The Strategic Imperative and the Strategic Risk
Generative AI has created a dual pressure on business leaders that is genuinely unusual in the history of enterprise technology. On one side, there is a legitimate fear of missing a transformative capability that competitors might exploit first. On the other, there is a very real risk of wasting significant capital on initiatives that produce impressive demonstrations but no meaningful business impact. Both risks are real, and navigating between them requires a framework that is more disciplined than the breathless enthusiasm that dominates most conversations about generative AI.
At Insightrix, we have worked with organisations across financial services, professional services, manufacturing, and the public sector on their generative AI strategies. The pattern we observe is remarkably consistent: organisations that approach generative AI with clear evaluation criteria, realistic expectations, and disciplined implementation methodology extract genuine value. Those that approach it with excitement, board-level pressure to do something with AI, and a vague mandate to explore and innovate almost always end up with a collection of pilots that never scale and a growing sense of disillusionment.
This article provides the evaluation framework we use with our clients. It is designed to help you identify where generative AI will create genuine, measurable value in your organisation and where it will not — regardless of what the technology vendors and consulting firms are telling you.
Research from multiple sources suggests that fewer than twenty per cent of generative AI pilot projects progress to production deployment. The gap between experimentation and production is not primarily technical; it is strategic. Organisations that fail to define clear success criteria before starting pilots almost never manage to define them afterwards.
Understanding the GenAI Hype Cycle
The generative AI hype cycle has followed a predictable but accelerated pattern. The initial wave of excitement, driven by the public availability of large language models, produced a flood of vendor claims, media coverage, and boardroom enthusiasm. Every software vendor rebranded existing features as AI-powered. Every consultancy published a generative AI strategy framework. Every conference keynote featured a demonstration of a chatbot doing something that looked impressive in a controlled environment.
Where We Are Now
As of late 2025, we are firmly in the transition from peak inflated expectations to the trough of disillusionment — and this is actually healthy. Organisations that invested early in generative AI are now confronting the operational realities: hallucination rates that are unacceptable for regulated industries, integration complexity that far exceeds initial estimates, data privacy concerns that legal and compliance teams are only now fully appreciating, and costs that scale in ways that were not anticipated when the pilot was approved.
This is not an argument against generative AI. It is an argument for approaching it with the same rigour that any significant technology investment deserves. The organisations that will extract the most value from generative AI are those that learn the right lessons from the current disillusionment phase and apply them to a more disciplined second wave of adoption.
The Productivity Paradox
One of the most important patterns to understand is the productivity paradox of generative AI. Many organisations report that their employees are using generative AI tools extensively, yet the expected productivity gains are not appearing in operational metrics. The explanation is straightforward: individual productivity gains are being offset by new forms of waste. Employees spend time crafting prompts, reviewing and correcting AI-generated output, and navigating the uncertainty of whether the output is reliable enough to use. In some cases, the total time spent on a task increases because the employee now performs the task and reviews the AI's attempt at the task.
This is not a permanent condition, but it does mean that the path from adoption to productivity is longer and more complex than the headline figures suggest. Organisations that measure only adoption rates — how many employees are using AI tools — without measuring actual productivity impact are deceiving themselves about the value they are extracting.
A Framework for Evaluating GenAI Use Cases
We evaluate generative AI use cases across five dimensions. A use case must score well on all five to justify investment; weakness on any single dimension is typically sufficient to predict failure.
1. Error Tolerance
The single most important question for any generative AI use case is: what happens when the AI gets it wrong? Generative AI models hallucinate. They produce confident, plausible, and entirely incorrect output. For use cases where errors are easily detected and cheaply corrected — drafting initial marketing copy, summarising internal meeting notes, generating code that will be reviewed and tested — this is manageable. For use cases where errors are difficult to detect, expensive to correct, or potentially harmful — legal advice, medical information, financial calculations, regulatory submissions — the hallucination problem is a fundamental constraint that no amount of prompt engineering will eliminate.
2. Data Availability and Quality
Generative AI performs best when it can be grounded in high-quality, well-structured organisational data. Retrieval-augmented generation (RAG) architectures that connect language models to curated knowledge bases consistently outperform models operating on their training data alone. The question is whether your organisation has the data infrastructure to support this: well-maintained knowledge repositories, structured databases, clean document archives, and the data engineering capability to keep these current.
