Enterprise AI success depends on redesigning how work gets done

Most enterprises are still deploying AI without transforming their businesses

As much as 90% of enterprises remain focused on deploying AI tools and pursuing narrow applications, while a smaller group reports substantial financial gains from broader transformation. Bain & Company estimates that the leading 10% have developed more disciplined approaches to AI. Among transformation leaders, the consultancy reports EBITDA growth of 10% to 25%. EBITDA measures earnings before interest, taxes, depreciation, and amortization.

The gap raises a question for business leaders: Why are so many companies struggling to capture AI’s value when the technology itself is improving so quickly?

Bain’s analysis, “The AI-Native Enterprise: Absorption Is the New Advantage,” points to how organizations integrate new technology into their operations. By 2024, leading enterprises had established transformation programs with executive sponsorship, defined business objectives, and coordinated implementation across functions. These programs linked AI deployment to changes in how work was organized and measured.

Companies taking a narrower approach often focus on individual productivity. An AI assistant might help employees write documents faster, summarize customer interactions, or complete routine tasks. Those improvements save time, but the financial benefit can remain limited when the surrounding business process stays unchanged.

Consider a workflow where AI reduces the time needed for one activity. If approvals, handoffs, and downstream operations continue at their previous pace, the organization may see little improvement in total processing time. The employee becomes more productive, but the process retains its constraints. Meaningful financial gains therefore depend on changing how work moves through the business.

Bain’s reported EBITDA growth describes the performance of enterprises pursuing broad transformations. The figure represents their overall financial performance rather than a separate measurement of AI’s contribution to earnings. Still, the difference between these organizations and companies running isolated experiments shows why implementation matters.

Bain, which advises enterprises on technology and organizational transformation, argues that the gap can widen over time. Every quarter spent in pilot mode gives more advanced competitors another quarter to redesign operations and improve execution. Organizations waiting for AI technology to mature may therefore be delaying management changes that already require considerable time and effort.

The biggest barriers to AI progress are often organizational

The difficulty of implementing AI is frequently explained through technical concerns. Data security remains challenging, generative models can produce incorrect information, and returns on investment can be uncertain. These issues matter particularly when AI systems handle sensitive information or make decisions affecting customers and business operations.

Yet Bain’s transformation work suggests that organizational barriers often determine whether companies move beyond experimentation. Its Bain GenAI Survey 2026 examines this gap in Figure 4, titled “Enterprises blame technology, but the real barriers are organizational.” The finding directs attention to management decisions that determine whether technically capable systems become part of everyday work.

A 2025 Massachusetts Institute of Technology study identified several obstacles: difficult change management, uncertain executive support, poor user experience, and employee resistance to unfamiliar tools. Each can prevent an AI application from becoming part of normal operations. Even a technically reliable system may deliver limited value when employees avoid it, managers preserve existing responsibilities, or no executive owns the business outcome.

Executive sponsorship becomes especially important when improving a process requires several departments to change together. Consider customer service involving support representatives, technical specialists, billing teams, and account managers. AI might help one team answer questions faster, but improving the complete process requires agreement on ownership, authority, escalation procedures, and performance expectations.

Without senior leadership, departments can preserve their existing approval structures while introducing separate AI applications. Each team may improve its own activities, yet delays between teams remain. The organization accumulates tools without developing a coordinated way to use them.

Bain identifies three common responses that reinforce this pattern. Companies remain in pilot mode, introduce AI into existing processes without fundamentally changing them, or postpone major decisions while waiting for technology platforms to stabilize. All three delay changes to the organization itself.

Bain describes the second approach as “pave the goat trails.” The phrase refers to automating activities within an existing process while preserving its underlying structure. AI may accelerate individual steps, but unnecessary reviews, repeated data entry, and fragmented responsibilities continue consuming time.

A broader redesign starts with the intended business outcome. Managers determine which activities remain necessary, which decisions require human judgment, and how information should move between people and systems. This approach can eliminate entire steps while changing responsibilities across departments.

Bain reports double-digit efficiency improvements across customer service, sales and marketing, software development, operations, and back-office functions in well-designed transformations. Figure 1, “Leaders are reporting significant improvements from AI in well-architected transformations,” links these gains to coordinated implementation and changes in operating processes.

Achieving such improvements requires substantial investment beyond software licenses and model access. Among companies Bain works with, organizations spend approximately four dollars on people and process changes for every dollar spent on technology. The consultancy benefits commercially when enterprises invest in transformation services, so its spending observations should be understood in the context of its client work.

