a16z partners Eric and Seema Amble published an essay this month arguing that legacy enterprise platforms — SAP, ServiceNow, and Salesforce — are not going away, but that AI agents will become the dominant interface layer sitting above them. The central thesis frames these systems as irreplaceable repositories of institutional memory: canonical data models, compliance controls, approval workflows, and custom pricing logic accumulated over decades. Migrating off them is prohibitively expensive — the essay cites an SAP ECC to S/4HANA upgrade costing $700 million, taking three years, and requiring a 50-person Accenture team. Rather than replacing these systems, a16z argues AI agents will make them dramatically more programmable, turning messy human intent into correct, auditable action across three lifecycle phases: implementation acceleration, day-to-day workflow automation, and bespoke extension building.
The target opportunity is the $380 billion system-integration market. The essay profiles a cohort of early-stage startups attacking each phase. On the implementation side, Axiamatic builds knowledge graphs from project artifacts to de-risk S/4HANA migrations, while Tessera evaluates existing ERP instances and automatically remediates issues mid-migration — both aimed at compressing timelines and cutting reliance on large consulting teams. For daily operations, the essay points to computer-use agents capable of <a href="/news/2026-03-14-markov-ai-releases-48k-screen-recording-dataset-for-training-computer-use-agents">operating at the UI level</a>. The motivation: the essay cites data showing digital workers toggle between application contexts roughly 1,200 times per day, losing around four hours of productivity per week. Other profiled startups include Conduct, Auctor, Supersonik, Factor Labs, Sola, and General Magic Cell; a16z discloses investment positions in several.
The strategic bet is that winning companies here will not pitch clean-slate replacements. They enter on transformation budgets, demonstrate measurable risk reduction, then expand into daily operations as a trusted "system of action" control plane. On Hacker News, that framing ran into a pointed objection from experienced SAP users: the dense UI is not a problem most power users want solved. The comparison to Bloomberg Terminal users is apt — people who know the interface well prefer it precisely because it is optimized for speed, not for newcomers. What AI pitches often miss, commenters noted, is the depth of edge-case coverage SAP provides: multi-shipment jobs with split due dates, complex exception handling built over years of real operational use. That institutional specificity is hard to replicate and easy to underestimate.
The 70% failure rate for large-scale ERP transformations, cited in the essay, is the sharpest argument for the agent wedge. If an agent can demonstrably reduce that failure rate on a $700 million project, the ROI case sells itself before anyone has to make a broader argument about AI. The real test for startups like Axiamatic and Tessera is whether they can land and hold on those transformation engagements long enough to prove it.