Your Company Is Not Integrated. Your Best People Are.
Why adding AI agents to disconnected systems may make the hidden coordination layer faster, busier, and harder to see
If you are new to the publication, Welcome to The Executable Company explains the larger question behind this work.
Consider a familiar sales opportunity. The CRM says it is open and moving forward. The customer’s hesitation is buried in an email exchange. A meeting transcript contains a promise that nobody formally approved, while the delivery plan was written before the scope changed. Finance knows that the proposed price can no longer support the work, and the project system shows that the people required to deliver it are already committed elsewhere.
Every system may be technically correct. Taken together, they describe a situation that is operationally false.
Someone inside the company usually knows this. They do not know it because they have access to a secret database or because their title grants them a complete view. They know because they remember why the customer hesitated, understand what the delivery lead meant when they questioned the estimate, recognize that the proposal and the current scope no longer describe the same work, and know which person has the practical authority to interrupt the process before the company makes a promise it cannot fulfill.
They then translate the situation for people who hold different pieces of it, bring the right people into the conversation, and follow through after the official workflow says that the decision has been made.
We describe this person as proactive, experienced, commercially aware, good with stakeholders, or simply someone who gets things done. Those descriptions may all be accurate, but they can obscure an architectural fact: the company looks integrated because the person is integrating it.
The integration layer hiding inside the organization
Organizations are usually described through the work that is easiest to see. A proposal is written, a budget is approved, a release is shipped, a ticket is resolved, or a deal is closed. Each of those activities depends on another layer of work that receives far less attention because it does not belong cleanly to a single role, workflow, or system.
Someone has to reconstruct the current state from representations that were created at different times for different purposes. Someone has to preserve the reason behind a decision after the meeting ends. Someone has to translate what sales calls a commitment into what delivery understands as scope and what finance recognizes as exposure. Someone has to remember which assumptions were still true when the original plan was made, notice when an exception has invalidated the normal route, and verify that the completion of a task produced the consequence the company intended.
Researchers of cooperative work have described part of this as articulation work, meaning the work required to align people, tools, schedules, representations, and contingencies so that the formal work can proceed. Kjeld Schmidt and Liam Bannon developed this argument in their 1992 paper, “Taking CSCW Seriously: Supporting Articulation Work”. Susan Leigh Star and Anselm Strauss later examined how organizations distribute visibility and silence around different forms of work in “Layers of Silence, Arenas of Voice”.
This kind of coordination cannot and should not disappear completely. Situations change, plans remain incomplete, people interpret the same evidence differently, and relationships matter. A difficult customer conversation is not technical debt merely because it consumes time. Negotiation may be the real work. A team may need to talk in order to discover what it believes rather than retrieve a pre-existing answer from a system.
The structural problem begins when the same integration has to be reconstructed repeatedly, remains invisible in the operating model, and depends on individual vigilance. At that point, the organization does not really possess the capability. A person is lending it to the company every day.
I use the phrase human middleware to describe this condition, not a category of worker. It is not a dismissal of producers, project managers, account managers, executive assistants, operations leaders, or anyone whose work crosses organizational boundaries. Much of what these people do requires judgment, trust, negotiation, timing, and care. The phrase describes what happens when the company also uses them as transport, synchronization, memory, format conversion, status reconstruction, monitoring, and exception routing because its systems cannot perform those functions reliably.
The distinction matters. We should not automate the human contribution that gives the work value, but neither should we romanticize the organizational failure surrounding it. Human capability is too valuable to spend on carrying fields between systems, chasing decisions that no workflow owns, and remembering commitments that the company never learned to preserve.
Why agents do not automatically solve this problem
The current AI conversation is beginning to acknowledge that the main constraint is no longer confined to model capability. Microsoft’s 2026 Work Trend Index reports that active agents in the Microsoft 365 ecosystem grew fifteenfold over the previous year, while also finding that many employees are moving faster than the organizations around them. In its analysis, organizational conditions such as culture, manager support, and talent practices were associated with more than twice the reported AI impact of individual factors. A capable person with a capable tool can still be trapped inside an operating environment that was not designed to use either of them well.
The products themselves are also moving beyond conversation. OpenAI’s workspace agents can gather context and take action across connected tools, continue work across multiple steps, remember what they have learned, and run on a schedule. The examples include agents that monitor feedback across channels, create prioritized tickets, prepare recurring reports, qualify leads, draft follow-ups, and update a CRM. This is a meaningful shift because software can now perform parts of the connective work that previously required a person to gather information, move between applications, and prepare the next action.
