Most companies spent 2024–2025 treating AI as a software procurement problem. By 2026, the real divide will be operational design. Being able to log into ChatGPT, Copilot, Gemini, or a local model is not the same as reliably completing pricing, customer service, legal review, procurement, or risk workflows end-to-end. One is tool access. The other is production capability.
That is why many enterprise AI programs disappoint. The model is rarely the real bottleneck. The failure point is usually elsewhere: no business owner with accountability, no process redesign, no governance boundary. What should have become a scalable operating capability gets trapped as a cluster of departmental experiments.
My core view is simple: in 2026, enterprises will not be constrained by model availability; they will be constrained by the absence of an AI operating system. Not a single product, but a management system built on three pillars: organization (who owns the outcome), process (which work gets rewritten and where handoffs change), and governance (what data can be used, what decisions require review, and what can be audited). If your AI agenda is still dominated by training sessions, prompt workshops, and scattered SaaS pilots, you are not just slightly behind. You are building a weaker cost structure and a slower decision system.
Are you deploying tools, or rewriting work?
Gartner repeatedly warned in 2024 that more than 30% of generative AI projects would be abandoned after proof-of-concept due to unclear business value, poor data quality, or inadequate risk controls. McKinsey’s 2024 The State of AI made a similar point: many organizations are using AI in at least one function, but the companies actually seeing meaningful bottom-line impact remain a minority.
In practice, we see the same mistake again and again: enterprises treat AI as a desktop productivity layer rather than a workflow execution layer. Desktop productivity can save employee time, but it does not automatically improve unit economics. A sales team using AI to draft proposals may save 30 minutes per rep; that does not guarantee a higher win rate. But putting AI into pricing approvals, inventory forecasting, customer segmentation, and support ticket routing can improve margin and cycle time simultaneously.
A useful maturity lens is what I call the three-layer AI transformation model. Layer one is human assist: employees use AI to write, summarize, search, and analyze. Layer two is embedded workflow: AI is inserted into CRM (customer relationship management), ERP (enterprise resource planning), service, finance, and legal processes. Layer three is governed automation: under clear permissions, audit logs, and risk thresholds, AI executes parts of decisions directly. Most companies think they are in layer two. In reality, many are still in layer one.
If the organization does not change, AI budget becomes digital stationery spend
Without a business owner, AI almost always gets absorbed into IT as an infrastructure program. The result: systems are delivered, but value is not. Deloitte’s 2024 enterprise surveys on generative AI consistently identified governance ambiguity, talent gaps, and weak cross-functional coordination as top obstacles. That is not just “culture.” It is operating design.
The answer is not simply to set up a central AI Center of Excellence (CoE) and call it done. The more effective model is a dual-core structure. A central AI governance office should own policy, vendor standards, security controls, architecture principles, and regulatory interpretation. Business units, meanwhile, must provide process owners with P&L accountability for time reduction, error reduction, conversion lift, or margin improvement. If the central team only produces demos and training, while business units only submit requests, the program stalls.
This matters even more in Asia-Pacific and Greater China. Enterprises in Hong Kong, Taiwan, Singapore, and mainland-linked operations face multilingual workflows, data residency constraints, and cross-border compliance complexity. It is not just about output quality in Traditional Chinese, Simplified Chinese, and English. It is about whether data can leave the jurisdiction, whether mainland personal information rules are implicated, and whether on-premise deployment is required. If legal, security, and business teams are not jointly involved in decision-making, procurement outruns governance very quickly.
If the process is untouched, ROI will not emerge from a chat window by itself
The companies that create AI ROI are not necessarily the ones with the highest employee usage. They are the ones with the fewest unmanaged exceptions in core workflows. IBM’s 2024 CEO research and Microsoft/LinkedIn’s 2024 Work Trend Index both showed heavy executive commitment to AI, but also a gap between broad employee experimentation and institutionalized operating change. That gap is where ROI disappears.
So where should companies start? Look for three traits: high frequency, clear rules, and manageable error costs. Frontline service responses, standard contract review, procurement classification, enterprise knowledge search, and meeting summarization usually pay back faster than “fully automated strategy.” Forrester and IDC have been directionally aligned on this for several years: bounded, repeatable use cases scale better than vague creative aspirations.
The overlooked design question is the human-machine handoff. What can AI draft? At what dollar threshold must pricing be reviewed by a manager? Which contract deviations trigger legal escalation? Which customer complaints can be auto-resolved and which require a human? Without these rules, “AI automation” is often just accountability blur. The real enterprise risk is not that AI makes one mistake. It is that nobody can explain how the mistake happened.
