The software we inherited
Enterprise resource planning systems became central to organisational life because organisations needed reliable ways to coordinate transactions, people, materials, finances and obligations. Over time, this produced an extensive landscape of forms, dashboards, workflow queues, approval chains and reporting surfaces. These interfaces are not superficial decorations. They have been how people interact with authoritative records and how institutions translate rules into daily operations. Yet it is worth distinguishing the durable functions from the particular human-operated forms through which those functions have been delivered.
That distinction becomes consequential when AI moves beyond assisting a person at a screen and begins performing bounded organisational work. An agent may gather information, compare records, initiate a transaction, test compliance conditions, reconcile an outcome and report exceptions. If those actions are legitimately delegated and technically reliable, reproducing every human click and handoff may be unnecessary. The question is not whether the organisation can dispense with enterprise software, but which parts of the software estate exist primarily to help humans operate it.
The disappearing interface is a hypothesis, not a forecast
Consider accounts payable. A conventional workflow may involve a person receiving an invoice, another validating a purchase order, a manager approving an exception and finance posting the transaction. An AI-native workflow could perform much of the routine matching and evidence gathering without the same human-facing sequence. But the ledger, payment authority, fraud controls, tax treatment, supplier identity, segregation of duties and exception resolution remain. The visible interface could shrink while the institutional requirements become more exacting.
The likely transformation is therefore uneven. A finance ledger has durable integrity requirements; an approval inbox may be more contingent on how work is organised. A human dashboard may become less central for routine execution but more important for investigating disputed actions or exercising legitimate oversight. Industries also differ. In healthcare, critical infrastructure, government and regulated finance, legal and human accountability constraints can limit what may be delegated even where a model performs well. There is no sound basis yet for asserting that a particular percentage of ERP functionality will disappear.
From application inventory to governed capability
Traditional enterprise architecture often begins with applications: what systems do we own, which vendor supplies them, who uses them and when must they be upgraded? That remains useful for asset management and cyber security. It becomes insufficient, however, if organisational work is increasingly delivered through replaceable combinations of models, agents, services, data and specialised software.
A capability-oriented approach starts with what the institution needs to accomplish, the information and authority required, and the conditions under which the outcome is acceptable. It can then evaluate possible implementations, including commercial suites, native model capabilities and open-source ecosystems. This does not justify a free-for-all in which agents silently download and operate arbitrary tools. The institution must still know its active dependencies, access rights, provenance, operational risks and substitution consequences. Dynamic discovery and disciplined runtime inventory are complements, not alternatives.
This direction is not unique to Silver Digital. SAP's published AI-native architecture describes enterprise systems evolving beyond isolated records toward context-rich, governed agents, and its 2026 Autonomous Enterprise announcement locates agents within business data and governance. These are important signals from an incumbent vendor, although vendor roadmaps are not proof that the resulting architectures will succeed at scale. The contested question is whether AI-native capability will mostly reinforce large integrated suites or allow a more modular and replaceable software estate.
Governance after the click
If an AI system proposes a payment, changes a supplier record or initiates a procurement action, conventional governance questions do not disappear. They move closer to the point of action. Who authorised this class of decision? What information was trusted? Which constraints were binding? What happened after the action? Could an independent party reconstruct and challenge it? Who remains answerable if the outcome was harmful?
This suggests a useful distinction between intelligence and authority. A system may be highly capable of identifying an optimal action without possessing legitimate permission to take it. Authority must be constituted and delegated; capability alone cannot create it. Governance of autonomous work therefore concerns legitimate purpose, scope of delegation, enforceable constraints, observability, proportional assurance, escalation and accountability. Some decisions can be machine-governable under evidence-backed limits. Others must retain deliberate human involvement because of law, rights, institutional legitimacy, uncertainty or irreversibility.
The NIST AI Risk Management Framework is relevant prior art: its govern, map, measure and manage functions treat risk management as continuous across the AI lifecycle. In applicable high-risk contexts, the EU AI Act also imposes oversight and logging obligations. Neither framework resolves every practical question about autonomous organisations, but both challenge the notion that governance can be reduced to approving a catalogue of AI products once and then leaving them alone.
Independent assurance and institutional memory
An AI-native institution cannot rely entirely on the agent's own explanation of its conduct. The system that performs a consequential action should not be the sole authority on whether that action was permitted, correctly executed or successful. Evidence may need to be captured from transaction records, independent controls, external observations and subsequent outcomes. Assurance should scale with consequence rather than demand identical human review for every routine action.
This is a demanding design problem. A claimed reduction in human approval work can be offset by increased costs of monitoring, testing, security, incident recovery and audit. More autonomy may require stronger independent checks, not fewer. The relevant measure is therefore not simply the number of screens removed or hours saved, but whether the institution can produce legitimate, reliable, reversible and accountable outcomes at a sustainable total cost.
The economics of the enterprise software estate
Enterprise software vendors sell more than interfaces. They provide domain rules, tested transaction processing, integration, support, security updates, compliance features and responsibility for maintaining complex products. These are real sources of value. At the same time, licence structures, implementation projects and consultancy arrangements can reward accumulated complexity and make substitution costly. It is reasonable to investigate those incentives; it is not reasonable to assume that every expensive enterprise system or consulting engagement exists because a vendor wants unnecessary work.
AI-native capability introduces competing economic possibilities. Incumbents may embed agents inside their existing suites, preserving trusted records and exploiting integration advantages. New providers may offer smaller specialised components that can be assembled around governed capabilities. Open-source software may improve replaceability and bargaining power but shift expenses toward engineering, support, security and assurance. A model that appears cheap in a demonstration may become costly when inference, exception handling and governance are counted. These alternatives should be compared on measured lifecycle outcomes rather than ideology about proprietary or open-source software.
A more demanding definition of governance
The emerging thesis is not that governance should disappear along with some human-oriented workflow. It is almost the opposite: as machine capabilities absorb routine organisational work, governance may become more central because action becomes faster, more frequent and less dependent on a person navigating an application. The institution must preserve what is legitimate and durable while questioning what is merely inherited.
That shift suggests a new evaluation question for enterprise technology. Instead of asking only whether a product has the necessary screens, workflows and AI features, ask whether the organisation can constitute authority, discover and replace capabilities, preserve trustworthy state, observe real consequences, assure decisions independently and intervene where legitimacy requires a human. None of this proves that ERP suites will vanish. It does suggest that the most valuable part of future enterprise software may be less visible to the routine user, even as it becomes more important to the institution.
The question remains open and empirical. How much human-facing software can disappear without degrading reliability or accountability? Where does modularity outperform integration? When is human intervention a necessary constitutional safeguard rather than a legacy workflow? Answering these questions requires comparative deployments and evidence over time. The enduring objective is not software reduction for its own sake, but reducing the distance between institutional intent and legitimate consequence without removing the friction that protects people.
PUBLIC CLASSIFICATION: GREEN
CANONICAL SOURCE: FTL-02-AI-NATIVE-ERP-GOVERNANCE · v1.0