AI and Enterprise Software: When Software Becomes the Worker

Enterprise software is entering its third major era. After decades as systems of record and systems of coordination, AI is enabling systems of execution. The challenge for enterprise leaders is no longer simply adopting AI, but redesigning operating models around software that can execute meaningful work.

Management, biology, and the hidden cost of coordination

Enterprise management evolved to compensate for human limits in attention, memory, communication, and scale. Modern enterprise software extended that management layer: most enterprise software coordinates work, it does not perform it. AI changes this economics by attacking the coordination burden directly rather than only speeding up individual tasks.

Management structures did not appear because organizations enjoyed complexity. They emerged because individual humans can track only a small number of priorities, remember a limited set of commitments, and hold only so much context in working memory. As teams grew, companies added layers of managers, checklists, meetings, and reports to keep work aligned. Enterprise software digitized those structures: task boards, queues, approval chains, and dashboards.

The cost of that coordination is now measurable at scale. Asana’s long-running "Anatomy of Work" research has repeatedly found that knowledge workers lose close to 60% of their time to what it calls "work about work"—status updates, chasing information, and switching between tools. Separate research from In Parallel on 247 managers across five countries estimates that managers lose 16.5 hours per week to coordination activities alone, at an annual cost of about USD $64,600 per manager. Coordination is not a side activity; it is the job.

For decades, the implicit assumption was that this overhead was unavoidable. Software could make it more transparent, but it could not make it disappear. Systems of record would store what happened. Workflow tools would route the next action to the next person. Meetings and emails would fill in everything the systems missed.

The entire operating model of the enterprise was built on one premise: people perform work, and software coordinates them. This is where AI introduces a discontinuity rather than another efficiency gain.

For the first time in enterprise computing, software is evolving from a system that tells people what to do into a system that can do the work itself.

Traditional automation tools assumed the process was fully specified in advance. They could move data, trigger alerts, or enforce simple rules, but they could not interpret ambiguous inputs or make context-sensitive trade-offs. Generative and agentic AI can read documents, synthesize histories, compare options, and propose actions in ways that look much closer to how humans manage exceptions today. 

As a result, the question is no longer only, "How do we make managers and staff more efficient?" It becomes, "Which parts of management exist solely because of human limits in attention, memory, and communication—and could now be executed by software?" That reframing is more radical than an architectural upgrade, because it challenges why entire categories of roles and systems existed in the first place.

From systems of record to systems of coordination to systems of execution

Over the last fifty years, enterprise software has moved through three broad eras: systems that store information, systems that coordinate people, and now systems that can execute work. The arrival of AI-executed work marks a structural break in how enterprises design operating models, not just in which tools they use.

In Era 1, software stored information. Mainframes, early ERP systems, and policy administration platforms digitized paper records. The value thesis was straightforward: replace filing cabinets with databases. Financial ledgers, customer details, policies, and claims histories all became machine-readable. These systems of record were built for reliability, auditability, and consistency, not for flexibility. They were the memory of the enterprise.

Era 2 added systems of coordination. As client–server and web architectures matured, software became the digital manager. CRM platforms routed leads to salespeople. Claims systems assigned work to adjusters. Workflow engines moved tickets between queues. Tools such as Jira, ServiceNow, and modern policy or claims platforms orchestrated who should do what next. As one commentator on the "AI Service Operator" model put it, software "never did the work; it coordinated the work." Decision-making, improvisation, and exception handling stayed with humans.

This era quietly reshaped management. Instead of shouting across a room or walking the floor, managers relied on dashboards, service-level agreements, and workflow reports. The system told them which queues were aging, which steps were overdue, and which approvals were stuck. Software deepened the management layer and made it measurable, but it did not change the underlying assumption that humans were the actors performing the work.

Era 3, now emerging, is different. AI systems can interpret context, reason over unstructured information, and propose or carry out actions across multiple tools. Research from firms such as BCG suggests that when organizations redesign processes for "agentic" AI—where agents own outcomes and orchestrate other systems—they can see up to 60% long-term cost reductions and an 80% reduction in cycle time, far above earlier incremental automation programs. The gains appear when software is allowed to execute work end to end, not only pass tasks between people.

This shift is more than rebranding "automation" with new language. In a traditional workflow, software waits for a human to read a file, interpret it, decide on the next step, and then click a button. In an execution-first model, software reads the file, interprets it, prepares the decision, and, in low-risk cases, carries it out under pre-agreed controls. Humans design, supervise, and refine the system; they no longer perform every intermediate step.

