For more than two decades, the enterprise software stack followed a familiar structure. Companies bought applications for finance, human resources, customer management, supply chains, and service operations. Employees logged into those applications, moved through menus and dashboards, entered information, and pushed work from one system to another.
Artificial intelligence (AI) is beginning to unsettle that model.
The change is not simply that business applications are gaining chatbots or better search functions. AI agents are starting to interpret requests, retrieve context, make recommendations, trigger actions, and move work across systems. In some cases, the software is no longer waiting for a human user to operate it. It is becoming an active participant in the workflow.
That shift is forcing enterprise software companies to rethink their products, interfaces, data architectures, and even the meaning of software as a service (SaaS). Salesforce is positioning customer relationship management as a context and execution layer. Workday is connecting probabilistic AI with deterministic business systems. Snowflake wants to become an operating layer for enterprise intelligence. Meanwhile, companies such as Actian, CloudMoyo, Cognizant, and HCLSoftware are focusing on the metadata, governance, orchestration, and control required to make this new stack work.
From systems of record to systems of action
Enterprise applications have traditionally served as systems of record. A customer relationship management platform stored customer interactions. An enterprise resource planning system maintained financial and operational records. A human capital management platform held employee information, organisational structures, payroll records, and permissions.
AI is pushing these platforms towards becoming systems of action.
At Salesforce’s Agentforce World Tour in Mumbai, Arundhati Bhattacharya, President and CEO, South Asia, Salesforce, used the evolution of self-driving cars to explain the change. Early autonomous vehicles retained the steering wheel, pedals, and familiar cabin design. Later versions began to question whether those interfaces were necessary at all.
The enterprise equivalent is significant. If an AI agent can retrieve customer information, check eligibility, apply business rules, and initiate an approved action, does the employee still need to navigate several application screens? Does every process need a dashboard, form, or queue designed exclusively for a human user?
Salesforce calls part of its emerging architecture Headless 360. The proposition is that humans, agents, data, and applications will interact through context layers, application programming interfaces, governance systems, and workflows. The traditional application screen remains, but it may no longer be the only doorway into enterprise software.
This does not make systems of record irrelevant. It may make them more important.
Aneel Bhusri, Co-founder, CEO and Chair, Workday, captured this tension by distinguishing between probabilistic AI and deterministic enterprise software. AI can reason, infer, and generate. But payroll, compliance, accounting, identity, and employee records cannot be approximately correct.
“AI only works in the enterprise when it’s connected to trusted, deterministic systems,” Bhusri said during a Workday leadership briefing.
The implication is that AI may change how employees interact with enterprise applications, but it still needs the rules, identities, permissions, and validated records held inside those systems. The interface may become conversational. The underlying need for control does not disappear.
Software stops being just a tool
For most of the software industry’s history, applications were tools operated by people. Users entered commands, clicked buttons, interpreted outputs, and decided what to do next.
Vala Afshar, Chief Digital Evangelist, Salesforce, believes that relationship is changing. He describes software as moving from a tool towards a “digital colleague”, capable of handling defined jobs, tasks, actions, and channels within specified guardrails.
That does not mean software acquires human judgement or organisational accountability. It means companies may increasingly assign repeatable work to agents in much the same way they assign responsibilities to teams.
In customer service, for example, an agent may retrieve account history, classify a request, check a policy, prepare a resolution, or complete a routine action. Human employees can then focus on exceptions, disputes, sensitive conversations, and situations requiring empathy or judgement.
Afshar calls this digital labour. His argument is that enterprises must decide which work should remain human, which tasks can move to agents, and where the two should collaborate. That makes AI adoption as much an operating-model decision as a technology deployment.
The same pattern is emerging inside Workday. Joel Hellermark, Senior Vice President and General Manager of AI, Workday, described a future in which employees ask for an outcome rather than manually co-ordinating every step.
“Instead of dozens of tickets and handoffs, you ask for an outcome and Sana delivers it,” he said.