The four-to-one relationship also shows how much management work implementation involves. Enterprises must define outcomes, assign accountable executives, redesign processes, modernize applications and data environments, train employees, and measure results. These activities determine whether AI changes business performance.

Enterprise knowledge determines what AI can do

Once responsibilities and workflows are defined, the next requirement is reliable access to business information. AI systems need formal data stored in applications, along with the knowledge employees use to interpret it. Without this context, a model may understand a request but lack the information needed to complete it correctly.

Enterprise knowledge includes systems, standards, workflows, exceptions, customer relationships, decision rights, and accumulated institutional judgment. An AI agent handling a customer request may need contractual terms, previous interactions, approved exceptions, and rules governing managerial authorization. An agent is an AI system that uses information and tools to carry out actions toward a defined objective.

Making this knowledge available requires engineering and governance. Information must be accessible to authorized employees and AI systems, while permissions prevent inappropriate disclosure or action. Applications also need reliable connections so agents can retrieve information and complete permitted tasks.

These foundations retain their value as models improve. A future model may reason more accurately or execute tasks at lower cost, but it will still need the company’s operating information and appropriate authority to act. Better intelligence increases the potential value of reliable enterprise context.

So waiting for the technology market to settle can carry a cost. Process ownership, employee adoption, application modernization, and organizational coordination take time to establish. Enterprises that develop these foundations can incorporate new model capabilities more quickly when they become available.

AI vendors are investing billions to close the implementation gap

The difficulty of turning AI into working business systems is changing how technology suppliers invest. Model development remains central to competition, but major vendors are also funding services and operating structures that help customers deploy AI within existing organizations. These investments reflect a commercial opportunity: suppliers benefit when enterprises move from experimentation to sustained adoption.

Five commitments show the scale of the shift.

These commitments cover different business structures and activities, making them separate investment programs rather than directly equivalent spending categories. Together, they show suppliers allocating substantial resources to the work required after customers gain access to AI technology.

A key role in this effort is the forward-deployed engineer, or FDE. These engineers work directly with customers to connect technology to operational systems, integrate applications and data, resolve implementation problems, and deliver functioning software. Working closely with customers helps them understand requirements that are difficult to capture in a standard product specification.

Palantir has made this approach prominent. Other suppliers are expanding embedded engineering teams because enterprise deployments depend on customer-specific data structures, business rules, application dependencies, and workflows. Engineers working alongside customers can address these requirements as they emerge during implementation.

Microsoft Frontier Company represents one version of this approach through an operating business that places engineering specialists with customers. Amazon Web Services is funding a dedicated forward-deployed engineering organization. Anthropic, OpenAI, and Google Cloud are developing their own structures to accelerate enterprise adoption.

The emphasis on customer implementation also appears in hiring patterns. Bain’s Figure 2, “Firms are hiring roles designed to build and deliver AI outcomes,” examines the functions companies are staffing to support AI deployment.

The analysis distinguishes revenue operations, covering sales, marketing, and customer service, from solutioning roles, which include technical presales, forward-deployed engineers, and other delivery specialists. It also includes general and administrative functions, commonly called G&A. The focus on solutioning reflects the technical work needed to turn general-purpose AI into specific business results.

A customer purchasing access to an advanced model still needs to connect it to internal applications, establish secure information flows, configure workflows, and validate performance. Vendor specialists can help complete those tasks and identify feasible applications. Their involvement can reduce the customer’s engineering burden and shorten deployment time.

But implementation services have an organizational boundary. A vendor’s engineers cannot independently decide how the customer’s departments should operate, which executives should own results, or how employee responsibilities and incentives should change. Those decisions require authority within the customer organization.

A supplier can help build an AI-supported customer-service workflow, for example, while the customer determines how it affects staffing, escalation procedures, quality standards, and accountability. Technical implementation and management redesign must therefore proceed together.

This creates a shared interest between suppliers and enterprise customers, alongside different commercial incentives. Vendors want customers to adopt and expand their use of AI services. Enterprise leaders need those deployments to produce measurable improvements in their own business performance.

Better models still matter. enterprises need the flexibility to use them

Implementation quality determines whether AI becomes useful in an organization, but the intelligence of the underlying model still affects what a system can accomplish. In demanding workflows, small differences in model accuracy can determine whether a process operates reliably enough for production.

Meanwhile, the model market is becoming more diverse. Open-weight models, whose parameters are available for organizations to use and adapt under applicable licenses, have narrowed capability gaps with proprietary systems across many enterprise applications.