3. Process Maturity
Automating a process that is poorly understood, inconsistently executed, or frequently changed is a recipe for failure, regardless of the technology used. Generative AI is most effective when applied to processes that are well-defined, consistently executed, and stable enough to train and evaluate against. If your organisation cannot describe the process clearly enough for a new employee to follow, it cannot describe it clearly enough for an AI system to perform.
4. Human-in-the-Loop Feasibility
For most enterprise use cases, generative AI should augment human decision-making rather than replace it entirely. This requires a human-in-the-loop workflow where AI output is reviewed, validated, and approved by a qualified person before it reaches its intended audience. The feasibility of this depends on the volume of output, the expertise required for review, and the time constraints of the process. If the volume is too high for meaningful human review, or if the reviewers lack the expertise to evaluate the AI's output critically, the human-in-the-loop model breaks down.
5. Measurable Business Impact
Every generative AI use case should be evaluated against a specific, measurable business metric: cost per transaction, time to completion, error rate, customer satisfaction score, revenue per employee, or similar. If you cannot define the metric before you start the pilot, you will not be able to demonstrate value after the pilot is complete. Vague benefits like improved efficiency or enhanced creativity are not sufficient justifications for enterprise-scale investment.
Do not evaluate generative AI use cases in isolation. The total cost of ownership includes model API costs, data infrastructure, integration development, ongoing prompt maintenance, human review time, error correction, and governance overhead. Many use cases that appear economically attractive at pilot scale become prohibitively expensive at production volume.
High-Value Use Cases That Deliver
Based on our experience across multiple sectors, the following categories of generative AI use cases consistently deliver measurable returns when implemented with appropriate governance and infrastructure.
Internal Knowledge Management
Large organisations accumulate vast quantities of internal documentation: policies, procedures, technical specifications, project reports, meeting minutes, and institutional knowledge that exists only in the heads of long-serving employees. Generative AI-powered knowledge management systems that allow employees to query this documentation in natural language and receive accurate, sourced answers consistently deliver significant productivity gains. The key is robust RAG architecture with high-quality source documents and clear citation mechanisms that allow users to verify answers.
Customer Communication Drafting
Generating first drafts of customer communications — email responses, proposal sections, report summaries, claims correspondence — is a strong use case because it satisfies all five evaluation criteria. The error tolerance is high (drafts are reviewed before sending), data is available (historical correspondence provides training examples), the process is well-defined, human review is natural and efficient, and the time savings are directly measurable.
Code Generation and Developer Productivity
Software development is one of the most mature and well-evidenced generative AI use cases. AI coding assistants demonstrably accelerate development for routine tasks such as writing boilerplate code, generating unit tests, documenting functions, and translating between programming languages. The evidence suggests productivity gains of fifteen to thirty per cent for experienced developers working on well-defined tasks, with smaller or negative gains for complex architectural work or unfamiliar codebases.
Data Analysis and Reporting
Generative AI systems that can query databases using natural language, generate analytical summaries, and produce formatted reports reduce the time that analysts spend on routine reporting tasks. This is particularly valuable in organisations where a large proportion of analytical work involves recurring reports that follow established templates and draw on structured data sources.
Low-Value Traps to Avoid
Equally important is identifying the use cases that consistently fail to deliver value at enterprise scale, despite their appeal in demonstrations and strategy presentations.
Customer-Facing Chatbots in Regulated Industries
Deploying generative AI chatbots as customer-facing interfaces in financial services, healthcare, legal services, or insurance carries disproportionate risk relative to the value delivered. The hallucination problem means that every response carries the potential for providing incorrect information that could harm customers and create regulatory liability. The governance overhead required to mitigate this risk — content filtering, response validation, continuous monitoring — frequently exceeds the cost of the human agents the chatbot was intended to replace.
Strategic Decision Support
Using generative AI to inform strategic decisions — market entry analysis, competitive positioning, investment evaluation — sounds compelling but is fundamentally constrained by the model's inability to reason about novel situations, its tendency to produce confident but unsupported conclusions, and the impossibility of verifying the reasoning behind its recommendations. Strategic decisions require judgement, contextual understanding, and accountability that current generative AI models cannot provide.
Creative Content at Scale
The promise of AI-generated marketing content, social media posts, and creative assets at scale has attracted enormous investment but delivered disappointing results for most organisations. The output is competent but generic, lacking the brand voice, cultural sensitivity, and creative insight that effective marketing requires. Organisations that have scaled AI content generation frequently report declining engagement metrics and brand dilution.
