It has been about a month since the last article, which feels both longer and shorter than it should. I may take the scenic route on this one, but it is my website and apparently I built it for exactly this sort of thing. Glad you are here. This is the good stuff.
July at Microsoft has a particular texture. The fiscal year turns over, everyone tries to remember which rhythm they are supposed to be in, and half the company seems to be somewhere with an out-of-office message doing the emotional labor of pretending not to check Teams.
I was one of them.
I have been off, taking time with family, and admittedly turning work mode off as much as I know how to do. Present, reset, not trying to turn every stray thought into a diagram. Mostly successful. Somewhat successful. Successful enough.
Coming back into the conversation, though, one pattern was hard to miss.
The enterprise AI conversation continues to evolve.
Over the past sixty days, the most useful conversations I have been in have not been about prompt output. They have not been about whether one model writes a cleaner summary than another, or whether the newest release feels a little more natural, or whether the benchmark chart moved in the right direction again.
Those things still matter.
They are also not where the serious enterprise buyers are spending most of their energy anymore.
The conversation has moved toward integration. How does the AI reach the systems where the work lives? How does it understand permissions? How does it act inside a workflow without creating a control problem? How does it connect to the business process cleanly enough that someone can measure the result against the investment?
That last part matters.
The real ROI is not in whether a model can produce a better paragraph. The real ROI is in whether it can reduce time in a process, improve the quality of a decision, close a loop that used to require five handoffs, or move work across systems without adding another layer of human coordination.
That has become a recurring theme.
Most AI discussions still assume adoption is constrained by intelligence. If the models get smarter, the enterprise use cases will arrive. If the reasoning gets better, the workflows will follow. If the benchmark scores keep climbing, organizations will eventually stop experimenting and start transforming.
That is true at the frontier.
It is less true inside the average enterprise.
Most large organizations are not stuck because the model cannot write a summary, draft a response, classify a record, reason through a policy, or turn messy inputs into a useful first pass. For a large class of enterprise work, the model has crossed the threshold of good enough. Not perfect. Not magical. Not something you leave unsupervised in a regulated workflow with a corporate credit card and a dream.
Good enough.
The constraint has moved.
The bottleneck is no longer whether the model can think. The bottleneck is whether it can reach the systems where the work actually happens, understand what it is allowed to see, act inside the right workflow, and leave behind an audit trail the business can defend.
That is the shift from the intelligence economy to the connector economy.
The industry is solving the visible problem
The model race is easy to watch.
A new model launches. It reasons a little better. It handles longer context. It writes more naturally. It passes a harder exam, codes a cleaner function, follows instructions with fewer strange little detours. The demos are visible and the comparison is clean. Anyone can open two tabs and feel the difference.
Enterprise AI does not work that cleanly.
Inside a large organization, intelligence is only one part of the work. The model has to find the right document, not just any document. It has to respect the permission boundary, not just answer the question. It has to know which record is trusted, which workflow is current, which system of record wins when three systems disagree, and which action requires a human approval before the machine touches anything.
That is not a model problem.
That is an access problem.
The work lives in CRM systems, ERP systems, HR systems, ticketing platforms, procurement tools, document repositories, meeting transcripts, chat history, custom applications, old line-of-business systems, and the internal workflows people politely call process because calling them institutional duct tape feels rude.
A generic model can sound intelligent while being functionally useless. It can produce a polished answer to the wrong question. It can summarize a policy it should not have been able to open. It can propose a workflow that makes sense in the abstract and breaks the minute it touches the actual business.
Enterprise AI is not bottlenecked by intelligence. It is bottlenecked by access.
That is the part of the story I think we keep underpricing.
The connector tax
Most organizations still price AI by looking at model consumption.
Tokens. Seats. Compute. Inference. Maybe some vector storage. Maybe a premium license if procurement is feeling brave and finance has not yet asked what the word agent means in a purchase order.
Those costs matter.
They are also not the whole cost.
The real expense starts when the business asks the AI to do something useful.
A simple request like “prepare the account brief” may require document retrieval, email context, meeting history, CRM records, entitlement checks, security trimming, data classification, summarization, citation, drafting, and routing. A request like “process this exception” may require identity validation, policy lookup, record updates, workflow execution, audit logging, notification, escalation, and possibly a human approval before the final write.
What looked like one AI interaction becomes a chain of backend calls.
That chain has economics.
