You run three technology estates. Governing them as one is how costs escape the budget.

Writing in Forbes, Peter Bendor-Samuel argues that most enterprises are no longer managing one technology estate but three distinct environments, each needing its own operating model, governance and expectations ^[1]. That matters because the default corporate reflex is to extend the policies, organisational structures and contracting models that worked for traditional IT across all of it, an approach he says is unlikely to succeed ^[9].

The first estate is the accumulated stack built over the past three to five decades: mainframes, ERP, and the applications piled up through successive digital transformations, run internally, outsourced or in some hybrid arrangement ^[2]. Its operating principles are well understood ^[2]. The second is the AI-infused stack, where most enterprise money is going now, adding capability to existing systems through products such as Microsoft Copilot or custom applications built on existing data and workflows ^[3]. The third is native agentic AI, which according to Bendor-Samuel attracts the most excitement and the most misunderstanding ^[4]. In that environment data structures evolve continuously, ontologies change, the agents themselves adapt, and even the relationship between technology and business operations becomes fluid rather than fixed ^[4]. He describes it not as an evolution of the current stack but as a different one ^[4].

The financial logic for moving into the second estate is straightforward. Organisations leaned on labour arbitrage for decades, and AI now offers cost reduction beyond what moving work to cheaper locations can deliver ^[5]. The limits are equally clear: adding AI modules to existing applications generally produces incremental improvement rather than transformation, and bolting AI onto legacy systems rarely delivers the business reinvention executives anticipate ^[6]. Where organisations have finished the harder work of redesigning how teams operate, the reported range is 20% to 40% productivity improvement, and those gains come from operational transformation as much as from the software ^[7]. Buying licences without changing how work is performed adds cost without benefit ^[8].

Note what is and is not quantified here. The only numeric benefit range in the piece attaches to the middle estate, the AI-enhanced one; the agentic environment carries no cost or benefit figure at all ^[11]. So the estate generating the most enthusiasm is also the one with no published economics, and it is the one whose behaviour changes underneath the people meant to be supervising it ^[4][11].

That is where the governance point bites. In traditional systems, people and processes are the control layer, keeping technology performing correctly and inside acceptable cost boundaries ^[10]. In AI-native environments, Bendor-Samuel argues, much of that accountability has to be engineered into the technology itself ^[10]. He describes operational accountability as having two essential components, though the text available to us breaks off mid-sentence before the second is specified ^[12]. Read plainly, a control model that depended on a human noticing an anomaly does not transfer to a system that rewrites its own data structures between reviews ^[4][10].

Things worth watching: whether agentic workloads are being priced and contracted on traditional IT terms at the next renewal cycle ^[9]; whether anyone inside the business is auditing claimed productivity gains against the 20% to 40% band and the operating changes that supposedly produced them ^[7]; and whether run-cost accountability for agents is engineered into the systems or simply asserted in a policy document ^[10].