Governance is not a brake; it is what gets AI into production
Many executives still hear “governance” and think “slower rollout.” That is backwards. Without governance, AI stays stuck in pilot mode. Stanford HAI’s 2024 AI Index showed rising enterprise interest in model capability alongside rising concern over safety, reliability, and accountability. In regulated sectors such as financial services, healthcare, telecom, and public services, large-scale deployment without auditability is simply not realistic.
At minimum, enterprises need four governance layers. First, data classification: public, internal, sensitive, and regulated. Second, model classification: public model, commercial API, private model. Third, decision classification: recommendation only, human-assisted execution, or fully automated action. Fourth, logging and traceability: who asked what, what the model returned, who approved the outcome, and when the workflow changed. This is not bureaucracy for its own sake. It is what protects the company when a customer, regulator, board member, or auditor asks how an AI-supported decision was made.
Vendor choice should also be made through this lens. The “best” model is not automatically the best enterprise fit. Data boundaries, integration capability, regional availability, and total cost matter more than benchmark headlines.
| Product / Approach | Indicative pricing model | Positioning | Strengths | Risks / Limits |
|---|---|---|---|---|
| Microsoft 365 Copilot | Around US$30/user/month (public pricing) | Productivity embedded in office suite | Strong integration with Office, Teams, Entra; manageable for enterprise IT | Limited on deep workflow redesign; ROI pressure if licensed broadly without targeted use |
| OpenAI Enterprise / API | Seat-based or token-based usage | General-purpose model capability with mature developer ecosystem | Fast to adopt; strong for knowledge assistants, service, and content generation | Costs can scale quickly with usage; governance and orchestration must be designed by the buyer |
| Google Gemini for Workspace / Vertex AI | Workspace add-on and cloud consumption | Productivity plus model-building environment | Good fit for Google-centric estates; strong search and ML integration | Higher adoption complexity in non-Google environments |
| Anthropic Claude Team / Enterprise | Seat or API pricing | Long-context enterprise assistant with strong safety positioning | Strong for document analysis and policy drafting; clear enterprise risk narrative | Narrower ecosystem and regional availability may need verification |
| Open-source model + on-prem deployment (e.g., Llama-based stack) | Higher upfront compute, integration, and operations cost | Data sovereignty and customization | Suitable for finance, manufacturing, government, and sensitive cross-border contexts | Requires MLOps (model operations), specialist talent, and disciplined governance; TCO may not be lower |
Not every company should go all-in on AI; every company does need a strategy with clear boundaries
This is not an argument against ambitious AI adoption. PwC’s 2024 estimates remained positive on AI’s long-term impact on productivity and GDP, while IDC continues to forecast strong double-digit growth in enterprise AI spending through the second half of the decade. But a fast-growing market does not mean every enterprise investment will work.
The winners will not be the companies that buy the most tools. They will be the companies that choose the right workflows, install the right governance, and assign the right ownership.
This matters especially for SMEs in Hong Kong, Taiwan, Singapore, and export-linked businesses across Greater China. Many are not large, but they operate complex cross-border supply chains, multilingual customer interactions, and fragmented back-office processes. Their most common mistake is buying three or four AI tools at once and getting none of them into formal operations. My advice is more conservative and more effective: choose one value chain, such as lead generation to quotation or customer service to renewal, and run a disciplined 90-day transformation. Fix four metrics upfront: labor time, conversion rate, error rate, and customer satisfaction. AI strategy is not a shopping list. It is a management hypothesis.
Key takeaways: in 2026, the winners will not have the most models. They will have the strongest operating discipline
If you take away one idea, let it be this: enterprise AI transformation is not fundamentally a model problem. It is the challenge of connecting organization, process, and governance into a measurable value chain. First assign accountable business ownership. Then redesign the workflow. Then use governance to make the system safe enough for production. Reverse the sequence, and you will spend heavily on pilots that look innovative but do not replicate.
A practical three-step agenda works well. First, appoint an AI business owner who is accountable for operating metrics, not just a CIO or IT manager. Second, select two or three high-frequency, rule-based workflows that can be measured within 90 days. Third, build minimum viable governance: data classification, vendor admission criteria, human review thresholds, and logging standards. If you can do these three things, AI has a chance to become an enterprise capability rather than a software experiment.
Self-check questions:
- Are we measuring tool usage, or actual workflow outcomes?
- Does every AI initiative have a business owner accountable for revenue, cost, or risk metrics?
- If a regulator, customer, or board member asked tomorrow how an AI-supported decision was made, could we produce a complete record?