Every retrieval call costs something. Every API call costs something. Every transformation costs something. Every permission check costs something. Every connector has to be built, tested, secured, monitored, updated, and owned. The model may be the part everyone can see, but the connective tissue is the part that quietly accumulates cost.
At scale, enterprises may spend more moving context than generating intelligence.
That is the connector tax.
It shows up in technical debt first. The enterprise has spent twenty years accumulating systems with different identity models, different permission shapes, different data quality, different integration patterns, different owners, and different definitions of truth. AI does not erase that complexity. It exposes it.
The model becomes a flashlight pointed directly at the mess.
A company that cannot answer which system holds the authoritative customer record will not magically answer it because a model is now asking politely. A company with inconsistent permissions will not become more secure because the assistant has a nicer interface. A company with brittle integrations will not become agentic. It will become brittle faster.
The hidden cost of AI is not inference. It is integration.
The scarce resource changed
Technology markets tend to misidentify scarcity at the beginning of a cycle.
In the early cloud era, the visible scarce resource was compute. Over time, the more durable value moved into the operating model around compute: identity, governance, deployment, observability, security, data services, procurement, and the control plane enterprises could actually run.
AI is moving through a similar curve, only faster.
The early scarce resource was intelligence. Who had the best model. Who had the best benchmark. Who had the largest context window. Who could reason, code, summarize, plan, and recover from ambiguity.
That still matters at the frontier. There are problems where a few points of model quality change the outcome. Scientific discovery. Deep technical reasoning. Novel design. Hard research. The frontier is real.
Most enterprise work is not frontier work.
Most enterprise work is summarizing messy information, grounding answers in company context, routing requests, drafting from known patterns, checking policy, preparing decisions, updating records, and moving work through a system without creating a mess behind it.
For that work, the scarce resource is different.
The scarce resource is access to the right context, under the right permission, inside the right workflow, with the right ability to act.
That is the connector economy.
The assets that matter start to change.
Model performance still matters, but data access matters more. Token efficiency still matters, but permission efficiency becomes just as important. Compute still matters, but workflow connectivity becomes the gating condition. Reasoning quality still matters, but context availability determines whether the reasoning is useful.
The question changes from “which model is smartest” to “which platform can connect intelligence to work with the least cost, least risk, and least operational drag.”
That is a very different market.
Connectors are not plumbing anymore
It is tempting to treat connectors as boring infrastructure.
That is usually a good instinct. Boring infrastructure is how enterprise technology becomes real. The exciting parts get the keynote. The boring parts get the budget.
But connectors are not just pipes in this cycle.
They are economic assets.
A connector determines what a model can see. It determines what an agent can do. It determines how permissions are interpreted, how actions are executed, how logs are created, how errors are handled, and how much friction exists between a user’s intent and a business outcome.
The better the connector layer, the cheaper it becomes to turn intelligence into action.
The worse the connector layer, the more every AI use case becomes a bespoke integration project wearing a nicer user interface.
That distinction matters because agents amplify the problem.
A chatbot connected to a document repository is one thing. An agent that can execute work across finance, HR, IT, sales, procurement, and operations is another. Now every workflow becomes a chain of systems. Every action has dependencies. Every dependency has a permission model. Every permission model has an owner. Every owner has a backlog.
An enterprise that deploys 200 agents across 150 integrations has not simply deployed AI.
It has created a new software estate.
That estate needs observability. It needs lifecycle management. It needs security review. It needs deprecation planning. It needs cost telemetry. It needs incident response. It needs a way to understand when the agent failed because the model reasoned badly, the connector broke, the API changed, the permission expired, the source data was wrong, or the workflow itself was never as clean as the process document claimed.
The hidden management burden is enormous.
It is also where the next market will form.
The graph beats the pile of APIs
The obvious answer to the connector economy is more connectors.
Build more integrations. Support more SaaS apps. Expose more APIs. Create more tool catalogs. Let agents call everything.
That will be necessary for a while.
It is not the end state.
The better architecture is not a model connected to twenty systems through twenty connectors, each with its own permissions, logs, schemas, failures, and governance posture. That works, but it is expensive. It also makes every AI workflow inherit the complexity of every system it touches.
The better architecture is a model connected to a unified enterprise graph.
Not a graph in the vague marketing sense. A real graph of people, teams, files, meetings, messages, business records, devices, workflows, permissions, policies, classifications, and relationships. A layer that does not simply store data, but understands how the organization fits together.
That graph becomes powerful for a simple reason.
Enterprise context is relational.
