Technology Report
|
Резюме
This article is part of Bain’s Technology Report 2026 By 2024, enterprise AI leaders had cracked the code on how to run AI transformations. They stopped treating AI as an IT experiment and instead ran business transformations with C-suite sponsorship and programmatic discipline. These leaders were rewarded with 10% to 25% EBITDA growth. However, our experience suggests that as much as 90% of enterprises today remain focused on tool deployment and narrow use cases, with little target or change-management discipline. This results in micro productivity but is mostly just “pave the goat trails” with little impact on revenue or earnings. These enterprises will eventually follow the taillights of the leading 10% in the coming years or risk being left behind. The lesson from these transformations has been consistent: The most important work lies in process redesign and modernizing the data and application environment. Companies we work with have found that for every dollar spent on technology, four are spent on people and process. And because those investments are required regardless of how the technology evolves, waiting for the dust to settle is a losing strategy. The real advantage will come from making the organization transparent to and understood by machines and employees: its systems, standards, workflows, exceptions, customer knowledge, decision rights, and institutional judgment, without making all of that information indiscriminately accessible. The evidence is difficult to ignore. Across customer service, sales and marketing, software development, operations, and the back office, well-designed transformations are delivering double-digit efficiency gains (see Figure 1).
Figure 1
Absorption enablement is a new competitive battlegroundWhile enterprises race to copy the winning business transformation playbook, the major labs and platforms are speeding up their release cycles faster than most enterprises can absorb them. That gap has turned enterprise absorption itself into a competitive variable. Recognizing this, vendors are racing to break the barrier, and that includes investing in the forward-deployed engineer (FDE) model leveraged by Palantir (see Figure 2). Anthropic has committed roughly $1.5 billion of joint ventures to accelerate enterprise and midmarket adoption. OpenAI’s DeployCo vehicle marshals more than $4 billion. Microsoft has committed $2.5 billion to set up Microsoft Frontier Company, a new operating business to embed engineering experts at customers’ locations. Amazon Web Services is investing $1 billion to set up an FDE unit, and Google Cloud is investing $750 million in its new Gemini Enterprise transformation program, all with the goal of helping customers get measurable business outcomes.
Figure 2
Notes: Revenue operations typically includes sales, marketing, and customer service; solutioning may include technical presales, forward-deployed engineers, and other delivery roles; G&A includes general and administrative roles Source: Bain analysisThe stack and the profit pool continue to evolveAs use cases scale, enterprises move beyond their first wave of innovation and begin to optimize for greater efficacy and autonomy, continuous improvement, lower token costs, and stronger data security. At the same time, open-weight models have narrowed the performance gap with closed models for many enterprise use cases, expanding choice for enterprises and intensifying competition between models (see Figure 3).
Figure 3
Note: Epoch Capabilities Index (ECI) is a composite metric aggregating results from multiple AI benchmarks into a single measure of overall capability Source: Epoch AISome observers see this growing competition and falling token prices as indicators that the modal layer is commoditizing and its profits pools disappearing. At Bain, we don’t come to the same sweeping conclusion about commoditization, though we clearly see price pressure on mature model tokens and the growth of value in other layers of the stack (infrastructure, agent platforms, context, applications) for several reasons. First, the boundary of the model will continue to evolve as capabilities in today’s harness are incorporated into the model and run natively. Second, leading model performance will remain valuable where intelligence matters most. Frontier models will tackle increasingly difficult problems that remain unsolved by humans, including disease treatments, fusion, and quantum gravity. Third, in multistep agentic workflows, even modest differences in model accuracy can compound dramatically. For example, a 97% accurate model would deliver just 36% accuracy in a 34-step process vs. 71% accuracy for a model that is 99% accurate. That seemingly trivial 2% advantage can make the difference between a viable process and an unacceptable error rate. We see the more likely scenario as a continuum of frontier and mature models. New and unproven use cases will initially favor frontier models, then likely be migrated to lower-cost alternatives as the use cases mature. Frontier providers may consolidate as they push toward more fundamental challenges, capturing value where superior intelligence matters most. Meanwhile, a rich ecosystem of lower-cost, specialized, and optimized models will efficiently serve more common tasks, providing the services and support that enterprise customers need. In short, the industry structure is far from settled, and competitors are increasingly competing across multiple fronts. We may be headed into a segmentation between frontier models and trailing models or an expanding definition of models rather than a classic commoditization pattern. Vendors are racing to sell the harnessAs this debate rages, vendors aren’t waiting; they’re building out the application and infrastructure layers. Palantir’s Alex Karp acknowledged the shift on an August 2026 earnings call when he said, “You own the weights, you own the alpha, you own everything.