Crossing a Department's Border Is What Makes Agentic AI Expensive

On October 2, MIT Technology Review published a piece arguing that enterprise AI has moved past the ambition stage into full operational deployment, with global AI investment set to reach $2.5 trillion in 2026, up 44 percent on the previous year. Buried in that piece is the more useful observation: all that capital is producing fragmentation. A sales agent runs well but has no idea a customer has an open support ticket. A marketing system personalizes outreach with no visibility into what finance already knows about the same account. Each function performs in isolation. The enterprise as a whole learns almost nothing.

That fragmentation is not a side effect of autonomous AI. It is the clearest evidence we have yet of where the real distinction lives between a system that answers questions and a system that does work.

The line is not capability, it is ownership

A system that answers questions can live happily inside a single team's data, because the human asking the question supplies whatever context is missing. A support agent chatbot that only sees support tickets is still useful. It answers the question in front of it. A system that is supposed to do work, by contrast, has to act on a decision that spans teams, and that means it needs access to data it does not own: the support ticket the sales bot never saw, the margin figure the marketing system was never shown. The moment an agent needs to read or write across that boundary, the project stops being a model problem and becomes an organizational one.

This is visible in the easier cases too. MTN Rwanda's CEO recently credited AI with improved fault management on the network, a result reported by Telecompaper. Fault management is a contained domain. Network operations owns the sensors, the logs, the escalation rules, and the agent that watches them. Nobody outside that team has to sign off on what the system is allowed to see or touch. That is precisely why it works cleanly: the agent never crosses a border it doesn't own.

Compare that to the scenario MIT Technology Review describes, where a system is supposed to coordinate sales, support, and finance. Each of those functions has its own owner, its own data quality standards, its own incentive to resist an outside agent writing into its system without review. The coordination the article calls for is not a data pipeline problem. It is three separate negotiations with three separate department heads, each of whom has a legitimate reason to say no first and ask questions second.

The objection: this is just integration work

The obvious reply is that this is nothing new, that enterprises have been solving cross-system integration for decades with APIs, data warehouses, and master data management, and that calling it an organizational problem is just a way of avoiding the engineering. There's something to that. Building a software bridge between a CRM and a ticketing system is a known problem with known solutions, and tools like South Africa's Exten AI, which lets non-technical founders generate full-stack applications from plain language, show how cheap that layer of software is becoming. If generating the connective code is nearly free, the argument goes, the organizational friction should dissolve along with it.

It doesn't, because the friction was never really about the code. An API connecting two systems is a technical artifact. Permission for an autonomous agent to act on another team's behalf inside that system is a trust artifact, and trust artifacts are not generated by a prompt. The support team's objection to a sales agent auto-closing or re-prioritizing tickets is not that the integration is hard to build. It's that they did not choose this, were not consulted on the failure modes, and will be the ones blamed when the agent gets a judgment call wrong on a customer they've been managing for two years. Cheaper code generation makes the technical half of the problem smaller. It does nothing to the half that requires a finance director, a support lead, and a sales lead to agree on what an agent is allowed to decide without a human in the loop, and what it must escalate.

What this costs in practice

For an organization operating in East Africa, this boundary-crossing cost is not smaller than elsewhere. If anything it is larger, because the teams on either side of the border often run on different systems with different data discipline, built at different times under different procurement decisions, and the people who own those systems have watched automation projects arrive before and change their job description without warning. A hotel group that wants one agent coordinating reservations, front desk, and billing is not buying a software feature. It is asking three departments that have never shared a system of record to agree on what an outside agent gets to see about their work and what it gets to decide. Uganda's stated push toward AI-driven tourism growth, reported by TV BRICS, will run into exactly this seam the moment anyone tries to connect a booking agent to the finance systems that actually reconcile revenue.

The practical answer is to stop pricing agentic projects as if the model and the integration layer are the expensive part. They are the part that has gotten cheap, fast, and will keep getting cheaper. The part that determines whether the project finishes on schedule is how many department owners the agent has to cross, and whether each of them has actually agreed to the crossing before the build starts, not after the demo.

When a vendor describes a system as autonomous, the question worth asking is not what model it runs or how many tokens it processes. It is whose data the agent touches outside the team that commissioned it, and who on the other side of that boundary signed off on letting it in. If nobody can answer that question, the system has not been designed to do work. It has been designed to answer questions inside one room, and someone is about to discover that the hard way, probably in month three.

Sources

Redefining enterprise intelligence with autonomous AI (MIT Technology Review, 2026-10-02)

South African startup Exten AI helps non-technical founders build full-stack apps in plain language (Disrupt Africa, 2026-10-01)