalates upward, which is the worst of both structures.
Write memory rules for people who move
When an employee changes teams, their agent's memory should change with a policy behind it. Some context should travel, because the person's judgment and relationships are why you moved them. Some should be sealed, because the old team's customer disputes and unreleased plans have no business in the new team's drafts.
Cisco is instructive here for a second reason. According to CFO Mark Patterson, speaking to Fortune on July 1, 2026, Cisco built much of its agent stack on-premises and routes each request to whichever model suits the task. "It knows which tool is most effective and most efficient," he said. Cisco owns the layer where scope and memory get set. A company that buys agent seats from a vendor and accepts the defaults has handed that layer, and with it a large share of its organizational design, to the vendor's product team.
Put the CHRO and the CISO at one table
Organizational design has always sat with HR and the executive team. Permissions have always sat with security and IT. The two groups never had to share a pen, because a permission used to be a narrow technical fact about who could open which folder.
That changed when the thing holding the permission started reading, summarizing and acting across the whole company on someone's behalf. Now a permission is a reporting line, and a reporting line is a permission. The CHRO who designs the chart without seeing the access graph is designing a diagram. The CISO who sets access without seeing the chart is designing the company without knowing it. Put them in one meeting, on a fixed schedule, with the access map on the table.
What the Cisco rollout should tell a board
Cisco's finance team already uses agents to write 80 to 90 percent of the first draft of its MD&A, according to Patterson. That is the management discussion section of a public company's filings, the part where executives explain the numbers to investors. A few years ago, that draft moved up a reporting line, from analysts to a controller to the CFO, and every hop was a person reading and cutting.
Some of those hops are now an agent reading the ledger directly, with whatever access someone granted it. The document still gets signed by a human. The path it took to reach that human has changed shape, and the org chart on the board's slide does not show the new shape at all.
This is what AI does to organizational design. It leaves the chart standing and routes around it. The boxes still exist. The lines that matter now run through permission sets, channel rosters and agent memory, and in most companies those were configured by default, in a hurry, by people with no mandate to shape the organization. The companies that redesign well over the next two years will be the ones whose leaders noticed where the pen went and took it back.
Take the pen back
A tool will not fix this. Dataiku will sell you an inventory in October, and you should probably have one. An inventory tells you what exists. It does not tell you what your company should look like, which functions should see each other, which memories should travel, or which human stands behind each agent's reach. Those are architecture decisions about how your organization thinks, and they need someone to make them deliberately, before the defaults harden into structure that nobody chose and everybody inherits.
That work sits at the join between strategy, people and systems, which is exactly where most companies have no owner. Agor AI Advisory works in that join. We map the access graph you actually have, compare it to the organization you mean to run, and design the permission, membership and memory rules that close the gap, so that the next reorg you announce changes how your company works and not only the slide. Every week you wait, your fastest builders draw another piece of the chart for you.
Sources
- Cisco Blogs, "MyAgent and the Rise of Ambient Intelligence," August 27, 2026
- PYMNTS, "Cisco Deploys Custom AI Agent to Entire 90,000-Person Workforce," August 27, 2026
- Fortune, Cisco CFO Mark Patterson on AI agents, July 1, 2026
- GlobeNewswire, "Ando Launches Agent-Native Messaging Platform, Announces $20 Million Seed," September 24, 2026
- SiliconANGLE, "Dataiku debuts cross-platform Agent Management," September 24, 2026
- Harvard Business Review, "AI Transformation Requires Redesigning Work, Not Cutting Roles," August 28, 2026
- Fortune, Coinbase replaces "pure managers" with player-coaches, May 5, 2026
