B(I)ack to the (AI) Future

By Assaf Araki

History repeats itself: a high failure rate from BI to Agents

The New AI Services Land Grab 

In May 2026, both OpenAI and Anthropic moved beyond building AI models and into the professional services arena. OpenAI launched OpenAI Deployment Co. (DeployCo), backed by over $4 billion in funding from a consortium that includes TPG, Capgemini, Bain & Co., McKinsey, Goldman Sachs, SoftBank, and Warburg Pincus, anchored by the acquisition of London-based AI consulting firm Tomoro. Around the same time, Anthropic established its own AI services company—backed by Blackstone, Hellman & Friedman, and Goldman Sachs—focused on helping mid-size enterprises adopt Claude across their operations. Microsoft recently launched the Microsoft Frontier Company, a $2.5 billion initiative that deploys 6,000 embedded Forward Deployed Engineers (FDEs) directly into customer organizations to build, deploy, and optimize AI applications and autonomous agents on-site.

These moves signal a broader strategic shift: AI is no longer content to sell API access and leave implementation to third parties. By building in-house services arms staffed with consultants and engineers, OpenAI and Anthropic are directly competing with major system integrators like Accenture and Deloitte, positioning themselves to own not just the technology layer but the deployment and transformation layer of enterprise AI adoption.

The consulting industry's pivot to AI services has become its most significant growth driver in over a decade. Accenture leads the pack with $5.9 billion in generative AI bookings in FY2025, nearly doubling year-over-year, with AI revenue tripling to $2.7 billion across more than 1,300 clients and 11,000 active projects (Accenture FY2025). BCG disclosed that 25% of its $14.4 billion in 2025 revenue, roughly $3.6 billion, came directly from AI work (BCG, Apr. 2026), while McKinsey now reports that approximately 40% of all its projects carry an AI component. Among the Big Four, EY saw AI-related revenue grow 30% in its FY2025, with more than 15,000 professionals working on AI-led client engagements (EY, Oct. 2025), and KPMG India reported AI-related work growing nearly 20x in just 14 months. 

The AI services channel has emerged as one of the fastest-growing segments in enterprise technology, with Omdia forecasting a $267 billion partner opportunity by 2030 (Omdia, Sep. 2025)—a figure that includes the rapidly expanding agentic AI layer. The economics of the channel are shifting fundamentally: partner profitability is moving away from resale margin and toward high-value services such as data preparation, model integration, workflow automation, and ongoing optimization. Vendors using AI-powered partner enablement are reporting double-digit increases in partner-sold revenue. Meanwhile, cloud hyperscalers are amplifying the channel multiplier effect. AWS now generates $7.13 in partner services revenue for every $1 of its own product revenue (AWS/Omdia, 2025), underscoring how the real economic value in AI is accruing not at the model layer but in the deployment, integration, and services layer that channel partners own.

What Business Intelligence Got Wrong

The rapid rise of AI services today carries an unmistakable echo of the business intelligence boom of the 1990s, when a wave of consulting firms, system integrators, and niche specialists rushed to help enterprises make sense of their data through tools like Cognos, Business Objects, and early SAP deployments. The promise was transformative: finally, decisions driven by data rather than instinct. Yet the majority of Business Intelligence (BI) implementations failed to deliver lasting value, not because the technology was flawed, but because organizations underestimated the change management, data quality, and organizational alignment required to make it stick. 

The consultants who thrived were not those who sold the most licenses or deployed the fastest, but those who embedded themselves deeply enough in a client's operations to drive genuine adoption. AI services today face the same crucible. The technology is real, the demand is genuine, and the market numbers are extraordinary, but the graveyard of failed BI programs reminds us that complexity, resistance to change, and the gap between a working demo and enterprise-scale deployment are what ultimately determine winners and losers. The consulting firms that carry those lessons forward—prioritizing outcomes over implementation speed, and adoption over deployment, are the ones most likely to build durable practices rather than ride a wave that eventually breaks.

The obstacles that slowed enterprise data analytics adoption twenty years ago stemmed from a set of recurring and deeply structural challenges: - 

• Misaligned Business Goals: Projects frequently start with a cool technological concept rather than an actual business problem, rendering the final insights irrelevant or unactionable to corporate stakeholders.

• Legacy Process Ignorance: Deploying algorithms or dashboards without redesigning the underlying legacy business processes causes friction. Teams continue to use old workflows alongside new tools, leading to low adoption rates.

• Siloed Data & Poor Quality: When data pipelines are disconnected or feed on inaccurate, outdated, and biased data, "garbage in, garbage out" scenarios break trust in the results.

• Communication Gaps: A disconnect between technical data scientists and corporate decision-makers creates translation errors, meaning users can't interpret the models and decision-making gets delayed.

• Scope Creep and Large-Scale Implementations: Instead of testing data architectures on specific operational workflows, companies attempted massive, enterprise-wide overhauls. Without a minimized, disciplined scope, projects quickly ran out of budget and "started to gather dust".

Twenty years ago, organizations rushed to invest massive budgets into data architecture. However, Gartner's own numbers on BI failure shifted over the years but never got optimistic: a 60% failure estimate at one point, with later commentary putting the range as high as 70-80%. Whichever figure you use, more than half of BI and data warehouse implementations failed to deliver their intended value (Gartner, via Enterprise Apps Today). Gartner also noted that more than half of all analytics projects failed to be completed on time, within budget, or with the features originally promised. The literature from twenty years ago consistently pointed out that these failures were rarely caused by flawed technology. Instead, they stemmed from a lack of strategic alignment and a failure to adjust business processes to match the new data capabilities.

