Agents vs. Model Training vs. Infrastructure: Where Should Your Company Put Its AI Budget?
Table Of Contents
- The Three-Way AI Investment Decision Every Executive Faces
- AI Infrastructure: The Foundation You Can't Skip
- Model Training: When Building Your Own AI Pays Off
- AI Agents: Where Business Value Shows Up Fastest
- The Hidden Costs That Kill ROI Projections
- A Practical Allocation Framework for Business Leaders
- How to Decide: A Decision Guide by Business Maturity
- The One Investment Most Companies Get Wrong
- Conclusion: Spend Smart, Not Just Big
Agents vs. Model Training vs. Infrastructure: Where Should Your Company Put Its AI Budget?
Every week, another executive team approves an AI budget. And every week, a surprisingly large number of them make the same expensive mistake: they invest heavily in the most visible part of the AI stack while underfunding the layers that actually determine whether anything works.
The three biggest buckets competing for your AI dollars right now are AI agents (autonomous systems that execute business workflows), model training (building or fine-tuning your own AI models on proprietary data), and AI infrastructure (the cloud compute, orchestration, data pipelines, and governance tooling that everything runs on). Each promises a different kind of return. Each carries a different risk profile. And each suits a different stage of AI maturity.
This article breaks down the real economics of all three, using the latest enterprise data to help you understand which investment is right for your business, what a smart allocation looks like, and where most companies quietly waste their AI budgets. Whether you're just getting started or scaling existing deployments, the framework here will help you spend with intention rather than urgency.
The Three-Way AI Investment Decision Every Executive Faces {#three-way-decision}
When most business leaders talk about 'investing in AI,' they're conflating three fundamentally different bets. The first is betting on AI agents — deploying autonomous systems that handle real business workflows like customer onboarding, claims processing, or sales outreach. The second is betting on model training — paying to build, fine-tune, or adapt an AI model on your proprietary data so it understands your industry, your customers, and your business logic better than any off-the-shelf model can. The third is betting on AI infrastructure — buying or building the compute, data pipelines, orchestration layers, security controls, and governance frameworks that make any AI initiative sustainable at scale.
These aren't mutually exclusive, but your budget forces prioritisation. And the stakes are getting higher: enterprise generative AI spending reached approximately $37 billion in 2025, more than tripling from the prior year, split roughly evenly between the application and infrastructure layers. Knowing how to allocate your share of that equation determines whether your AI program delivers real competitive advantage or quietly becomes a cost centre.
The decision is complicated by the fact that the AI landscape shifts faster than most annual planning cycles. A workflow that fails the economics test today may become a strong investment in months. A custom model that seemed like a differentiator in 2023 may now be outperformed by a foundation model API at one-tenth the cost. Against that backdrop, a clear framework matters more than a fixed formula.
AI Infrastructure: The Foundation You Can't Skip {#ai-infrastructure}
Infrastructure is the least glamorous AI investment category and consistently the most underfunded relative to its impact. It encompasses cloud compute for inference and storage, data pipelines that connect AI to your enterprise systems, orchestration platforms that coordinate multiple agents or models, and the governance and observability tooling that keeps everything auditable and secure. It's the plumbing — and like all plumbing, nobody notices it until it fails.
The scale of infrastructure investment tells you how seriously the industry takes this layer. Global AI infrastructure spending is on track for roughly $497 billion in 2026, and Gartner projects total worldwide AI spending to reach $2.59 trillion in 2026, with infrastructure forming the largest single slice at over 45 percent. These aren't vanity numbers. They reflect what companies actually need to make AI work in production rather than in a proof-of-concept environment.
The reason infrastructure commands such a large share is partly about inference economics. In 2026, inference — actually running an AI model to process a query or execute a task — now accounts for roughly 80 to 90 percent of the lifetime compute cost of a production AI system, compared to just 10 to 20 percent for training. As agentic deployments multiply, that proportion only increases: agentic AI consumes 5 to 30 times more tokens per task than a standard chatbot interaction, because each task fans out into multiple reasoning steps, tool calls, and retries. Every agent you deploy adds to your inference bill.
For business leaders, the implication is clear: infrastructure is not a one-time capital decision. It's an ongoing operational cost that compounds with every AI initiative you add. A practical budget allocation benchmark puts infrastructure at 30 to 40 percent of total AI spend, framing it as the highest-ROI category because it reduces the cost and increases the effectiveness of everything else built on top of it. A strong infrastructure foundation makes your agent deployments cheaper to run, your model fine-tuning faster to execute, and your governance defensible to regulators and boards alike.