The value is not just in the document. It is who created it, who edited it, which meeting led to it, which account it relates to, which policy governs it, which team owns it, which permission boundary protects it, and which workflow depends on it.
A pile of APIs can retrieve data.
A graph can explain why the data matters.
That is the operating system shape for enterprise AI. The model does the reasoning. The graph supplies the context, permission inheritance, organizational meaning, and paths to action.
The optimal platform does not maximize connectors forever. It minimizes the need for external connectors by making more of the enterprise natively understandable to the AI layer.
That is where hyperscalers and large enterprise platforms have an advantage that is easy to underestimate.
They already sit across identity, security, productivity, infrastructure, data, workflow, and admin surfaces. They already know where people work, where files live, how permissions are assigned, which devices are managed, which apps are trusted, and which controls the enterprise has put in place.
An outside AI tool can integrate into that world.
A platform that already owns large parts of that world can inherit it.
That difference is economic.
Why hyperscalers are positioned better than they look
There is a contrarian version of the AI market that says model-first companies will capture the value because they define the intelligence layer.
They may capture a lot of value.
But the enterprise value pool may not settle where the early attention went.
The hyperscalers and major enterprise platforms have a different advantage. Potentially less elegant in a demo, certainly more durable in production.
They host the infrastructure. They manage the identities. They secure the devices. They store the data. They run the databases. They provide the productivity surfaces. They manage the admin controls. They sit inside procurement. They already have the compliance posture and the enterprise trust muscle.
That matters because every external connector has a cost.
It adds latency. It adds a security review. It adds maintenance. It adds another vendor surface. It adds another place where permissions can be misread, logs can fragment, and data can move in ways the governance team has to chase later.
Internal graph calls are different.
They can be cheaper. They can be more observable. They can inherit existing permissions. They can respect classification. They can feed existing audit trails. They can sit inside the same admin model the organization already uses.
This does not mean one ecosystem will own everything. Enterprises are too messy for that, and anyone who has seen a real application estate knows better.
But it does mean consolidation has a gravitational pull.
If a platform can combine identity, security, context, workflow, AI experience, and governance into one operating layer, it reduces the connector tax. It gives the enterprise fewer seams to manage. It turns AI from a set of clever tools into an extension of the control plane the organization already trusts.
That is why the cloud giants may benefit from AI in a way the market still describes too narrowly.
Not just by selling compute.
By turning enterprise context into an economic moat.
The governance requirement gets heavier
The connector economy also changes the governance problem.
When AI mostly answered questions, governance meant controlling what it could read and how it responded. Important, but bounded.
When AI starts doing work, governance has to cover action.
What can the agent see. What can it change. Which systems can it call. Which records require approval. Which actions are reversible. Which are not. What happens when the agent is uncertain. Who owns the exception. Who reviews the logs. Who pays for the workflow. Who shuts it off when it stops being useful.
These are not theoretical concerns.
They are the practical questions that determine whether a proof of concept survives contact with security, legal, finance, and operations.
Enterprises will need a more mature discipline around connector governance. Not just whether a connector exists, but whether it is approved for certain data classes, whether it supports security trimming, whether it logs at the right level, whether it handles permission changes correctly, whether it fails safely, and whether someone owns it when the upstream API changes on a Thursday night because software vendors enjoy keeping everyone humble.
This is where technical debt becomes governance debt.
A poorly documented integration was already a risk. Add an agent that can reason across it and act through it, and the risk becomes larger. Not because the model is malicious. Because it is capable enough to expose the weak points in the operating model.
The organizations that do well here will not be the ones with the most agents.
They will be the ones with the cleanest boundaries.
The maintenance burden nobody wants to price
Every connector has a lifecycle.
It gets built. It gets tested. It gets approved. It gets deployed. It gets monitored. It breaks. It gets patched. It gets versioned. It gets deprecated. It gets replaced. Somewhere in the middle, the person who understood it best leaves the company and everyone discovers the documentation was more aspirational than descriptive.
Now put agents on top of that.
The maintenance burden compounds because the connector is no longer just serving a reporting dashboard or a simple automation. It is feeding an intelligent system that may make decisions, propose actions, execute steps, and interact with humans.
A broken connector does not just return an error.
It may produce stale context. It may omit a key record. It may cause the agent to route work incorrectly. It may make the AI appear less capable when the real issue is that the system behind it was never ready to be depended on at this level.
This is the part of the AI cost model that most early business cases miss.
They calculate licenses. They estimate model usage. They forecast productivity gains. They rarely price the ongoing cost of keeping the connective tissue healthy.
