Enterprise software incumbents are often described with one of two convenient conclusions. Their installed base is either an unbeatable advantage because they already hold the data and workflow, or an immovable burden because legacy systems slow every change.
The same asset can carry either sign.
An installed base gives an AI deployment known users, interfaces, identities, data, controls, budgets, and partners. It also inherits every unresolved question inside those systems: who may access what, which record is authoritative, how exceptions work, who approves an action, and who operates the result.
Advantage and deployment tax arrive together. Operational readiness determines which one the team encounters.
The advantage ledger
An incumbent platform rarely enters an empty environment. The customer already has accounts, relationships, data structures, administrative controls, support paths, and people who know how work moves. Distribution is present. Context exists. Governance mechanisms may already be approved.
Current vendor materials make the deployment surface visible. Microsoft's Agent 365 FastTrack documentation connects agent deployment with identity, least-privilege access, data protection, network controls, monitoring, lifecycle workflows, sponsors, and owners. Salesforce's consulting-partner material describes strategy, data mapping, multi-system integration, governance, deployment, change management, and optimization around Agentforce.
These are company descriptions, not independent performance evidence. They still show that an AI feature enters an operating estate. The provider connects new behavior to systems and controls that already carry enterprise responsibility.
That can reduce uncertainty. An existing identity system offers a known mechanism for authentication and access. A system of record may hold relevant customer or transaction context. An established partner may understand the integration surface. A familiar interface can reduce part of the adoption burden.
The deployment-tax ledger
Each advantage can reverse sign.
Take identity. Suppose an agent needs to read a customer record, propose a change, and submit that change for approval. Existing identity infrastructure helps if the enterprise has decided which principal the agent uses, who sponsors it, what it may read, what it may write, where human approval occurs, how activity is logged, and how access is revoked.
The same infrastructure becomes a constraint when the request is simply “give the agent access.” Security cannot approve an unspecified action boundary. Engineering cannot build a stable interface. Operations cannot monitor behavior against an undefined policy. The technology is present; the authority model is not.
Systems of record behave similarly. They accelerate work when fields are usable, meanings are shared, and exceptions are known. They slow it when context lives across incomplete fields, emails, local files, and human memory.
Governance is not the tax. Unusable governance is. Review becomes deployment infrastructure when the decision-maker, evidence requirement, escalation path, and turnaround expectation are explicit. It becomes a queue when every function can object and nobody can decide the next bounded step.
Five assets, two signs
Workflow context helps when the decision, user, exceptions, and baseline are known. It hinders when the team tries to automate a process it cannot describe.
Identity and access help when action boundaries, sponsors, privileges, review, and revocation are explicit. They hinder when access is requested before authority is defined.
Data helps when sources, meanings, quality limits, retention, and permitted use are understood. It hinders when volume is mistaken for usable context.
Integrations help when interfaces, responsible parties, failure behavior, and observability are established. They hinder when custom glue becomes invisible production infrastructure.
Governance and adoption help when decision rights, evidence, training, escalation, and operation are designed together. They hinder when governance is a late gate and adoption is left unassigned.
The point is not to repair every inherited system first. Sequence the readiness work around the selected workflow. Resolve the identities it needs, the records it trusts, the interfaces it crosses, the decisions it changes, and the controls that govern those decisions. This converts a broad modernization problem into a bounded deployment path while exposing any dependency that truly requires wider change.
What the evidence can show
Current research supports caution about translating AI availability into operating value. McKinsey's 2026 State of AI survey reports uneven enterprise scaling across AI tools. PwC's 2026 CEO survey reports that only a minority of surveyed CEOs saw both revenue gains and cost reductions from AI. The surveys use different populations and measures; neither establishes why a particular company succeeds.
An Accenture session published by AI Engineer offers a practitioner interpretation: enterprise agent programs face tensions across organizational speed, funding, delivery method, trust, and feedback, and can use graduated autonomy to build evidence. That remains a named speaker view, not a universal finding.
Together, the sources show uneven scale and a deployment surface broader than model capability. They do not establish that incumbents are universally faster, slower, safer, or more valuable than AI-native vendors. The paired ledger is my analysis of the operating mechanisms the public evidence exposes.
Review one workflow before selecting the model
This is not a call for enterprise-wide modernization before any AI work begins. Run a context-and-control readiness review on one consequential workflow:
- Who can decide how the workflow changes?
- Which systems and data are required, and who accepts each interface?
- What may the agent observe, recommend, or do, and who approves that authority?
- What evidence permits expansion, reshape, or stop?
- Who operates the result and accepts the handoff?
If the answers are weak, changing models will not resolve the operating gap. The next useful work is to name decision-makers, test interfaces, record controls, and bound the first release.
The installed base is not a verdict. It is context plus obligations. Make both usable in one bounded deployment, and existing assets can accelerate the work. Leave decisions implicit, and the same assets create delay, rework, and ownerless pilots.