Build vs Buy: The Strategic Decision
One of the most consequential decisions in any generative AI strategy is whether to build custom solutions or adopt commercial products. The answer depends on the use case, the organisation's technical capability, and the strategic importance of the AI capability to competitive differentiation.
When to Buy
Commercial products are the right choice for use cases that are common across industries and do not require deep customisation: coding assistants, meeting transcription and summarisation, document search, and basic content drafting. These are commodity capabilities where the vendor's scale advantage in model training and product development far exceeds what an individual organisation can achieve.
When to Build
Custom development is justified when the use case requires deep integration with proprietary data, when the AI capability is a source of competitive differentiation, or when regulatory requirements demand a level of control over the model's behaviour that commercial products cannot provide. Financial services firms building AI-powered advisory tools, pharmaceutical companies developing AI-assisted drug discovery platforms, and legal firms creating AI-powered contract analysis systems are examples where custom development delivers strategic value that justifies the higher investment.
The most effective approach for most organisations is a hybrid strategy: buy commercial products for commodity use cases, build custom solutions for differentiated capabilities, and invest heavily in the data infrastructure and integration architecture that supports both. The data layer is the true strategic asset, not the AI models themselves.
Governance and Risk Management
Generative AI governance is not optional, and it is not something that can be bolted on after deployment. The risks are substantive and well-documented: hallucination-driven errors, intellectual property infringement, data privacy violations, bias amplification, and reputational damage from AI-generated content that is offensive, inaccurate, or inappropriate.
Establishing an AI Governance Framework
Every organisation deploying generative AI at scale needs a governance framework that addresses at minimum: acceptable use policies that define what generative AI may and may not be used for; data governance policies that specify what data may be shared with AI systems; quality assurance processes that define how AI-generated output is reviewed and validated; incident response procedures for when AI systems produce harmful or incorrect output; and monitoring systems that track model performance, usage patterns, and risk indicators.
Intellectual Property Considerations
The intellectual property implications of generative AI remain legally unsettled in most jurisdictions. Organisations must consider whether AI-generated output is protected by copyright, whether training on copyrighted material creates liability, and whether using AI tools that were trained on competitors' data creates competitive intelligence risks. Until the legal framework matures, a conservative approach to IP risk is advisable: treat AI-generated output as a starting point that requires human modification and review, not as a finished product.
Regulatory Compliance
The EU AI Act, the UK's sector-specific regulatory frameworks, and emerging regulations in other jurisdictions all have implications for generative AI deployment. Organisations that build governance frameworks aligned with the most stringent applicable regulations will be best positioned to adapt as the regulatory landscape continues to evolve. The cost of retrofitting governance into an ungoverned AI deployment far exceeds the cost of building it in from the outset.
Building a Strategic Roadmap
A credible generative AI strategy is not a list of use cases. It is a phased roadmap that builds capability incrementally, demonstrates value at each stage, and maintains the flexibility to adapt as the technology and the regulatory landscape evolve.
Phase One: Foundation (Months 1-3)
Establish governance frameworks, acceptable use policies, and data infrastructure. Deploy commercial tools for low-risk internal use cases: coding assistants, meeting summarisation, internal knowledge search. Measure adoption and productivity impact rigorously. This phase builds organisational familiarity with generative AI whilst keeping risk contained.
Phase Two: Targeted Value (Months 4-9)
Identify three to five high-value use cases using the evaluation framework described above. Build or configure solutions for these use cases with full governance and monitoring infrastructure. Run controlled pilots with defined success criteria and measurement frameworks. Scale the use cases that demonstrate clear value; terminate those that do not.
Phase Three: Strategic Integration (Months 10-18)
Integrate proven generative AI capabilities into core business processes and customer-facing workflows. Invest in custom development for differentiated capabilities. Build the internal expertise to maintain, monitor, and evolve AI systems over time. Establish continuous improvement processes that refine AI performance based on operational data and user feedback.
The organisations that will extract the most value from generative AI are not those that move fastest. They are those that move most deliberately — with clear criteria for what constitutes value, rigorous governance that manages risk, and the discipline to stop investing in use cases that do not deliver measurable returns. Speed without direction is just expensive chaos.
Aru Bhardwaj, Founder — Insightrix