The Hype Cycle Has Ended. The Work Has Begun.
In 2023, enterprise generative AI conversations were dominated by possibility. In 2024, by pilots. By 2025, the organisations that moved beyond pilots were separating into two distinct populations: those that had built sustainable, measurable value from generative AI, and those that had accumulated a collection of underutilised tools, frustrated users, and governance debt.
The difference between the two populations is not technical sophistication. It is not budget. It is not the models they chose. It is whether they treated generative AI as a business transformation problem — requiring process analysis, measurement, change management, and governance — or as a technology deployment problem, solvable by installing the right software. Building a durable AI strategy and governance capability is what separates the organisations sustaining real results from those accumulating stalled pilots.
This is a 2026 state-of-the-practice guide. We cover what has worked, with evidence. What has disappointed, and why. What governance requirements look like at scale. How to think about model selection. And a five-step adoption framework derived from what separates the first population from the second.
The Real Adoption Numbers
Enterprise generative AI adoption is deep but uneven. As of early 2026, most large organisations in financial services, professional services, healthcare, and government have deployed at least one generative AI capability in production. Fewer than a third have deployed more than three capabilities with measured business outcomes. Fewer than one in five have a coherent enterprise AI strategy that connects individual deployments to organisational priorities.
The pattern this creates is identifiable: organisations with pockets of genuine success in isolated domains, surrounded by a much larger number of pilots and partial deployments that never reached production or never demonstrated measurable value. The governance and change management infrastructure needed to scale is typically absent even where individual use cases work.
This does not mean generative AI is overhyped in the sense of failing to create real value — it does create real value. It means the implementation discipline required to realise that value at enterprise scale is harder and different than most organisations anticipated.
Use Cases with Proven ROI
These categories have produced consistent, measurable value across multiple organisations and sectors. The ROI is not theoretical — it has been validated in production deployments with measured baselines.
Content Operations
Generative AI has materially changed the economics of content production for organisations with significant content output: financial services firms producing regulatory filings, client communications, and research; government agencies producing public communications and policy documentation; healthcare organisations producing patient communications and clinical documentation; legal firms producing client memos and transaction documents.
The proven model is not fully autonomous content generation — it is AI-assisted production with human review and editing. AI drafts from structured inputs, applying organisational style, regulatory requirements, and relevant precedent. Human editors review, refine, and approve. The result is 40–60% reduction in time-to-draft for content that previously required a blank-page start, with consistent quality maintained through the review process.
What makes it work: structured input templates, clear quality standards, editor training on effective review of AI output, and governance for content that requires disclosure or attribution.
Code Development and Review
The productivity impact of AI coding assistance on software development teams has been the most consistently documented generative AI ROI across industries. Experienced engineers using AI coding tools show 25–40% reduction in time for routine development tasks — code completion, test generation, documentation, refactoring — freeing them for more complex design and architectural work.
Code review assistance — tools that flag security vulnerabilities, performance issues, and compliance with coding standards — provides additional value independent of generation: consistent review quality without reviewer fatigue.
What makes it work: integration into existing developer workflows rather than parallel tools, training on prompt techniques that get better results, and clear standards for what human review must cover that AI review does not.
Customer and Stakeholder Service
AI-assisted and AI-first customer service has produced measurable improvements in response time and quality for organisations with high-volume, structured service contexts. The key word is structured: generative AI performs well in service contexts where the range of queries is known, the underlying knowledge base is maintained, and the consequences of errors are manageable with appropriate escalation.
First-contact resolution rates improve when AI can accurately answer more queries without escalation. Resolution time falls when agents have AI-drafted response suggestions rather than starting from scratch. Customer satisfaction holds or improves when AI handles routine queries well and escalates complex ones appropriately.
What makes it work: a maintained, authoritative knowledge base; clear escalation logic; agent training; and honest expectation-setting with customers about when they're interacting with AI.
Document Processing and Intelligence
Extracting structured information from unstructured documents — contracts, clinical notes, regulatory filings, financial statements, correspondence — at scale is a proven generative AI use case. The AI reads documents, extracts defined data points, identifies clauses or terms of interest, flags anomalies, and summarises key information.
The ROI is strongest where document volume is high, documents are variable in format, and the cost of manual processing is significant. Legal due diligence, contract management, clinical documentation, regulatory correspondence handling, and financial analysis are all in this category.
What makes it work: clear extraction templates, validation against structured sources where possible, human review for high-stakes decisions, and ongoing accuracy monitoring.
Use Cases That Have Disappointed
These categories have attracted significant investment and produced systematically weaker results than initially projected. Understanding why matters more than the list itself.
Fully Autonomous Decision-Making
The promise of AI systems that autonomously make operational decisions — approve loans, determine benefits eligibility, set prices, allocate resources — without meaningful human oversight remains technically possible in narrow, well-defined contexts and operationally premature for most enterprise applications.
The disappointment here is not that AI cannot make accurate decisions in some contexts — it can. The disappointment is that the governance, explainability, and accountability requirements for autonomous AI decision-making in regulated industries are substantially more complex than most organisations anticipated. Regulatory frameworks in Canada, the EU, and increasingly the US require human oversight at decision points with meaningful consequences. Building the infrastructure to comply with these requirements while maintaining the efficiency benefits of automation requires careful architecture — and that architecture is still being developed across most sectors.
General-Purpose AI Assistants Without Workflow Integration
Deploying a general-purpose AI assistant — a standalone chat interface with access to a model — and expecting employees to spontaneously integrate it into their work at scale has produced consistently weak results. Usage concentrates among a small population of early adopters; most employees find intermittent use of a disconnected tool less valuable than a focused assistant embedded in the workflow they already use.
