The Return of Traditional AI: Organisations Are Rethinking Their LLM-First Strategies

According to S&P Global research, forty-two percent of companies abandoned the majority of their AI initiatives this year, up from 17% in 2024. Forrester predicts generative AI will orchestrate less than 1% of core business processes in 2025 — rule-based systems and robotic process automation (RPA) will still do the heavy lifting. The World Quality Report confirms the initial rush has given way to more grounded strategy, with 60% citing hallucination and reliability concerns. The pattern is consistent: organisations applied generative AI to problems better solved by traditional machine learning or deterministic automation, and are now recalibrating.

The evidence mounts

The value gap is widening rather than closing. BCG’s October 2025 research, surveying more than 1,250 firms, found only 5% achieve AI value at scale. While 89% of organisations are piloting generative AI-augmented workflows, only 15% have achieved enterprise-scale deployment. Non-adopters increased to 11% from 4% in 2024 — organisations are now choosing not to deploy rather than deploying poorly. For Boards who have watched their organisations invest heavily in large language model (LLM) capabilities, these statistics demand attention.

In June 2024, I argued that focusing solely on generative AI risked overshadowing the diverse and equally transformative types of AI that had been quietly driving business innovation for years. In January 2025, I opened my piece on selecting an enterprise LLM with a question that many considered unfashionable at the time: “Is your use-case a good fit for generative AI?”. The data now emerging suggests this question deserves more prominence in Board discussions.

The year of recalibration

The pattern extends beyond statistics to concrete product decisions. Coverage last week in The Information revealed that Salesforce, one of the most valuable software firms, has been incorporating “deterministic” automation into Agentforce to improve reliability. “We all had more trust in the LLM a year ago,” acknowledged Sanjna Parulekar, the company’s senior vice president of product marketing. The company’s CTO for Agentforce, Muralidhar Krishnaprasad, explained the technical reality: “If you give an LLM more than, say, eight instructions, it kind of starts dropping instructions.” Salesforce’s website now states that Agentforce can help “eliminate the inherent randomness” of LLMs, “guaranteeing that your critical business workflows follow the exact same steps every single time.” This isn’t a single vendor’s challenge — it reflects a market-wide recognition that LLMs struggle with precision-dependent workflows.

The ISG State of Enterprise AI Adoption report from September 2025 captures the paradox: production deployment doubled to 31% of use cases compared to 2024, yet “expectations that AI would cut costs and boost productivity are underdelivering.” Organisations are getting better at deploying AI — but not necessarily at matching the right AI to the right problem.

Several patterns have emerged from this recalibration. The first is the proliferation of “guardrails” language across enterprise AI discussions, implicitly acknowledging LLM unpredictability in production environments. The second is growing sophistication about where generative AI adds genuine value versus where it introduces unnecessary complexity and cost. The third is the emergence of hybrid architectures that blend generative AI with rules-based automation.

The underlying reality is straightforward. LLMs excel at tasks requiring contextual understanding, ambiguity handling, and creative content generation. They struggle with precision-dependent workflows, numerical reasoning, and multi-step processes requiring consistent execution. LLM inference also costs significantly more than traditional machine learning or rules-based approaches — a difference that compounds quickly for high-volume tasks. And the inherent randomness that makes LLMs powerful for content generation makes them problematic for deterministic business processes that must execute identically thousands or millions of times daily.

This isn’t failure — it’s maturation. The market is learning to deploy generative AI strategically rather than universally. The World Quality Report’s framing is apt: organisations are moving from “initial rush” to “grounded strategy.”

The right tool for the problem

The capability matching question I originally posed — “Is your use-case a good fit for generative AI?” — is one that more organisations are now asking retrospectively. Tasks like ticket categorisation, where fixed criteria apply, often achieve faster and more accurate results with straightforward rules engines. Fraud detection remains better suited to traditional machine learning models trained on historical transaction patterns. The industry data suggests these distinctions matter more than the initial enthusiasm acknowledged.

So, when does traditional AI or deterministic automation outperform LLMs? Classification tasks with fixed criteria represent the clearest case. Customer support ticket routing, document categorisation, and approval workflows benefit from the speed, cost efficiency, and consistency that rules engines or traditional machine learning provide. These systems don’t hallucinate. They don’t produce variable outputs for identical inputs. They don’t require the governance overhead of managing probabilistic responses.

Pattern recognition in structured data presents another clear domain. Fraud detection, anomaly identification, and predictive maintenance all rely on supervised learning models trained on historical patterns. These models deliver higher accuracy at lower cost than LLMs for these specific tasks — and provide the deterministic reliability that risk functions require.

Precision-dependent calculations — financial modelling, actuarial analysis, inventory optimisation — remain problematic for LLMs. Numerical reasoning continues to be an acknowledged weakness. For high-volume processes requiring consistent execution, the Forrester finding that rules and RPA “remain the primary orchestrators of mission-critical flow” validates what operational leaders have discovered through experience.

