Your company just approved a major AI investment. Most of it is about to be spent making a broken architecture more comfortable to live with.
Every enterprise in the world is racing to deploy AI. Hundreds of billions in investment. Agentic systems that can reason, see, read, and act. The most powerful general-purpose technology since the internet.
And most of it is being pointed at the wrong target.
Somewhere in your organisation right now, a person is sitting at a desk with two application windows open. They are reading data from one screen and typing it into another. The two systems they are bridging were built by different vendors, run on different technology stacks, and have never been taught to speak to each other.
Somewhere else, another person is looking at a PDF on their screen. They spot a particular value and already know what will happen if they enter it into the target system. It will trigger an exception. So before they type anything, they split the value across multiple entries, or route the record to a different application entirely, or apply a conversion that no system has been programmed to perform. The rules live nowhere except in their head.
The first person is bridging a gap between systems. The second person is the system — performing the function of an application that was never built.
Both are doing what I call software-driven labour: the manual operation of software to compensate for gaps in the enterprise’s application architecture. The global economy spends upward of $400 billion a year paying people to do exactly this.
Here is the question your AI strategy should be answering: are you using AI to eliminate the need for this labour — or are you using AI to perform it faster?
Because those are two fundamentally different things. And nearly every enterprise AI deployment I see is doing the latter.
The Long Tail
Every enterprise runs on an archipelago of software systems — dozens or hundreds of applications that work well individually but were never taught to communicate. When a business process needs multiple systems to interoperate and even one connection is missing, the default response has been the same for thirty years: put a human in the gap.
If you rank every point where two systems need to interoperate by business value, you get a long-tail distribution. The high-value integrations get built. But the long tail — hundreds of smaller integrations, each individually modest but collectively enormous — gets left to humans. The cost of conventional integration is too high for each individual case. So the organisation hires people in lower-cost geographies to bridge the gaps manually.
Here is what the cost-per-transaction logic misses: the moment you place a human in the process chain, you lose straight-through processing. This is not a minor efficiency trade-off. It is a categorical difference in operational capability — seconds versus hours, zero error versus re-keying mistakes, unlimited scalability versus shift ceilings. You are not comparing two ways to do the same thing at different price points. You are comparing two fundamentally different architectures.
Now AI has arrived with the potential to change these economics entirely. An intelligent agent that can interact with any system, through any interface, adapting in real time — this should make the long tail viable for the first time. But only if the AI is pointed at the integration problem. If instead it’s pointed at the human in the gap, you’ve built a faster workaround. The ceiling remains.
The Missing App
The conventional understanding is that this is purely an integration failure: systems that can’t talk to each other. That understanding is incomplete — and it explains why current AI deployments are falling short.
Go back to the second person I described at the opening. They are not bridging two systems. They are executing business logic — rules that determine how data should be transformed, when exceptions should be raised, which records should be routed where. These rules are real. They govern real business outcomes. And they have never been written into software.
This is the missing application in your enterprise architecture.
Consider what these operators actually do:
They look at a row in a spreadsheet and determine that a certain type of explanatory note must be prepared — based on a pattern they’ve learned to recognise, applying rules that exist in no system. They apply exchange rate conversions or threshold adjustments that should be codified but never were. They spot a field value and already know it will trigger an exception workflow — so they pre-emptively restructure the data before entering it.
These people are not performing data entry. They are performing application logic — the function of a software system that was never built. The rules they carry in their heads are its business rules. The exception-handling they perform is its error-handling logic. The pattern recognition they apply is its decision engine.
Software-driven labour has two dimensions: integration labour — moving data between systems that cannot interoperate — and logic labour — executing business rules that have never been codified into software. One is a missing connection. The other is a missing application.
This is precisely where AI should be revolutionary. A multi-modal agentic system can observe the logic that operators perform, understand the rules they’re applying, and codify those rules into a functioning application — not by mimicking the human, but by building the software that should have existed all along.
The BPO industry bundles both integration and logic labour under a single headcount and calls it “operations.” AI vendors are now bundling both under “intelligent automation” and calling it progress. In both cases, the actual problem — the incomplete architecture — remains untouched.
Three Waves, Same Target
I’ve written extensively about the Box A / Box B framework — the distinction between the core enterprise layer where data processing should happen (Box A) and the workaround layer that exists because Box A is incomplete (Box B). (For the full argument about where AI is currently being aimed within this framework, see Rethinking AI for Automation: Where the Light Is.)
Here is the pattern that has repeated three times:
In the 2000s, the BPO industry poured investment into making Box B’s human workforce more efficient — better training, lower-cost geographies, optimised shift patterns. The gap in Box A remained.
In the 2010s, the RPA movement poured billions into giving Box B digital workers — bots that mimic human actions on screens. The gap in Box A remained.
In the mid-2020s, the AI industry is pouring hundreds of billions into giving Box B intelligent agents — systems that can read documents, extract data, and operate interfaces with superhuman speed. The gap in Box A remains.
Three waves of technology. Each more powerful than the last. All three pointed at the same wrong target.