Epoch AI tracks this development through the Epoch Capabilities Index, or ECI. The index combines results from multiple AI benchmarks into a measure of overall model capability. Its analysis, referenced in Figure 3, “Open-source models are narrowing the capability gap with proprietary models,” indicates increasing competition between openly available and closed models.

Greater competition gives enterprises more choices about performance and cost. Applications that previously required an expensive frontier model may eventually run effectively on a lower-cost or specialized alternative. This can reduce the cost of processing AI requests, commonly measured through tokens, the units of text and other content that models process.

As models improve, applications with mature, well-understood requirements may move toward efficient systems optimized for cost and reliability. New applications involving difficult reasoning, uncertain environments, or demanding scientific problems may continue to require the strongest frontier models.

The economic value of model intelligence therefore depends on the task. Falling token prices and improving open-weight models create pricing pressure for established capabilities. But a more capable model can retain substantial value when weaker systems cannot complete the required work reliably.

Bain expects frontier models to tackle increasingly difficult problems, potentially including disease treatments, fusion, and quantum gravity. These possibilities show why leading model performance may remain valuable even as common capabilities become cheaper.

Small accuracy differences become large workflow differences

The importance of model quality becomes clearer when AI performs a sequence of dependent actions. In an agentic workflow, an AI system uses tools, information, and intermediate decisions to complete a multistep task. Each step introduces another opportunity for failure.

Consider a process containing 34 dependent steps. If every step must succeed, the probability of completing the full process depends on the reliability of the individual steps.

Bain illustrates the effect with two models.

The calculation assumes every step must succeed and that success probabilities multiply across the workflow. Under those assumptions, end-to-end success equals the individual-step success probability raised to the 34th power.

A two-percentage-point improvement in step accuracy therefore produces a substantial difference in overall reliability. In this example, the stronger model completes the workflow successfully roughly twice as often.

That difference can determine whether automation is economically practical. A process that fails most of the time may require extensive human correction, reducing or eliminating the expected savings. Greater reliability lowers the need for intervention and makes longer workflows more suitable for production.

Model selection should therefore reflect each task’s requirements. A stronger model may be justified for difficult decisions, while routine activities can use cheaper alternatives that meet the necessary accuracy threshold.

The next challenge is preserving the ability to change those selections as models improve.

Model flexibility gives enterprises control

Palantir’s Alex Karp emphasized enterprise control over model selection during the company’s August 2026 earnings call:

“You own the weights, you own the alpha, you own everything.… We have a product that allows you to switch out models.”

Karp’s statement reflects Palantir’s commercial interest in providing software that connects enterprise systems with different AI models. The company benefits when customers use its platform to manage those connections and retain flexibility over underlying model providers.

Microsoft’s Satya Nadella described a related objective during its July 2026 earnings call:

“[T]he goal is to have the firm be in control of their own destiny.… And the models are an input.… That means any given model at any given time is swappable.”

Microsoft also has a commercial stake in enterprise AI infrastructure and platforms. Although Microsoft and Palantir compete in parts of the enterprise AI market, their executives both emphasize customer control over model selection. Their statements support similar architectural principles while reflecting each company’s commercial position.

The practical requirement is an architecture that allows enterprises to select models according to business needs and revise those choices over time. This reduces dependence on any individual model provider and creates opportunities to adopt improvements in accuracy, speed, and cost.

One way to provide that flexibility is through an LLM harness. A large language model, or LLM, is an AI model trained to process and generate language. A harness is the software surrounding the model that connects it to business systems and manages how it performs work.

The harness provides the context, tools, and controls needed for enterprise execution. It can retrieve data, maintain memory, enforce permissions, coordinate actions, evaluate results, and route tasks to suitable models.

For example, an AI agent receiving a customer request may need account information, previous interactions, and access to several internal applications. The harness retrieves relevant context, checks which systems the agent is authorized to use, and supplies the tools needed to complete the request. It can then evaluate the result and record information about the execution.

Model routing becomes useful when different stages of a process have different requirements. An inexpensive model might classify a request or extract structured information, while a stronger model handles a difficult interpretation or decision. Orchestration software coordinates these actions and manages dependencies between steps.

Evaluation mechanisms allow the enterprise to learn from execution. Completion rates, errors, and other results can inform future model choices, instructions, tool configurations, and workflow changes. This connects architectural flexibility to measurable performance improvement.