A seemingly simple activity such as onboarding an employee may involve human resources, IT, procurement, calendars, email, identity management, service desks, and collaboration platforms. Today, users often carry the process from one application to another. In the emerging stack, an orchestration layer may co-ordinate those actions while applying the permissions and business rules of the underlying systems.
The interface begins to fade
The software industry has spent decades improving screens. It moved from desktop applications to browser-based SaaS, then to mobile applications, dashboards, low-code tools, and conversational assistants.
Agentic AI raises a more fundamental possibility: the most important software interaction may no longer begin inside an application.
A user could state an objective in Slack, email, WhatsApp, voice, or a workplace assistant. The agent could interpret the request, gather information from authorised systems, perform approved actions, and return the result through the same channel.
Salesforce’s architecture places Slack within the engagement layer, while its applications, data platforms, integration products, and agents sit underneath. Workday is similarly positioning Sana as an experience that can move across Workday and connected applications such as Salesforce, ServiceNow, Gmail, Jira, and Microsoft Outlook.
The application does not vanish. It becomes less visible.
This changes the competitive question for SaaS providers. The winner may not be the company with the most polished interface or the largest collection of features. It may be the platform that can provide reliable context, expose secure actions, integrate with other systems, and allow agents to complete work without losing control.
The new application experience may therefore sit above multiple products rather than inside a single product.
Data becomes context
AI models attract much of the attention, but enterprise agents cannot operate on models alone. They need business context.
Christian Kleinerman, Executive Vice President of Product, Snowflake, told Dataquest that Snowflake sees the “operating system for enterprise intelligence” as the direction in which it is moving. But he qualified that ambition. Snowflake, he said, must be strong in both AI and data.
“We do think that we will power the enterprise, power applications for the enterprise, and power workflows for the enterprise,” Kleinerman said.
Snowflake’s position reflects a larger contest within the software stack. Data platforms no longer want to remain passive repositories. They are adding agents, application-development capabilities, streaming, metadata, governance, and connections to multiple AI models. Their ambition is to become the layer where data, models, applications, and workflows meet.
But storing enterprise data is not the same as understanding it.
Ole Olesen-Bagneux, Vice President and Chief Evangelist at Actian, an HCLSoftware division, argues that agents need a much stronger metadata layer than traditional analytics systems required.
Analytics data was generally prepared for human interpretation. A person viewing a dashboard could notice missing information, question an unusual figure, or apply knowledge that was not explicitly present in the system. An agent acting directly on data may not have that advantage.
It needs to understand what the data means, where it originated, how it relates to other information, who is authorised to use it, and whether it is suitable for a particular action. If the metadata is fragmented, the agentic layer will inherit that fragmentation.
For Olesen-Bagneux, this makes metadata a trust layer for enterprise AI. The challenge is not merely connecting databases. It is connecting the organisation’s scattered semantic knowledge so that agents can interpret data coherently.
Workflows become the real battleground
The early enterprise AI market was dominated by copilots, assistants, and standalone experiments. These tools could summarise documents, answer questions, and help employees produce content.
The next phase is moving into workflows.
Manish Kedia, Co-founder and CEO, CloudMoyo, argues that enterprise AI must solve three connected problems: data unification, real-time intelligence, and intelligent operations.
Many pilot projects fail to scale because they begin with the visible AI tool rather than the systems underneath it. Data remains fragmented, governance is incomplete, and the workflow is not ready for automated execution.
“If you do not solve those three layers, your AI effort remains cosmetic,” Kedia said.
CloudMoyo’s work in contract intelligence illustrates the issue. Contracts may contain obligations, rights, pricing terms, penalties, and compliance conditions. But their value remains limited if the information is isolated from customer management, procurement, finance, supply chain, and operational systems.
The opportunity lies in connecting the process from end to end. An agent should not merely identify a contract clause. It should help determine what that clause means for a purchase order, supplier obligation, customer commitment, or operational decision.
That makes workflow integration, rather than conversational ability, the more important measure of enterprise AI maturity.