Is Agentic AI Repeating It?

There are early reports, like MIT’s 2025 NANDA research, that report very high failure rates and claim that enterprise AI implementations are experiencing a catastrophic failure epidemic, with 95% of generative AI pilots failing to deliver measurable financial returns (MIT NANDA, 2025). This represents billions in wasted investment across a $500 billion global AI market, with 42% of companies now abandoning most AI initiatives. The pressure is not easing: Gartner projects that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Gartner, June 2025). The crisis stems not from technological limitations but from fundamental organizational and strategic execution failures. However, the 5% of organizations achieving success provide a clear roadmap for transformation. We identify several obstacles that will slow AI adoption; not all are known, but we already identify some challenges: 

• The “Science Experiment” trap - The introduction of ChatGPT in November 2022 disrupted the AI market, proving that you can build an application on top of an LLM. That drove most of the developers to start experimenting with LLMs, which creates the science experiment trap: projects get stuck in "pilot paralysis." They work well in forgiving, controlled lab environments but buckle under the reality of messy, real-world operational demands.

• The “Mercedes in the Parking lot” - The intense top‑down pressure from management teams and boards to “implement AI,” often faster than organizations are structurally or culturally ready for, creates a dynamic where leaders want visible, impressive AI projects they can point to, sometimes more for bragging rights than for genuine business impact.

• Weak Data Foundations - When data is fragmented across systems, poorly governed, and low quality, even strong models produce unreliable results. Rather than crashing like traditional software, AI generates confidently inaccurate outputs (hallucinations), resulting in flawed decision-making and corrupted business logic.

• Cultural and Workflow Changes - Enterprises assume AI success is 100% technical, allocating disproportionate budgets to algorithms rather than user enablement. Successful adoption hinges on workflow redesign and training. When employees aren't brought along in the transformation, they resist adoption, rely on manual workarounds, or worry about job displacement.

• Monitor, governance, and control - Enterprises don't want to give unconstrained autonomy to AI systems without human-in-the-loop oversight or basic sanity checks.  They are already aware that lack of monitoring causes model degradation over time. More disastrously, it results in high-profile PR and legal nightmares. Building a robust control platform for AI is not easy and involves many checks and balances, it delays the move to production. 

What the 5% Do Differently

Not every organization is stuck. MIT's own research on the GenAI Divide found that the 5% of companies extracting real value share a consistent pattern (MIT NANDA, 2025): they buy or partner rather than build everything in-house, they target back-office and operational friction instead of flashy front-office pilots, and they measure success by whether a workflow actually changed, not by how many licenses got adopted. The same pattern shows up if you flip the obstacle list above on its head.

Where the Science Experiment trap keeps AI in the lab, the 5% pick one painful, well-scoped workflow and take it to production before touting a second one. Where the Mercedes in the Parking Lot chases visibility, the 5% fund projects a line manager asked for, not one a board member wants to announce. Where weak data foundations produce confident hallucinations, the 5% treat data readiness as a prerequisite, not a parallel track. Where culture gets bolted on as an afterthought, the 5% budget for workflow redesign and training alongside the technology itself. And where monitoring gets skipped to hit a launch date, the 5% build human-in-the-loop checkpoints from day one, even if it slows them down.

None of this is exotic. It's the same discipline that separated the BI programs that stuck from the ones gathering dust twenty years ago: narrow scope, real business sponsorship, and patience with the unglamorous data work underneath.

Why the Pattern Keeps Repeating

When the cloud was introduced, software became a service and forced process alignment among customers rather than customization, primarily because its architecture is built on shared, multi-tenant infrastructure that demands a high degree of standardization to ensure stability, faster updates, lower maintenance complexity, and scalable performance across all customers. To protect this model, most vendors deliberately draw a line between configuration—adjusting settings, workflows, and templates within predefined limits—and true customization, which would mean changing core logic or system behavior, and they intentionally restrict the latter as a permanent constraint of the delivery model rather than a temporary limitation.

SaaS created alignment in process among organizations, but it did not create alignment among organizations in how they represent data. BI and AI have always been connected, going back to the 1970s when BI maturity first began to evolve, giving rise to what became known as the BI maturity curve. This framework assesses how effectively an organization collects, manages, and transforms its data into actionable, strategic decisions. As an organization progresses along the curve from bottom left to top right, the sophistication required to implement the solution increases, but so does the business value it delivers, outlining a progression from basic reporting to advanced dashboards, and eventually to machine learning that enables prediction and automated, data-driven operations.

The data misalignment we saw in the early 2000s BI era, which drove high failure rates in BI projects, is the same misalignment that today creates a gap between a pre-trained model and an organization's actual data, requiring fine-tuning and harnessing to that organization's context. Different organizations may be able to align their processes, but they will not be able to align their underlying data representations.

The organizations closing that gap aren't waiting for data alignment to solve itself. They're doing the unglamorous work of fitting the model to their own context, one scoped workflow at a time, the same work that separated durable BI programs from the ones that got shelved. That work is slower and less announceable than a $4 billion services launch. It's also the only part of this pattern, twenty years running, that has ever actually predicted who wins.