The crucial point is that these eras now overlap in the same organization. Core systems of record still hold the data. Coordination tools still route exceptions and regulatory edge cases. But a growing share of the routine execution can move into AI-driven services that sit above or alongside those cores. The challenge for leaders is to intentionally decide which capabilities belong in each era, rather than allowing a patchwork of assistants and agents to emerge without an operating model.

When software becomes the worker: claims as the proving ground

Property and casualty claims offer a concrete example of software becoming the worker. Historically, claims platforms coordinated adjusters; now AI can read documents, propose decisions, and in some segments execute end-to-end journeys, while humans govern outcomes and edge cases.

Consider a typical auto physical damage claim. In the Era 2 model, the system logs first notice of loss, assigns a claim number, routes the case to an adjuster, and tracks each step in a workflow: coverage verification, liability assessment, estimate review, payment, and closure. The actual work—reading policy language, interpreting repair estimates, drafting letters, and negotiating settlements—sits with the adjuster. The platform is the manager; the adjuster is the worker.

In an AI-executed model, that balance starts to invert. An AI service can ingest photos, repair estimates, telematics data, and prior claim histories. It can compare those against thousands of similar claims, propose a liability split, generate a reserve range, and draft communication to the policyholder and repair shop. For low-complexity losses within defined thresholds, the system can often reach a settlement proposal that is as consistent as a human, but far faster.

Early programs in claims automation already demonstrate the potential. Case studies and consultancy analyses report that when carriers combine AI-driven document processing and triage with redesigned workflows, simple claims can move from days to minutes, with 30–50% reductions in handling costs for high-volume segments. Leakage reduction is also material: highlighting missing deductibles, duplicate payments, or subrogation opportunities can reduce soft leakage that often ranges from 7% to 14% of paid losses in many lines.

The subtle but important distinction is where accountability sits. The AI is not a bolt-on tool sitting in a separate sandbox; it becomes the primary execution layer for a defined claim band. The human adjuster or claims leader designs the guardrails, tunes the models, and governs exceptions. They move from performing each step to managing the behavior of a software worker operating at scale.

This reframing clarifies why many "copilot" pilots stall. When copilots are added as side-panel helpers on top of legacy cores, they remain advisory. They summarize files or draft emails, but the underlying process is unchanged: a person still acts as the central worker. The economics of management do not shift. The real discontinuity appears only when leaders design for specific journeys where software can own execution, with humans as supervisors.

Claims is an ideal proving ground because it is document-heavy, rules-bound, and outcome-driven. But the same pattern applies to other domains: invoice processing, simple underwriting endorsements, standard customer service workflows, and even parts of HR and finance. In each case, the test is whether software can perform the majority of steps under clear rules, within an acceptable risk envelope, while leaving humans to handle ambiguity, edge cases, and continuous improvement.

How leaders should redesign management for AI-executed work

When software executes work, management shifts from supervising people to designing, monitoring, and improving systems. Leaders must redefine roles, metrics, and coordination mechanisms so teams govern software workers instead of compensating for human limitations.

In a people-executed process, much of a manager’s energy goes into allocation and oversight: assigning tasks, resolving conflicts, chasing status updates, and escalating exceptions. Coordination time expands with team size. That is why research like Asana’s and In Parallel’s finds such large fractions of the week consumed by meetings, updates, and re-explaining context. The management layer is effectively a manual operating system of work.

In an AI-executed model, the locus of coordination moves into software. Agents and services orchestrate tasks between systems, apply business rules, and track outcomes in real time. Managers are still essential, but their job changes. Instead of asking, "What is the status of this claim or case?" they ask, "How is the system behaving? Where is it uncertain? Which patterns of exceptions signal that we need to adjust our rules, training data, or thresholds?"

This shift calls for different skills. Managers need to become comfortable with concepts such as confidence scores, guardrails, policy-as-code, and experiment design. Frontline leaders participate in configuring decision boundaries: which transactions can be fully automated, which require human review, and which must always be handled manually. They also take responsibility for monitoring fairness, leakage, and customer outcomes across large cohorts, not only individual cases.

Team structures change as well. Instead of organizing solely by function—claims, underwriting, service—organizations can create cross-functional "journey teams" responsible for an outcome such as "settle simple auto claims under USD $5,000 with high satisfaction and minimal leakage." Those teams own both the AI workers and the human experts who handle exceptions. Their weekly rituals focus less on task allocation and more on system performance: reviewing dashboards, exploring outliers, and deciding which improvements to deploy next.