That is going to change.
FinOps will expand from compute and cloud spend into AI workflow economics. Security teams will expand from human and non-human identities into agent action governance. Platform teams will be asked to maintain connectors as first-class production assets, not side projects. Business owners will be asked to justify not just whether an agent works, but whether the workflow it depends on is worth maintaining.
The connector economy is not just a technical market.
It is an operating model tax.
The market implications
A few things follow if this thesis is right.
First, model differentiation compresses for a large share of enterprise use cases. Not disappears. Compresses. Buyers will still care which model is underneath, especially for hard reasoning and high-risk workflows. But the buying question will become more practical. Can this AI reach our context, respect our controls, operate inside our systems, and be managed without creating another shadow estate?
Second, connector marketplaces become more strategic. The quality, security, and governance posture of connectors will matter more than the raw count. A vendor boasting thousands of integrations may be less valuable than one with fewer integrations that actually understand enterprise permissions, logging, data classification, and workflow execution.
Third, platforms with native enterprise graphs gain pricing power. The more context and action can happen inside a governed graph, the less the enterprise pays in external orchestration cost, integration drag, and operational complexity. That advantage may not show up in a benchmark. It will show up in deployment velocity and total cost.
The winners will not just make AI smarter.
They will make it governable enough to use.
The honest limits
Worth being clear about what this is not.
This is not an argument that models no longer matter. They do. Better reasoning expands the set of workflows AI can handle. Lower cost expands adoption. Longer context changes design patterns. Reliability matters enormously when the system is being asked to act.
This is also not an argument that every enterprise should consolidate into one ecosystem and call it strategy. Real enterprises are heterogeneous. They always have been. The average large organization has too many systems, too many acquisitions, too many regional requirements, and too many specialized workloads for any single vendor fantasy to survive inspection.
The point is narrower and more useful.
As models become sufficiently capable for common enterprise tasks, the value shifts toward the layer that connects intelligence to reality. That layer includes identity, permissions, enterprise context, workflow execution, observability, governance, and cost management.
Some of that will live inside hyperscaler ecosystems. Some of it will live in specialist tools. Some of it will live in vertical platforms. Some of it will remain ugly custom integration because the enterprise is a museum of business decisions with an API gateway attached.
But the direction is clear.
The company that reduces the cost of connection will capture more value than the company that only improves the answer.
What this looks like a few quarters out
If I had to make a few small bets, they would be these.
The immediate wave of enterprise AI disappointment will not come from models being too dumb. It will come from models being smart enough to reveal that the organization’s systems are not ready to be acted on.
The following wave of buying will move from AI features to AI operating platforms. Buyers will ask less about the demo and more about identity, permission inheritance, auditability, workflow ownership, connector health, and cost per completed action. Many organizations are here now. This is one of the conversations I keep running into with early adopters. They are beyond concept. They are building, and having to answer real questions from security and finance.
For the most AI-mature organizations, these are not clean sequential waves. They are branches of the same decision tree.
One branch is graph advantage. Platforms that already understand the relationships between people, files, meetings, messages, devices, applications, policies, and workflows will make AI feel more useful with less integration effort. That will look like magic to users. It will look like architecture to everyone who has to run it.
Another branch is governance. Mature buyers are already treating governance as a product requirement, not a compliance afterthought. If an AI system cannot show what it accessed, why it accessed it, what it changed, whose permission it used, and who approved the action, it will stay trapped in low-risk work.
The third branch is economic. Enterprises are beginning to calculate the cost per AI-mediated workflow, not just the cost per token or license. That is when the connector tax becomes visible on a spreadsheet, which is usually the moment a hidden problem becomes a real market.
The closing read
The AI era will not be won only by the company with the smartest model.
It will be won by the company that most efficiently connects intelligence to enterprise reality.
That means access. Permissions. Context. Workflow. Action. Governance. Maintenance. All the unglamorous machinery that turns a capable answer into a completed piece of work.
Models are becoming more abundant.
Context is not.
Reasoning is getting cheaper.
Connection is not.
The connector economy emerges because every useful enterprise AI system eventually has to reach information, understand authority, and do something inside the business without making security, finance, legal, and operations regret the whole project.
The next enterprise AI advantage will not come from knowing more in the abstract.
It will come from knowing what matters here, reaching what is allowed here, and acting in a way the business can trust here.
That is the work now.
Not smarter answers.
A shorter bridge between intelligence and action.
A bridge over the moat.
Views expressed are explicitly that of my own.