Interoperability supports changing model choices

A flexible architecture also depends on connections between components. Model Context Protocol helps standardize how AI applications connect to external tools and data sources. Agent2Agent supports communication between agent-based systems.

These standards make it easier to combine components from different providers. Yet replacing one model with another can still require adjustments to prompts, output handling, tool use, and evaluation. Model portability reduces integration work, although each implementation still needs testing.

The responsibilities of the harness may also change over time. Some functions currently handled by external software could become native model capabilities. Application platforms may take responsibility for other parts of execution as their products evolve.

Enterprises therefore benefit from organizing their systems around business requirements while keeping implementation choices adaptable. Models, infrastructure, agent platforms, enterprise context, and applications remain areas of active competition. No single supplier currently controls every layer.

The architectural decision matters because enterprises must balance performance, cost, security, and supplier dependence across changing technologies. A system designed to evaluate and replace components can make those trade-offs repeatedly as the market develops.

HubSpot shows how model flexibility becomes business performance

HubSpot provides an example of how flexible model selection can support working AI applications. Its approach combines an AI abstraction layer, customer-specific information, and agents that execute tasks within its business software.

An abstraction layer separates application logic from the details of the underlying model. Developers can then evaluate alternatives and change providers without rebuilding the entire application. This separation matters because models improve at different rates and perform differently across tasks.

Without this layer, changing models may require developers to revise integrations, adjust prompts, modify output handling, and retest workflows that depend on provider-specific behavior. HubSpot’s architecture reduces that dependency by allowing teams to evaluate emerging models and substitute suitable alternatives while preserving the surrounding application.

The arrangement gives HubSpot a repeatable process for incorporating advances in AI. When a new model becomes available, engineers can compare its accuracy, speed, and cost with existing options. They can then determine whether it improves a particular workflow and introduce it through the established system.

Customer context makes general-purpose models useful

Model flexibility becomes more valuable when applications connect intelligence to relevant business information. General-purpose models contain broad knowledge and reasoning abilities, but customer-specific tasks require information from the applications where work takes place.

For HubSpot, that context includes customer interactions, account records, sales activities, and other information needed to complete specific tasks. Connecting models to these records allows agents to interpret requests and act according to the customer’s actual business situation.

Consider a support agent responding to a customer issue. It needs to understand the request, retrieve relevant account information, identify the appropriate resolution, and complete permitted actions. Model intelligence helps the agent reason through the problem, while customer context supplies the facts needed for an accurate response.

HubSpot reports that its AI agents now resolve more than 70% of support tickets. This means agents handle a substantial share of work in an established customer-facing function. The measure describes ticket resolution by AI agents, while customer satisfaction and resolution quality remain separate performance considerations.

The company has also applied agents to sales prospecting. Its HubSpot Prospecting Agent generated nearly twice as many meetings as it had in the previous year. That result connects AI use to an output that matters directly to the sales organization.

Prospecting involves understanding potential customers, identifying opportunities, preparing outreach, and supporting engagement that leads to meetings. Agents become more useful when these activities draw on customer records and operate within the existing sales workflow.

The increase in meetings shows how AI can affect the volume of work a business process produces. The benefit extends beyond time saved by individual employees into the performance of the sales function.

A repeatable system makes improvement easier

HubSpot’s reported results reflect several working components: model evaluation, proprietary customer context, application integration, and agent execution. The abstraction layer supports these activities by making it easier to introduce new models as requirements and capabilities change.

That flexibility also affects product development. HubSpot reports that its architecture has accelerated development by allowing teams to incorporate new AI functionality more quickly. Engineers can spend more effort evaluating useful capabilities and less effort rebuilding connections to individual providers.

The underlying advantage is repeatability. Integrating one model into one application produces a working implementation. Establishing a process for evaluating models, connecting them to business information, updating workflows, and measuring performance creates a system that can improve across successive releases.

This is what enterprise AI absorption means in practice. An organization identifies a useful new capability, evaluates its performance, incorporates it into existing operations, and measures the result. The surrounding technology and management processes support repeated improvement.

Once an updated workflow is deployed, it generates evidence about accuracy, cost, completion rates, and other business outcomes. Engineers and business owners can use those findings to revise model choices, instructions, integrations, or operating rules. Each cycle builds experience that helps the organization manage later changes.

HubSpot’s approach also shows why proprietary enterprise information remains valuable as model access becomes more widely available. Competitors may use similar general-purpose intelligence, but its usefulness depends on the information and workflows connected to it.