The success cases for AI assistant deployment are specific and integrated: an AI assistant embedded in the document management system, surfacing relevant precedents while a lawyer drafts. An AI assistant in the claims management platform, presenting relevant policy language while an adjuster reviews a claim. Integration drives adoption; adoption drives value.
Customer-Facing AI Without Adequate Knowledge Infrastructure
Deploying customer-facing generative AI without a maintained, accurate, well-structured knowledge base produces a specific failure mode: the AI confidently answers questions incorrectly, citing non-existent policies, outdated pricing, or information from the wrong product version. The customer experience damage from this pattern is significant and disproportionate to the operational savings the deployment was intended to generate.
The knowledge infrastructure requirement for reliable customer-facing AI is more substantial than most organisations plan for. It is not a one-time build — it is an ongoing operational commitment.
Governance Requirements at Scale
Organisations that have scaled generative AI beyond isolated use cases have built governance infrastructure that organisations just starting their journeys should plan for from the beginning.
Model risk management. For any AI system making or informing decisions with material consequences, a model risk management framework adapted from financial services practice is appropriate: formal model validation before deployment, documented performance expectations, ongoing monitoring against those expectations, and a defined process for model updates and version control.
Data governance. Generative AI systems consume data. The governance of what data is used to train or fine-tune models, what data is passed to external model APIs, what data is retained in model context, and what data protection obligations apply is a non-negotiable governance requirement. In Canada, PIPEDA and provincial privacy legislation constrain the use of personal information in AI systems. In the EU, GDPR requirements interact with the EU AI Act's risk classification framework.
Explainability and audit. For regulated applications, the ability to explain why an AI system produced a particular output is an increasingly hard regulatory requirement, not a nice-to-have. Build explainability into system design — logging inputs, intermediate reasoning where accessible, and outputs — from the beginning. Retrofitting audit capability to deployed systems is substantially harder and more expensive.
Human oversight design. The design of where and how humans review, override, or approve AI outputs is a governance decision, not a UX detail. It determines your liability exposure, your compliance posture, and your ability to maintain accuracy over time. It should be reviewed by legal, compliance, and risk functions — not decided by an implementation team.
Model Selection: OpenAI vs. Anthropic vs. Google vs. Open Source
The model selection decision has become both simpler and more complex in 2026. Simpler because the leading closed-source models from OpenAI (GPT-4 family), Anthropic (Claude family), and Google (Gemini family) are all highly capable for most enterprise use cases, and the choice between them is less consequential than it was two years ago. More complex because the open-source alternatives — Meta's Llama family, Mistral, and others — have become genuinely competitive for specific use cases, changing the build/buy calculus significantly.
| Consideration | Implication for Model Selection | |---|---| | Data sovereignty and privacy | On-premise or private cloud deployment may favour open-source models | | Regulatory compliance (EU AI Act, US Executive Order) | Closed-source frontier models have clearer documented testing; open-source requires more internal validation | | Cost at scale | Open-source with self-hosting is significantly cheaper at high volume; closed-source APIs are cheaper at low volume | | Specific language requirements | Test performance in your specific language and domain before committing | | Multimodal requirements | Frontier models maintain advantage over open-source for combined text/image/audio tasks | | Customisation needs | Open-source allows fine-tuning on proprietary data without data leaving your environment |
The decision framework we recommend: start with the specific use case and its performance, privacy, cost, and governance requirements; test the leading candidates against real examples from your domain; and make the selection on evidence, not on brand or marketing claims about benchmark performance.
A Five-Step Adoption Framework
This framework reflects what the organisations with sustainable generative AI programmes have in common. It is not a sequence of technology deployments — it is a sequence of organisational capabilities.
Step 1: Inventory and prioritise use cases. Conduct a structured assessment of AI opportunities across the organisation, evaluate each against a consistent framework (value, feasibility, risk, urgency), and produce a prioritised portfolio. Do not start with a technology selection — start with the use case portfolio. The technology follows from the use cases, not the reverse.
Step 2: Build the data and integration foundations. Identify the data assets that your priority use cases require, assess their quality and accessibility, and build or repair the foundations before building AI applications on top of them. AI systems cannot compensate for missing or poor-quality data, and discovering data problems during deployment is expensive.
Step 3: Deploy a high-value, high-visibility first use case. Choose a first deployment that is genuinely valuable, demonstrably successful against measurable criteria, and visible to the stakeholders whose support you need. Nothing builds organisational confidence in AI investment like a specific, credible success story. Nothing erodes it faster than a deployment that overpromised and underdelivered.
Step 4: Establish governance infrastructure. Build the model risk management, data governance, explainability, and human oversight frameworks that will govern your AI programme — before you need them for your fifth deployment, not after. Retrofitting governance is significantly harder than designing it in from the beginning.
Step 5: Scale with a centre of excellence model. Create an internal function — a centre of excellence, an AI programme office, or an equivalent structure — that owns the AI strategy, maintains the governance framework, supports individual use case teams, and manages the knowledge base of what works and what doesn't across the organisation. Distributed AI deployment without a coordinating function produces fragmentation, inconsistent quality, and governance gaps. Agentic AI systems typically require this kind of governance infrastructure before they can operate safely at scale.
Remolda's Enterprise AI Practice
Remolda works with enterprise clients on AI strategy and governance, agentic systems, and integration across regulated industries. Our engagements are grounded in what demonstrably works, not in what generates vendor revenue.
Our clients include financial institutions, healthcare systems, government agencies, and professional services firms across Canada and internationally. We work in finance, healthcare, legal services, and government — sectors where governance and measurable outcomes are non-negotiable requirements.
Where Does Your Organisation Stand?
If you're trying to determine whether your current generative AI programme is on the right track, or where to invest next, we'd welcome a candid conversation — no pitch, just an honest assessment.
Contact Remolda to speak with our enterprise AI practice.