One Analyst Retracts The "AI Eats Software" Trade, And CIOs Should Read The Footnotes

Peter Bendor-Samuel used a Forbes column to walk back a forecast he made a year ago: that AI-driven task compression would meaningfully shrink the market for software development and technology services ^[9]. His revised position is that AI is splitting the enterprise estate into two markets rather than eliminating one ^[3], which matters to anyone building next year's vendor budget on the assumption that the line simply goes down.

The trade he is arguing against is familiar. Capital has flowed into chips and infrastructure while investors discounted software companies and technology services firms, on the assumption that AI will make development and implementation so efficient that both categories lose relevance ^[1][2].

His counter on the legacy side is unglamorous and hard to dismiss. Faster code generation does not remove the need to understand complex business processes, architect systems, or manage an enterprise technology estate ^[4]. Large enterprises are not going to replace decades of accumulated stack investment quickly, because the risk is high and the effort is routinely underestimated ^[5]. There is also a sequencing trap: before AI-native applications can replace existing systems, someone has to document every function those systems perform in enough detail to describe them, and building that understanding is itself an enormous undertaking ^[6]. His conclusion is evolution, not disappearance, with modest growth for years ^[7].

On services, he keeps the mechanism and drops the conclusion. Task compression is real: coding, testing and documentation get faster, so the revenue attached to an individual task falls ^[8]. But he now argues that demand from maintaining, extending, securing, governing and modernising enterprise systems is strong enough to offset much of that productivity gain, leaving the traditional services market relatively stable and possibly growing modestly ^[10].

The growth claim sits in the new layer, and it is the part operators should test hardest. A native agentic environment is not an extension of today's applications; agents, ontologies, orchestration layers, governance frameworks and supporting infrastructure change continuously, which he compares to redesigning a house while living in it ^[11]. That work needs continuous engineering, implementation, monitoring and refinement, plus tight coupling between business and technical teams, so on his reading native AI increases dependence on outside expertise rather than reducing it ^[12]. Taken with the legacy picture, the implication is a mix shift and a second spending line, not a contraction ^[1].

The operating-model point is where this becomes a procurement problem. Legacy maintenance and extension still fit delivery built on offshore talent, Global Capability Centers and established managed services contracts, and AI improves productivity inside those models without changing how they work ^[13]. He identifies the common error as treating both markets as if they run the same way ^[14]. A CIO who benchmarks agentic build-out against a legacy rate card, or governs it with the same statement-of-work cadence, will get the wrong price and the wrong accountability ^[2].

Two caveats. This is one analyst's directional argument, and the column offers no figures for either market's size or growth rate beyond words like modest and fast-growing ^[3]. It is also a reversal, which cuts both ways: it may reflect twelve months of evidence, or it may be the same forecast rebuilt around a more comfortable conclusion ^[9].

What to watch: whether services firms start reporting AI-native work as a separate line rather than folding it into digital revenue, and whether your own renewals show legacy managed services holding flat in absolute terms while a new agentic spend line appears beside it. If legacy contracts start repricing downward as agentic budgets grow, the bifurcation thesis is wrong and the discount was right.