Crucially, metrics must evolve. Traditional productivity metrics count cases per adjuster or tickets per agent. In an AI-executed environment, leaders need metrics that capture system-level performance: straight-through processing rates, exception volumes, cycle times by journey, and coordination time saved. Studies of AI-enabled operations already show that when organizations track and optimize these measures, they can move beyond 10–20% task-level gains to step-change improvements in end-to-end economics.

Risk, governance, and accountability when software acts

As software takes on decisions and actions, governance becomes central. Executives must design transparent, auditable, and regulator-ready control frameworks so AI execution remains explainable, fair, and aligned with policy intent.

Regulated sectors such as insurance, banking, and healthcare face a dual challenge. On one hand, AI offers the possibility of faster, more consistent decisions and lower operational risk from manual errors. On the other, regulators and oversight bodies expect clear accountability, traceability, and control. Supervisory guidance on AI governance—from insurance supervisors to financial regulators—consistently emphasizes transparency, human oversight, and the ability to reconstruct decisions ex post.

An execution-first model can support these expectations when it is built on an event-driven architecture. Every AI-influenced action can emit structured events that capture who initiated it, which model and version were used, what evidence was considered, what was suggested, and what ultimately happened. For example, when an AI service proposes a reserve change on a claim and a human accepts it, the system can log both a "reserve.proposed" and "reserve.updated" event with detailed metadata.

Over time, these events create an auditable trail that is often richer than traditional manual notes. Risk and compliance teams can reconstruct contentious decisions, compare how different cohorts are treated, and monitor override patterns. If they notice that a particular model’s suggestions are frequently rejected, they can trigger a review. If they observe that certain customer segments consistently experience longer cycle times or higher denial rates, they can investigate for bias.

Governance is not only about logs. It also requires clear role definitions. Executives need to decide which decisions can be delegated to software under policy, which require human sign-off, and which are prohibited from automation. They should maintain catalogs of AI use cases that specify allowed actions, risk ratings, controls, and accountable owners. This discipline mirrors traditional model risk management but extends it to operational AI agents and services.

External guidance can help shape these frameworks. Industry analyses, including resources from regulators and professional bodies, increasingly point to practices such as model inventories, impact assessments, and continuous monitoring. What is new in the AI-executed era is the scale and speed of decisions. That makes upfront design and ongoing oversight not just regulatory necessities but strategic capabilities.

Practical first steps: reframing AI programs around execution

To capture the benefits of AI-executed work, organizations should move beyond scattered copilots and instead target specific journeys where software can own execution under clear guardrails, starting small and scaling through evidence.

The first step is conceptual. Leaders should explicitly adopt the framing that for fifty years, enterprise software has been built on the assumption that people perform work and software coordinates them. AI now makes the reverse plausible in selected domains: software can perform work and people can coordinate it. This is not a slogan; it is a design principle that shapes where to invest.

From there, practical moves are straightforward but demanding. Identify one or two high-volume, rules-bound journeys where end-to-end execution is feasible—such as low-severity auto claims, standard invoice processing, or routine customer requests. Map the current process, including manual rekeying, handoffs, and approval steps. Quantify the coordination load: meetings, status checks, and time spent searching for information across systems.

Next, design an execution-first target state. Define which steps the AI service will perform, which decisions it can make autonomously, and where human review is mandatory. Establish clear thresholds for value at risk, customer importance, and regulatory sensitivity. Build or expose the minimum interfaces needed for the AI to act: retrieving records, updating statuses, sending communications, and triggering payments within existing control frameworks.

Crucially, embed measurement from the start. Track cycle time, handling cost, leakage, and satisfaction for the journey before and after the new model. Use findings from studies such as BCG’s agentic deployments—which report up to 60% cost reductions and dramatic cycle time improvements when processes are redesigned for AI—to set ambitious but realistic targets. Share results with operational leaders and risk committees in precise, non-technical language.

Finally, treat the first journey as a template, not a one-off. Document the governance model, event structures, escalation paths, and operating rituals. Reuse these patterns as you expand into adjacent journeys and domains. Over time, the organization builds a portfolio of software workers governed by a consistent management and risk framework.

The strategic opportunity is clear. For half a century, enterprises have accepted coordination tax as the price of doing complex work at scale. AI makes it possible to design operating models that reduce that tax dramatically by moving execution into software. The organizations that succeed will not be the ones with the most impressive demos. They will be the ones that recognize that software has stopped being only the manager—and has started to become the worker.

For half a century, enterprise software evolved by helping people coordinate increasingly complex work. AI marks the first era in which software itself can execute meaningful portions of that work. The strategic question for executives is therefore no longer which workflows to automate, but which work should continue to require humans at all. Organizations that answer that question deliberately will redesign management itself. Those that don’t will simply layer AI onto operating models built for a different era.