AI enters the software factory
The new software stack is also changing how applications themselves are built.
Singaravelu Ekambaram, Senior Vice President and Global Head of Delivery, Americas, Cognizant, told Dataquest that more than 30% of Cognizant’s code was already AI-generated. Agents are increasingly involved in planning, writing, testing, and refactoring software.
This is more than developer assistance. It suggests that AI is entering the software-development lifecycle as an execution layer.
Ekambaram sees the shift extending beyond coding. AI agents are beginning to close service tickets, process claims, complete parts of Know Your Customer procedures, trigger replenishment, resolve incidents, and initiate recovery actions in production environments.
However, greater autonomy creates a greater need for controls. Cognizant’s emphasis is on governed autonomy, which includes decision boundaries, identity, least-privilege access, monitoring, human escalation, rollback, and the ability to stop an agent.
Ekambaram’s broader point is that the autonomous enterprise will not be won by deploying the largest number of agents. It will depend on “engineered trust at scale”.
SaaS faces a control question
AI does not necessarily mean the end of SaaS. But it may weaken some of the assumptions on which the SaaS era was built.
The cloud-first model encouraged enterprises to consume standardised applications and infrastructure as services. AI is now pushing organisations to ask harder questions about model choice, data control, intellectual property, residency, continuity, and dependence on a single provider.
Kalyan Kumar, Chief Product Officer, HCLSoftware, describes this as a wider move towards reclaiming control across the AI stack.
Enterprises are not rejecting public cloud or SaaS. They are becoming more selective. Commodity workloads may remain in public services, while proprietary data, sensitive processes, and critical AI workloads may require private or hybrid environments.
Kumar argues that sovereignty involves more than the physical location of data. It also includes who owns the intellectual property, who can revoke access, where the technology can run, and whether the organisation can continue operating if a commercial relationship changes.
This introduces a new tension into enterprise software. Companies want the convenience, scalability, and continuous innovation associated with SaaS, but they also want flexibility over models, data, deployment, and control.
The result may be a more distributed software stack, with public services, private AI environments, open models, commercial models, enterprise databases, and orchestration layers working together.
Olesen-Bagneux takes the argument further by describing a possible shift from software as a service towards “services as software”. Instead of users opening a fixed application and navigating its features, software could increasingly assemble itself around the task that needs to be performed.
That would turn software from a static product into a more dynamic execution capability.
Governance moves into the architecture
As applications begin to act, governance can no longer remain a policy document reviewed after deployment. It has to sit inside the architecture.
An enterprise agent needs an identity. It needs clearly defined permissions. Its actions must be logged. Its access should be limited to what the task requires. High-risk decisions may need approval. Incorrect actions must be reversible. Enterprises also need to know who owns the outcome when an agent makes a mistake.
These requirements explain why identity systems, metadata, data lineage, observability, integration platforms, and policy engines are becoming central to the AI-era software stack.
The trusted system of record, the model, the context layer, and the workflow engine must work together. Remove one of them, and the enterprise either loses intelligence or loses control.
The stack is not disappearing. It is being rewired
The emerging enterprise stack will not be built around one model or one application. It will combine systems of record, AI models, enterprise data, metadata, integration, orchestration, security, identity, governance, and human judgement.
/filters:format(webp)/dq/media/media_files/2026/06/12/software-stack-2026-06-12-06-13-09.png)
Applications will remain, but users may interact with them less directly. SaaS will remain, but customers will demand more deployment choice and control. Data platforms will remain, but they will increasingly compete to provide context and power workflows. Developers will remain central, but AI will take on more of the coding, testing, and operational work.
Most importantly, workflows will no longer be designed only for human users.
The defining question for the next generation of enterprise software is therefore not whether an application has AI. It is whether the software can understand context, act across systems, respect enterprise controls, and know when to bring a human back into the process.
That is the real shape of the new software stack.
(Note: This feature draws on conversations Dataquest has had with industry leaders over recent months, along with relevant insights from our published reports and interviews.)