One area that routinely gets less than 5 percent of AI budgets despite outsized impact is the context layer — the systems that connect AI to your enterprise knowledge bases, CRM, ERP, and workflow data. An agent with excellent model access but no enterprise context is, in practice, a powerful engine with no fuel. Investing in this layer early pays dividends across every other AI initiative.
If you want to explore how to build this foundation correctly for your business, the Business+AI consulting program works with organisations to assess infrastructure readiness and design AI stacks that scale without runaway costs.
Model Training: When Building Your Own AI Pays Off {#model-training}
Model training — whether building a foundation model from scratch, fine-tuning an open-source model, or adapting a pretrained model on your proprietary data — is the AI investment category with the widest range of outcomes. Done right and in the right context, it creates a durable competitive moat. Done wrong, or chosen for the wrong reasons, it's one of the most efficient ways to burn several hundred thousand dollars with little to show for it.
The honest starting point is that for 95 percent of companies, training your own large foundation model is not on the table. That's a frontier AI lab problem, requiring billion-dollar compute clusters and machine learning teams measured in hundreds. What is on the table for most businesses is fine-tuning or domain adaptation: taking an existing foundation model and training it further on your proprietary data so it understands your specific products, customers, regulatory context, or industry terminology far better than a general model.
The case for this investment is real when the conditions are right. Bloomberg, for example, developed a custom LLM trained on financial data at an initial investment of over $10 million, and achieved approximately 30 percent better financial analysis accuracy than GPT-4 — a defensible competitive advantage for a data business where precision translates directly to customer value. The trade-off was a two-year development timeline and ongoing annual maintenance costs exceeding $2 million. That's a rational trade-off for Bloomberg. It may not be for your business.
Fine-tuning at the enterprise level carries its own cost structure worth understanding. Building the training datasets alone can cost $200,000 to $500,000 for regulated industries where labelled, clean data doesn't come ready-made. Then there's the MLOps infrastructure — training pipelines, version control, deployment orchestration, and monitoring — which adds $500,000 or more annually. And after all of that investment, you have a model that still requires prompting, still needs human review, and can't take autonomous action on its own. That last point is critical: a fine-tuned model is not the same as an agent. It needs to be paired with the right infrastructure and workflow design before it delivers business value.
The business case for model training is strongest when you have genuinely unique proprietary data that no foundation model has seen, when performance accuracy in your domain directly drives revenue or risk, when your use case requires compliance with data residency rules that prevent sending data to third-party APIs, and when you're operating at a volume that makes the fixed investment cost-effective over time. Outside those conditions, you are almost certainly better served by a well-designed agent built on top of a foundation model API.
Want to explore whether fine-tuning or custom model development makes sense for your specific use case? The Business+AI masterclass series covers AI strategy decisions like this in depth, with practical frameworks built for business leaders rather than ML engineers.
AI Agents: Where Business Value Shows Up Fastest {#ai-agents}
If infrastructure is where you build the foundation and model training is where you create differentiation, AI agents are where business value actually shows up in your P&L. Agents — autonomous systems that can plan, use tools, retrieve information, and execute multi-step workflows with minimal human intervention — are the most direct path from AI investment to measurable business outcome.
The market reflects this. Global spending on enterprise AI agent platforms is projected to reach $47.8 billion in 2026, up from $28.1 billion in 2025, a 70 percent year-over-year increase that outpaces every other enterprise software category. ROI expectations are high: executives anticipate an average 171 percent ROI on agentic AI deployments, and SAP and Oxford Economics research suggests enterprises expect average agentic AI ROI to reach $17.6 million within two years.
Those headline numbers deserve context. The distribution of outcomes is wide. The top quartile of deployments have achieved returns exceeding 800 percent, while the bottom quartile saw returns below 200 percent. What separates the top performers from the underperformers isn't the sophistication of the technology — it's the quality of workflow selection, the volume of transactions running through the agent, and the discipline applied to total cost of ownership rather than just token costs.
The total cost of running an agentic workflow is frequently surprising to business leaders who budget at the token level. For a customer-service agent in banking, for instance, token costs often represent just 20 to 25 percent of variable run costs. Human oversight — the risk and functional experts who review exceptions — accounts for 70 to 75 percent of variable costs. Multiply that across the thousands of runs a multiagent workflow executes, and customer-facing agents in some banks can cost $20,000 to $30,000 to run a single-agent workflow, and $100,000 to $200,000 for a full multiagent team. These are not small numbers. But for high-volume workflows where the fully loaded cost of a completed transaction falls from $50–$150 per customer to $10–$30, the economics can still be compelling.