Conversely, generative AI genuinely excels when contextual understanding matters more than precision. Document summarisation, semantic search, and content analysis represent natural LLM territory, as does creative generation for marketing content, design ideation, and code suggestions where variation adds value rather than introducing risk. Customer interactions requiring interpretation and conversational flexibility play to generative AI’s strengths, as does knowledge synthesis that combines information across sources for understanding rather than calculation.

The sophisticated approach emerging from industry analysis is hybrid architecture: deterministic systems handling execution while generative AI provides contextual intelligence. Previously, I explored how the most powerful agentic systems orchestrate multiple disciplines — computer vision, natural language processing, reinforcement learning, traditional machine learning — creating compound capabilities that exceed any single approach. The organisations capturing greatest value aren’t those deploying LLMs universally, but those developing nuanced understanding of where different AI capabilities add genuine advantage.

The Board’s evaluation framework

Armed with this understanding of capability matching, Boards can apply structured evaluation to every AI initiative seeking approval. The Complete AI Framework provides systematic guidance, but three diagnostic categories prove particularly valuable for distinguishing appropriate generative AI applications from expensive mismatches.

Problem domain assessment asks fundamental questions about the nature of the work. Tasks requiring consistent execution, reliable repetition, and adherence to defined rules rarely benefit from generative AI’s strengths in contextual understanding and creative variation. When the answers point toward precision over flexibility, generative AI often represents an expensive solution to a problem that simpler approaches could address more effectively.

Capability matching requires honest comparison. Before approving a generative AI initiative, Boards should understand what the solution would look like using traditional machine learning, rules engines, or existing automation — and what specific capability generative AI adds that alternatives cannot provide. Too often, organisations default to generative AI because it represents the newest technology rather than the most appropriate one, without evaluating total cost of ownership including inference costs at scale.

Risk and reliability assessment addresses operational realities. With 60% of organisations citing hallucination and reliability as top concerns, Boards need clarity on how inconsistent outputs will be handled, how drift or degradation will be detected over time, and what governance mechanisms keep outputs within acceptable bounds. These questions often surface assumptions about LLM reliability that don’t survive contact with production environments.

The maturity arbitrage opportunity deserves particular attention. Computer vision, machine learning, and RPA have years of production refinement behind them and deliver immediate, proven value. Generative AI adds capability at the edges whilst these established technologies handle core operations — a portfolio approach that reduces implementation risk by not betting everything on the newest discipline.

This maturity arbitrage maps directly to the AI Stages of Adoption. Organisations at the Experimenting stage can leverage proven traditional AI to build confidence and demonstrate value, whilst reserving generative AI experiments for bounded use cases. Those at Adopting and Optimising stages benefit most from hybrid architectures — using mature capabilities for scaling whilst testing generative approaches in parallel. Organisations attempting to Transform or Scale with generative AI alone face the reliability challenges the data reveals. Those orchestrating multiple AI disciplines across maturity levels achieve more sustainable progress.

The recalibration underway is partly organisations recognising they’ve been trying to skip stages — deploying transformative technology without the foundational capabilities to support it. The ISG finding that production deployments doubled whilst value delivery underperformed suggests organisations are getting better at deploying AI but not at matching the right AI to the right problem, or sequencing adoption appropriately.

The economic logic is straightforward: reserving LLMs for tasks that genuinely require them reduces inference costs, improves production reliability, and accelerates deployment by leveraging mature capabilities.

Strategic implications

Every transformative technology follows a predictable arc: initial enthusiasm, universal application, recalibration, and finally strategic deployment. Generative AI is no exception. The World Quality Report’s observation that “the initial rush has given way to a more grounded and complex strategy” captures this transition precisely. The organisations capturing greatest value aren’t those who deployed LLMs everywhere, but those developing sophisticated understanding of where different AI capabilities add genuine advantage.

The future belongs to organisations that orchestrate multiple AI disciplines strategically. When computer vision handles inspection, machine learning manages prediction, RPA executes processes, and generative AI provides contextual intelligence, each capability operates where it excels. This orchestration — what I’ve described as compound loops coordinating complementary technologies — delivers multiplicative value that single-discipline approaches cannot match.

As AI budgets face increasing scrutiny — and with 42% of companies having scrapped initiatives this year — Boards have both opportunity and obligation to ensure investments match capabilities to problems. The question isn’t whether to invest in AI, but whether each investment deploys the right AI for the specific challenge. Organisations that develop this discrimination will outperform those still forcing generative AI into every use case.

The recalibration underway isn’t retreat from AI ambition — it’s the maturation that enables sustainable scaling. Boards asking “Do we actually need generative AI for this?” aren’t being conservative; they’re asking the question that separates strategic deployment from expensive experimentation.