At the same time, the emerging discussion about agent sprawl suggests what happens when companies scale activity before they have designed the surrounding organization. Gartner reported in April 2026 that only 13 percent of surveyed organizations believed they had the right agent governance in place. Its warning focused on proliferation, ownership, connectors, information risk, and lifecycle management. An AWS analysis of agent sprawl across business units describes what that proliferation can look like in practice: different units building duplicate capabilities, agents taking conflicting actions on shared systems, credentials multiplying, and costs remaining hidden inside separate budgets.
These are serious governance problems, but the organizational question underneath them is broader. What exactly are the agents being added to?
An agent may summarize the meeting, compare it with the proposal, retrieve the current policy, identify a contradiction, draft an update, and create the next task. All of this can be useful. It can also produce more material for the same human integration layer to verify if the company has not made its state, commitments, decision rights, and expected consequences explicit enough.
The person who previously carried the work may now receive more summaries to check, more recommendations to validate, more exceptions to resolve, and more actions performed at a speed that makes errors harder to catch. The organization may call them an approver, supervisor, or human in the loop, but the old dependency remains. It has simply been placed behind a faster production system.
This is why an AI employee is not the same thing as an integrated company. Recreating the sales function, finance function, product function, and management hierarchy with agents may produce a plausible software topology, but it does not tell us whether the underlying responsibility has been designed. Activity can be delegated before anyone has clarified which result the company is responsible for producing, who can legitimately commit to it, what evidence makes action warranted, how the result will be verified, or who can stop the system when the situation no longer matches its assumptions.
I will take those questions further in the next issue, Responsibility Before Automation, because they cannot be solved simply by naming an agent after a role.
The hidden cost is also a hidden form of power
The most obvious cost of human middleware is time, although time is only the beginning. The person carrying the organization’s context has to remain reachable across systems and teams because a signal from any one of them may change the whole picture. Their attention becomes fragmented, and they develop a form of anticipatory monitoring in which they remember what could fail before a dashboard has enough evidence to report it.
There is emotional work in this as well. A reminder has to be shaped differently for a customer, a peer, a manager, or a team already under pressure. An escalation is never merely routed; it is interpreted and presented in a way that makes action possible without unnecessarily damaging a relationship. The person may then carry accountability for an outcome governed by other people’s authority, incentives, and capacity.
When that person leaves, the documented work often remains while the operating capability disappears. The company discovers that what looked like an individual performance advantage was actually an undocumented dependency.
The same position creates quiet power. The person who prepares the summary decides which ambiguity becomes visible to leadership. The person who carries context decides which context travels and which remains local. The person who routes exceptions learns where the real authority sits, regardless of what the formal process says.
That power may be exercised responsibly, and often is. It can also become a bottleneck or a form of gatekeeping. This is one reason that simply removing the intermediary is not a serious solution. Mediation does not disappear when it moves into software. The system that assembles context also selects context. A model that scores opportunities influences which customers receive attention. A summarizer affects what leadership believes to be happening. A routing agent determines which exception reaches a person and which disappears into the ordinary path.
Moving the integration layer into software therefore redistributes visibility and power. It should be designed with the same seriousness as any other movement of organizational authority.
Begin with the person everyone calls
The most useful place to start is usually not the system architecture or the catalogue of possible agents. It is the person whom everyone calls when they need to know what is actually happening.
The first question should not be how to replace them. It should be what capability the organization is borrowing from them.
Watch what they have to reconstruct before they can act. Notice which systems, conversations, relationships, and remembered decisions they combine. Ask which commitments they preserve because no operational object owns them, which exceptions reach them because the formal process cannot cope with reality, and which outcomes they verify after everybody else believes the work is complete.
Some of what you find should become durable organizational memory. Some should become clearer state, a governed workflow, or an explicit decision right. Some will remain human because relationship, judgment, negotiation, care, or legitimate authority is part of the value. You may also discover that the person has been holding together a disagreement the company has avoided resolving. That conflict should not be quietly encoded into software as though a single official answer already existed.
The goal is not to extract everything the person knows and build a digital copy of them. Much of their knowledge will be contextual, relational, tacit, or legitimately private. The goal is to stop requiring one person to be everywhere for the company to remember what it promised, understand its current situation, and notice whether its actions produced the intended result.
Your best people may be holding the organization together. That does not mean holding it together is the best use of them.
Before adding another agent, look for the human integration layer already running inside the company. It will show you where the formal system ends, where context is repeatedly lost, and where responsibility has never been made explicit. It may also reveal that integration is not achieved merely when systems can exchange data. A company becomes more integrated when the state, commitments, authority, and consequences that matter can travel through the organization without having to be reconstructed from individual memory every time.
That is not yet an automation problem. It is an organizational design problem that automation is forcing us to see.
If your organization has one person whom everybody calls to learn what is really happening, I am interested in what that person is carrying that the company itself does not yet know how to hold.