The most important economic insight about AI agents is one that McKinsey's analysis makes clearly: completed work ROI is the metric that matters, not cost per token. What did it cost — fully loaded, including humans, agents, and deterministic systems — to finish the job? That's the number to measure against the value generated. Businesses that anchor their agent strategy to this metric tend to select the right workflows, design appropriate oversight models, and avoid the trap of deploying expensive agents on low-volume or low-value tasks.
Business line leaders looking to identify which workflows to agentify first, and how to build the business case, will find Business+AI workshops a practical starting point — bringing together real use cases, cost frameworks, and peer experience from across industries.
The Hidden Costs That Kill ROI Projections {#hidden-costs}
One of the most consistent patterns in enterprise AI programs is the gap between projected and realised ROI. A large part of that gap is explained by hidden costs that sit across all three investment categories and rarely appear in the initial business case.
On the infrastructure side, organisations with dedicated FinOps resources still underestimate AI infrastructure costs by up to 30 percent, according to IDC's FutureScape 2026 analysis. AI agent compute costs — LLM API calls, vector database queries, orchestration overhead — scale with usage in ways traditional software costs don't, and without proper usage monitoring and cost-per-transaction tracking, the margin gains from automation can be entirely offset by infrastructure spend.
On the agent side, cybersecurity, compliance monitoring, and agent operations (AgentOps) — maintaining, updating, and overseeing deployed agents in production — add costs that most initial deployments underestimate. Production agents often need adjustment every few days as new foundation models, model context protocols, and business requirements emerge. Budget for evolution, not just the build.
Perhaps the most underfunded hidden cost category is people and training. BCG's framework for AI investment allocates 70 percent of AI investment to people and process change — not technology purchase. Despite representing only 3 to 6 percent of total AI budgets in most enterprises, employee AI training shows the strongest correlation with realised ROI across BCG and Deloitte enterprise surveys. Programmes that fund training above 5 percent of budget significantly outperform those that fund below 3 percent. This means that the workforce capability to use, oversee, and improve AI systems is often the binding constraint on returns — not the quality of the technology.
EY estimates that full enterprise AI costs can reach roughly three times the token bill, and Gartner expects inference costs per agentic workflow to increase more than fivefold through 2028. For CFOs and CIOs, deploying more agents is therefore not automatically the same as creating more value. The discipline is in measuring completed-task cost against completed-task value, consistently and at the workflow level.
A Practical Allocation Framework for Business Leaders {#allocation-framework}
With the three investment categories and their cost dynamics understood, the practical question becomes: how should a typical business allocate its AI budget across infrastructure, training, and agents?
Based on multiple enterprise AI surveys and allocation frameworks, a directional benchmark for most organisations looks like this:
- Infrastructure (30–40% of AI budget): Cloud compute, AI gateway, data pipelines, orchestration platform, governance and observability tooling, and the context layer connecting AI to enterprise systems. This is the foundation — the highest-ROI category because it improves the efficiency and effectiveness of every other layer.
- AI agent development and operations (30–35%): Engineering time building and deploying agents and workflows, plus ongoing AgentOps costs — monitoring, updates, compliance checks, model routing, and exception management. This is where operational value is realised.
- Model fine-tuning and adaptation (10–15%): Reserved for organisations with genuinely proprietary datasets, domain-specific accuracy requirements, or data sovereignty constraints. Most companies should start closer to zero here and grow this allocation only when the case is proven.
- People, training and change management (10–15%): Workforce AI capability building, AI literacy programmes, process redesign, and internal governance development. This is the most underfunded category relative to its impact on ROI.
- Experimentation (5–8%): Pilots, proofs-of-concept, and exploratory investments in emerging capabilities. Keeps the portfolio forward-looking without over-committing to unproven approaches.
This framework is directional, not prescriptive. Actual allocations vary significantly by industry, AI maturity, and strategic intent. A financial services business with strong proprietary data and high regulatory requirements will invest more in fine-tuning and governance. A consumer business prioritising speed-to-market will weigh more heavily toward agents on top of foundation model APIs.
Expore peer benchmarks and allocation approaches from companies across your industry at the Business+AI Forum — an annual gathering of executives, AI solution providers, and consultants sharing real-world AI investment strategies.