… We have a product that allows you to switch out models.” Microsoft’s Satya Nadella conveyed a similar sentiment on a July 2026 earnings call, saying that for customers, “[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.” These application layers act as harnesses that connect AI to the enterprise and turn model intelligence into business outcomes. Agents require context, data, memory, tools, permissions, routing, orchestration, evaluation, and learning loops. The harness determines how effectively an enterprise can put AI to work, operating over a pool of models and routing each request to the best-suited model that can meet the threshold for quality and correctness. How well an enterprise assembles this harness determines how fast it can absorb new AI capability and compound its gains. HubSpot illustrates the potential. Its AI abstraction layer allows it to evaluate and swap models without rebuilding the underlying architecture while its proprietary customer context helps agents outperform generic models. The results are tangible: HubSpot’s agents now resolve more than 70% of support tickets while its Prospecting Agent books nearly twice as many meetings as it did the previous year. The same architecture also helps HubSpot absorb new AI capabilities faster, accelerating product development. No single vendor yet owns every layer, and open protocols such as Model Context Protocol and Agent2Agent give enterprises the flexibility to compose across them, which means the harness is becoming a strategic choice, not a vendor lock-in. Enterprise AI leaders are compounding their gains with AI-native workflowsWhile vendors race to enable enterprise absorption, enterprise leaders are compounding their gains with another wave of business redesign. Having cleaned up their data and redesigned their first workflows, they are now building processes that resemble what an AI-native company would design from a blank sheet, unencumbered by legacy org charts. Consider Klarna's campaign production function: By deploying AI agents across image generation and creative development, the company compressed what had previously taken six weeks down to seven days while simultaneously increasing output volume. Weekly updates to images aligned with faster-paced retail events and created a better customer experience. The work moved from getting ready to getting better. Some companies are redesigning their product launch processes to be AI native. Instead of automating the existing roles across design, marketing, and sales, they condense the process around three senior leaders and a suite of agents that specialize in capabilities such as communication, competitive analysis, channel analysis, and sell-through analysis. Among the teams that were engaged in multi-month preparations for launch, the emphasis now moves to post-launch iteration based on continuous evaluation and improvement. The three leaders presiding over the product launch process spend less time on mundane analytics and troubleshooting and more time reviewing markdown files and iterating to improve the product launch’s effectiveness. Executives blame the tech, but organizational barriers matterLeading enterprises have charted a clear roadmap and set of best practices for deploying AI in ways that pay back the investment and deliver competitive advantage. Even so, most enterprises remain in the pilot phase of AI absorption. Asked what holds them back, they point mostly to technical and talent worries: data security, hallucinations, unproven ROI (see Figure 4). Our work with companies suggests that the real reasons for slow progress are broader and better represented by those found in a 2025 study by the Massachusetts Institute of Technology. Organizational barriers such as challenging change management, tentative executive sponsorship, poor user experience, and a simple unwillingness to adopt new tools.
Figure 4
The next era of AI is when companies begin turning the technology into a true competitive advantage. It’s defined by blank-slate redesigns of processes and organizations, and it’s powered by continuously improving hybrid human-agent workflows (see the Bain Brief “Proprietary Intelligence: How to Win with AI”). For CEOs looking to accelerate their AI programs, the starting point is clarity on the stakes and the ambition. Leaders need a rigorous view of where AI can create the most value, where it could reshape competitive advantage, and where inaction could leave the business vulnerable. From there, they can set ambitions bold enough to force genuine redesign rather than incremental improvement, concentrate leadership attention on the few domains in which AI-native ways of working could create breakaway advantage, and establish clear ownership for delivering results. The goal is not just to do more with AI but to make deliberate choices about where AI can fundamentally change the trajectory of the business. To help executive teams move in the right direction, several imperatives stand out.
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. Waiting for the dust to settle is the one strategy guaranteed to ensure it never does. More from the report
Read our Technology Report 2026 |