For a support agent, company records and established resolution procedures help determine whether a task is completed correctly. For a prospecting agent, customer context and sales integration influence whether its activity produces useful meetings. The advantage comes from combining adaptable model selection with the specific environment where work happens.

That combination creates opportunities for deeper change. Once AI can reliably access information and execute tasks, enterprises can reconsider the structure of the process itself.

Klarna shows how redesigning work creates business value

Klarna reduced its campaign production process from six weeks to seven days by introducing AI agents into image generation and creative development. The company also increased creative output and established a weekly schedule for updating campaign imagery.

The shorter production cycle changed how quickly marketing could respond to commercial conditions. A lengthy production schedule requires teams to make creative decisions well before the relevant retail activity. Faster production allows content to reflect more recent information about products, promotions, and customer interests.

Klarna’s AI agents supported visual creation and other creative-development activities. These systems made it practical to produce and revise more campaign material within a shorter period. The company could then refresh imagery weekly to align with retail events.

The operational improvement involved the entire production schedule. Faster individual creative tasks contributed to a workflow that could respond more frequently to market changes. More current campaign material also improved the customer experience by keeping imagery aligned with relevant promotions and products.

The reported results establish a substantial improvement in production speed and output. They also show how process redesign can support a more responsive customer experience, while the financial contribution of those changes remains a separate business measure.

For marketing functions, responsiveness matters because customer behavior, competitor activity, product availability, and commercial events change continuously. A process built around infrequent production cycles limits how quickly those changes can influence campaigns.

Klarna’s revised workflow creates more opportunities to respond while commercial activity is underway. AI-supported production makes repeated creative updates practical, allowing marketing decisions to use more current information.

Bain captures this change in emphasis:

“The work moved from getting ready to getting better.”

The same principle can shape more complex activities, including product launches that involve several departments.

AI agents can change how product launches are managed

Traditional product launches often require months of preparation across design, marketing, and sales. Each function develops its contribution, coordinates with other teams, and prepares analyses and materials before launch.

Much of this work attempts to anticipate future conditions. Teams estimate customer responses, channel performance, competitive developments, and market demand before the product becomes available. Those assumptions shape the launch plan, even though actual performance information arrives later.

AI-supported workflows create an opportunity to move more analysis and adjustment into the period after launch. Agents can gather information continuously, prepare findings, and support decisions as commercial conditions become clearer.

Bain describes an emerging operating model in which product-launch leadership is concentrated around three senior executives supported by specialist AI agents. The agents handle communication, competitive analysis, channel analysis, and sell-through analysis. Sell-through analysis examines how products move through sales channels to customers, helping teams understand demand and commercial performance.

This arrangement changes how work is divided. Agents repeatedly collect information, conduct analysis, prepare materials, and report results. Senior leaders use those outputs to interpret performance, set priorities, and decide which changes to make.

In this model, the three leaders review markdown files produced through the AI-supported process. Markdown is a simple text format for organizing structured documents. The files provide a practical way to present findings, analyses, and recommendations for executive review.

With routine information gathering handled by agents, leaders can focus on judgment and oversight. They spend more time evaluating the launch and refining execution, while supporting systems monitor developments between decisions.

The timing of decisions changes as well. Extensive preparation previously attempted to resolve many questions before launch. Continuous monitoring allows teams to adjust plans using information collected after the product reaches the market.

Competitive analysis might identify changes in rival positioning. Channel analysis can reveal differences in performance across distribution routes, while sell-through information indicates where demand is stronger or weaker than expected.

Those findings can guide decisions about communication, marketing allocation, sales priorities, and product positioning. Their value depends on connecting ongoing analysis to leaders with authority to change the operating plan.

The three-leader structure illustrates one emerging organizational design. It shows how departmental responsibilities can change when AI performs activities that previously required substantial coordination between teams.

The same reasoning applies across enterprise workflows. Managers can begin with the desired business outcome, identify activities that agents can perform reliably, and define where human intervention provides the greatest value.

This changes how responsibilities are assigned. People remain accountable for strategy, priorities, customer relationships, and material business risks. Agents support those responsibilities by producing timely information and completing authorized tasks.

The division of work also determines how the process can improve. If agents capture information about each execution, the organization can use that evidence to refine future decisions and actions.

Continuous learning turns execution into improvement

An AI-native workflow can record what happened during execution, evaluate the result, and use the findings to improve subsequent activity. Over time, the organization can adjust models, instructions, tools, and operating rules based on observed performance.