How to Decide: A Decision Guide by Business Maturity {#decision-guide}
The right AI investment mix also depends heavily on where your organisation sits on the AI maturity curve. A rough guide:
Early stage (exploring AI for the first time): Prioritise infrastructure fundamentals and AI literacy above all else. Only 22 percent of organisations believe their current architecture can support AI workloads without modifications. Fix that before deploying agents or fine-tuning models. Focus your agent investments on one or two high-volume, well-defined workflows where the ROI is measurable. Avoid the temptation to fine-tune — foundation model APIs will serve most needs, and building training datasets alone can consume your entire year-one AI budget.
Scaling stage (running AI pilots and beginning to expand): Infrastructure investment becomes critical as agent volumes increase and variable costs begin to compound. Build your AgentOps capability now — the cross-functional discipline of continuously managing agent spend, performance, and compliance. Begin evaluating whether any of your workflows have the volume, accuracy requirements, and proprietary data to justify fine-tuning. Incorporate AI unit economics into quarterly business reviews so leadership can see where value is being created.
Mature stage (AI deployed at scale across multiple functions): The total cost of ownership mindset becomes essential. At scale, decisions about model routing — which tasks go to expensive frontier models versus cheaper lightweight models — can materially affect your cost base. Agent reuse, where a single agent is deployed across multiple workflows rather than rebuilt from scratch, becomes a significant efficiency lever. Fine-tuning investments can be evaluated rigorously with production data rather than projections.
Across all stages, the most reliable principle is this: the competitive moat in AI will not be access to cheap inference or even access to the best models. As infrastructure becomes commoditised, what remains is the quality of your agent architecture, the enterprise knowledge embedded in your workflows, and the organisational capability to iterate faster than your competitors.
The One Investment Most Companies Get Wrong {#most-companies-wrong}
If there's a single investment category where enterprise AI budgets are most consistently misallocated, it isn't infrastructure or agents. It's the people layer.
More than 40 percent of agentic AI projects are projected to be cancelled by 2027, according to Gartner, primarily due to unclear ROI and weak governance. Neither of those failure modes is a technology problem. They are organisational problems — the result of deploying powerful AI systems without the human capability to define meaningful success metrics, oversee autonomous workflows, or adapt processes to capture the value that agents can create.
Employee AI training, despite showing the strongest correlation with realised ROI in enterprise surveys, receives only 3 to 6 percent of AI budgets in most organisations. Leaders expect employees to be training 41 percent of agents and managing 36 percent of them within five years, turning agent supervision into a material workforce development requirement — yet the investment in building that capability consistently lags behind the investment in the technology itself.
The companies that will generate the most durable value from AI are the ones that treat workforce AI capability as an infrastructure investment in its own right, not a line item to be trimmed when budgets tighten. This means structured AI literacy programmes, hands-on workshops that build judgment rather than just awareness, and governance frameworks that keep humans meaningfully in the loop as agent autonomy increases.
Building that capability across your organisation is precisely what Business+AI's workshops and masterclass programmes are designed to accelerate — practical, applied learning that turns AI investment into demonstrable business outcomes.
Conclusion: Spend Smart, Not Just Big {#conclusion}
The race to invest in AI is real, but the race to invest intelligently in AI is the one that actually determines outcomes. Agents, model training, and infrastructure are not competing priorities — they are interdependent layers of a coherent AI strategy, each with a distinct role, a distinct ROI profile, and a distinct set of conditions under which it earns its budget allocation.
Infrastructure is the foundation that makes everything else work. Model training creates differentiation when the conditions genuinely justify it. Agents are where value shows up in your operations and your P&L. And the people layer — training, change management, and AI literacy — is the multiplier that determines whether your technology investment delivers or merely disappoints.
The executives who will build the most durable AI advantage are not the ones who spend the most. They are the ones who understand the total cost of ownership of each layer, measure the right metrics at the workflow level, and build the organisational muscle to iterate as the technology evolves. That's not a technology strategy. It's a business strategy — and it's one that's very much still being written.
AI economics are changing fast. Make sure your strategy is changing with them.
Ready to Build a Smarter AI Investment Strategy?
Business+AI connects executives, consultants, and AI solution providers across Singapore and the broader APAC region through hands-on workshops, masterclasses, consulting programmes, and the flagship annual Business+AI Forum.
Whether you're deciding where to put your first AI dollar or optimising a portfolio of deployed agents, our community and programmes help you invest with confidence.