The BPO industry will tell you that the easy work has already been automated and what remains requires human judgment. This deserves scrutiny. What they automated was the simplest screen-to-screen data carrying. What remains is the logic labour — and it resists their tools not because it is inherently beyond automation, but because neither RPA nor current AI deployments are designed to be applications. They are designed to mimic what humans do in the gap. That is not evidence the work requires humans. It is evidence the technology is aimed at the wrong layer.
We are building AI agents to read documents that should never have been documents, extract data from PDFs that were generated from digital data in the first place, and automate workarounds that exist only because the underlying architecture was never completed. This is the most expensive misdirection in enterprise technology today.
The Onus to Comply
There is an architectural reason why integration has remained hard across every technology generation. I call it the onus to comply: in conventional integration, every participating application must include code to communicate with the integration infrastructure. You’re not just building the bridge — you’re retrofitting every island to accept the connectors. When even one application can’t comply, the chain breaks.
CORBA, SOA, ESBs, APIs — each generation imposed this onus on individual applications. Each time, the onus defeated the long tail.
The technology that became known as RPA — which a few of us built in the early 2000s as a UI integration tool — solved this differently by shifting the onus from the applications to the integration layer itself.
Agentic AI takes this further. An intelligent agent can interact with any system through whatever interface it presents — UI, API, document, message queue — adapting dynamically. This doesn’t eliminate the onus to comply. It shifts it to a far more capable layer.
But this capability only changes the game if the agent operates in Box A, as part of the integration architecture. An AI agent in Box B, reading screens on behalf of a human operator, is a self-driving car being used to run errands that wouldn’t be necessary if the house were in the right location.
The Seduction
If the problem is so clear, why does it keep growing?
Because outsourcing is the most seductive decision in enterprise management — and AI is making it more seductive.
A process costing $100,000 a year can be moved offshore for $30,000–$40,000 within two to three months. The CFO sees a 60% cost reduction. The CEO tells the board operations have been streamlined. The savings appear in the next quarterly report. Everyone is rewarded within the timeframe they are measured on.
Now add AI. The outsourcing provider offers AI-augmented operations — intelligent document processing, agentic bots handling exceptions. The cost drops further. The pitch is irresistible: labour arbitrage and AI capabilities, without building anything yourself.
The architectural fix offers none of this short-term gratification. It requires upfront investment, takes six to eighteen months, and the benefits compound over years, not quarters. No one gets promoted for preventing a structural dependency that would have cost ten times more over the next decade.
Outsourcing is operational debt. The $60,000 annual saving per role is the yield you think you’re earning. But the principal — competence drain, knowledge extraction, architectural lock-in, permanent forfeiture of straight-through processing — compounds silently. By the time the true cost surfaces, unwinding the dependency costs more than the original problem.
Every CFO understands financial debt. Very few recognise that their outsourcing portfolio carries the same compounding structure — with no balance sheet entry to make it visible.
The Real Cost
When an organisation outsources its process operations, it begins a slow transfer of institutional knowledge. The operators who handle exceptions, apply undocumented rules, and navigate architectural gaps accumulate a deep understanding of how the business actually works — not how it’s supposed to work, but how it actually works.
Over years, this knowledge migrates to the outsourcer. The client hollows out.
And then the BPO providers, having accumulated this knowledge across dozens of clients, begin to codify it. They build software platforms. They productise the knowledge their clients paid them to accumulate — and sell it back as software. With AI, this extraction accelerates. The outsourcer trains models on patterns learned across hundreds of engagements. The client funded the training data. The outsourcer owns the model.
Here is the test no one applies: has your organisation ever successfully brought a significant outsourced operation back in-house? At what cost? Over what timeframe?
Outsourcing takes two to three months. Insourcing takes two to three years — and often fails. The ease of entry was the trap. The cost of exit is the true price.
The cheap workaround is the most expensive decision the enterprise ever made. AI-augmented outsourcing is the same decision — at greater velocity and greater lock-in.
What Needs to Change
The path forward is not more AI in Box B. It is the completion of Box A — using AI as the means to achieve what was previously too expensive, too slow, or too complex.
First, digital data exchange must become the norm. We are building sophisticated AI to solve a problem created by sending digital data through analogue channels. Self-service portals, B2B gateways, and standardised interchange formats exist. AI should make digital exchange easier to set up, not make analogue workarounds more tolerable.
Second, the missing applications must be built. The business logic in operators’ heads can now be captured, reasoned about, and codified by agentic AI. Multi-modal models that process documents, data, screen states, and contextual rules simultaneously can build the application that should have existed all along — and place it in Box A. Not reading PDFs — writing the software that eliminates the need for PDFs.
Third, the integration layer must become intelligent. When the onus to comply shifts to an AI-capable layer that negotiates connections dynamically, the long tail becomes viable for the first time. The barrier that defeated every previous integration paradigm is dissolving.
I’ve spent nearly three decades inside this problem. I built a UI integration product in the early 2000s, before the term RPA existed. I watched it get acquired, repackaged, and redirected into the workaround layer it was designed to eliminate.
And now I am watching it happen a third time — the most capable technology we have ever had, deployed at extraordinary scale to build a more sophisticated version of the same workaround.
The $400 billion is not inevitable. The technology to fix it exists. With agentic AI, the ability to build the missing integrations, the missing applications, and achieve straight-through processing has never been greater.
We have the technology. We have the intelligence. We have the investment.
We just need to point it at the right target.