Consider the product-launch process. Agents analyze channel results after launch, identify patterns in customer response, and generate recommendations. Leaders review those findings, approve appropriate changes, and evaluate whether they improve performance.

The information produced during execution then becomes an input to the next cycle. Frequent feedback reduces dependence on periodic reviews and allows organizations to respond as conditions change.

Some parts of this process can run automatically. Systems can capture completion rates, errors, customer interactions, and other performance data, then feed that information into evaluation mechanisms. Within predefined limits, software may also adjust how tasks are routed or executed.

Other improvements require human judgment. An agent might identify an underperforming campaign, but a manager may need to decide whether to change its audience, creative direction, budget, or business objective.

The design challenge is deciding which adjustments are permitted automatically and which require explicit approval. Permissions, evaluation criteria, and oversight connect automated activity to business priorities and acceptable risk.

This requires more than collecting performance dashboards. Operational information must influence subsequent execution through a defined process. The organization needs a way to identify problems, determine their causes, introduce changes, and evaluate the results.

A customer-support agent that repeatedly fails to resolve a particular type of request provides a useful example. The failures may result from missing customer context, inadequate tool permissions, unsuitable instructions, or insufficient model capability.

An evaluation system can help identify the cause. Engineers and business owners can then modify the relevant component and measure whether subsequent requests are resolved more successfully.

The same process applies to campaign production and product launches. Each completed activity generates information that can improve the next one, provided the organization captures results accurately and has mechanisms for acting on them.

Bain describes the intended direction of this development:

“The goal is not a faster process; it is a process that gets faster on its own.”

The ambition extends beyond the improvements demonstrated by Klarna and the emerging product-launch structures. It describes workflows where feedback influences execution increasingly quickly, with automation handling appropriate adjustments and people governing consequential decisions.

Bain explores the importance of enterprise-specific knowledge in its related brief, “Proprietary Intelligence: How to Win with AI.” The argument connects proprietary business information with the ability to make AI systems more useful in the specific environments where they operate.

That information becomes more valuable when an organization learns from repeated execution. Performance data reveals where processes fail, which interventions work, and how subsequent activity should change.

The mechanism requires deliberate design. Feedback must be captured accurately, connected to the appropriate workflow, evaluated against business objectives, and translated into changes that can be tested.

For enterprise leaders, this creates a concrete management responsibility: establish clear ownership of outcomes, give technical teams the systems needed to evaluate and improve execution, and define the authority under which agents can act.

Bain expresses the competitive significance of this ability directly:

“The companies pulling ahead are not the ones with the best models; they are the ones that have learned to absorb and act faster than their competitors.”

Bain’s position reflects its consulting focus on organizational transformation, while the operating mechanism is visible in the examples of HubSpot, Klarna, and AI-supported product launches. Each connects technology to business information, execution, and measurable results.

As AI systems become more capable, the scope of work they can perform will expand. Enterprises that capture evidence from completed tasks can use it to improve instructions, model choices, permissions, and workflows. Those changes can then be tested during the next execution, allowing each round of work to improve the one that follows.

Recap

Enterprise AI performance increasingly depends on how effectively organizations turn new capabilities into changes in business operations. Access to advanced models is becoming more widespread, but the ability to integrate those models into workflows, connect them to proprietary knowledge, and measure their impact remains uneven.

For executives, the priority is to move beyond evaluating AI tools in isolation. Each investment should connect to a defined business outcome, with clear accountability for implementation and performance. That requires coordination across technology, operations, finance, and business functions, particularly when improvements depend on changing responsibilities or removing unnecessary steps.

Technology decisions also need to preserve flexibility. As models improve and costs decline, enterprises benefit from architectures that allow them to evaluate alternatives without rebuilding entire applications. Reliable access to business information, effective governance, and continuous performance measurement provide a foundation for adopting new capabilities as they emerge.

The experiences of HubSpot and Klarna demonstrate how these principles translate into operational results. Their reported improvements come from integrating AI into established business processes, changing how work is executed, and creating opportunities for faster adjustment.

The next stage of enterprise AI will depend increasingly on continuous improvement. Organizations that capture execution data, evaluate outcomes, and use those findings to refine workflows can build systems that become more effective over time.

For business leaders, the competitive question is shifting from how quickly AI can be deployed to how quickly the organization can turn AI capabilities into measurable improvements. The enterprises that develop this ability will be better positioned to benefit from advances in models, agents, and automation without repeatedly rebuilding their operations around each new technology.