Agentic AI has rewritten the economics of software development: the unit of cost is shifting from interactions with AI to outcomes. The same capabilities that make agents valuable also make their costs harder to predict, pushing AI cost management and value realization to the top of executive priorities.
IBM Institute for Business Value research finds that 79% of executives report seeing productivity gains from AI, but only 29% can measure its ROI confidently. Closing that gap starts with understanding where AI costs are created and where organizations can control them. In AI-assisted software development, the total cost extends beyond model usage and licensing to the decisions developers make, the incentives leadership creates, and the systems and processes that support them.
How AI-assisted development workflows changed
Two years ago, AI in software development meant autocomplete and a chat window: a developer asked a question, got an answer and decided what to do with it. Today, AI agents execute entire workflows with dozens of steps before a human sees the result. Goldman Sachs Research forecasts that agentic AI will increase token consumption 24-fold between 2026 and 2030.
A chat interaction is a single exchange. An agent completing tasks can consume 10 to 100 times as much because every planning loop, tool call, file read, retry and self-correction adds to the cost. The cost profile has also changed. When a chatbot is wrong, you lose one answer. When an agent goes wrong mid-workflow, you pay several times: for the wrong turn, for detecting it and for the recovery.
Pricing models have shifted alongside these workflows. Per-seat licensing gave way to consumption-based or hybrid models. As consumption pricing meets increasingly capable agents, predicting the cost of AI-assisted software delivery has become a major concern for engineering and finance leaders.
The stakes are large and growing. Gartner forecasts worldwide AI spending of USD 2.59 trillion in 2026, up 47% year over year, and argues that improved predictability of ROI is a precondition for enterprises scaling AI. According to IBM's 2026 CEO Study, 69% of CEOs say that AI is changing their core business. AI has also moved from the innovation budget to the operating budget, which generally faces greater scrutiny.
A discipline has formed around this problem. 98% of FinOps teams now manage AI costs, up from 31% two years ago. An unmeasured shadow economy also runs alongside: MIT's NANDA research found that employees rely on personal AI tools in over 90% of firms, creating costs and risks that do not appear on a dashboard.
Measuring what matters
Direct costs such as tokens and licensing are the easiest to measure. Less visible are review time, rework, validation, governance, training, infrastructure and integration effort. These costs surface later in the development lifecycle and in different teams, making the total cost of AI harder to assess.
METR's mid-2025 randomized controlled trial illustrates the challenge: experienced open source developers took 19% longer to complete real tasks with AI tools, despite believing they worked about 20% faster. Much of the slowdown came from reviewing, correcting and integrating AI-generated code rather than generating it. METR's follow-up study with late-2025 agentic tools found that overall productivity improved.
The takeaway: measuring AI productivity involves the entire development workflow, not just code generation, and requires reassessments instead of relying on one-time evaluations.
The three levers where costs are made or saved
AI costs are not determined when a contract is signed. Three interconnected layers shape AI efficiency: the decisions developers make, the incentives leadership creates and the systems that support them. A weakness in any one layer limits the effectiveness of the others.
1. Developers and development. The choices developers make every day determine how efficiently AI operates and how often teams repeat work. Individually, each request costs little. Multiplied across thousands of developers and millions of requests, small inefficiencies become a significant operational expense.
Three decisions matter most. Context quality: poor context (vague instructions, missing constraints, stale documentation) can cause rework and unnecessary iterations. DORA's research finds that AI effectiveness depends heavily on documentation and accessible internal data. Model-task matching: not every task deserves the same intelligence. McKinsey's controlled study found time savings of 45-50% on documentation but under 10% on high-complexity tasks. Routing repeatable work to a heavyweight reasoning model wastes money. Execution efficiency: agentic tasks chain many stages and inefficiency multiplies through the chain. The classic failure is the retry loop, where an agent stuck rerunning a failing test consumes resources at full rate while producing nothing. Interrupting it early rather than late can reduce costs by an order of magnitude.
2. Leadership and mandates. Technology alone rarely determines AI costs. AI doesn't fix a team. It amplifies what's already there. Strong organizations use AI to get stronger while fragmented ones discover their fragmentation faster and at higher cost.
The leadership levers include governance, which prevents duplicated work and uncontrolled experimentation; visibility into both usage and outcomes, which FinOps practitioners consistently rank as their top challenge; incentives that reward the right behaviors without pushing developers toward either wasteful usage or underutilization; and standardization and procurement decisions around vendor selection, model flexibility and overlapping tools.
These leadership decisions become easier when organizations have a centralized view of AI usage, costs, governance and outcomes. AI software delivery partners such as IBM Bob™ can help organizations connect consumption, governance and outcomes in one place.
3. Processes and systems design. The system sits between the developer and leadership: the workflows, platforms and architectural decisions that determine how efficiently AI operates.
Effective AI development systems incorporate workflow design that lets expensive reasoning be performed once and reused multiple times through prompt caching and shared context; determinism, so predictable steps like linting and compiling don't consume tokens; and validation at each stage to catch agent mistakes before they spread. DORA found that AI-driven increases in change volume led to instability where automated testing, version control maturity and fast feedback loops were weak.
At scale, architecture determines efficiency. AI costs increasingly come from orchestrating many interconnected tasks rather than from isolated prompts.
How the three layers work together
No single layer determines total AI efficiency: improvements at one are capped by inefficiency at the others. Skilled developers without governance produce high-quality work that is hard to standardize or scale. Leadership policy without practical workflows struggles to influence daily engineering practice. Even sophisticated platforms cannot compensate for inconsistent habits.
When the layers reinforce one another, the benefits compound: policy becomes tangible when the platform makes efficient behavior the default rather than an act of individual discipline.
Making tradeoffs: The economics of cost optimization
The three levers give organizations multiple ways to influence cost, quality and productivity, but improving one objective often affects another. Cost optimization is less about minimizing a single metric than about making informed tradeoffs. Six principles help organizations avoid common mistakes:
- Every optimization has a tradeoff. A smaller model is cheapest only until a task exceeds its capabilities and rework outweighs the savings.
- Beware of optimizing the wrong metric. Goodhart's law applies: token usage became a target because it's easy to measure, yet trimming context produces visible savings and invisible errors downstream.
- Costs rarely disappear-they move. Telemetry across approximately 22,000 developers found median PR review time up 441% as AI-assisted throughput rose. GitClear found duplicated code up 81% since 2023 while refactoring dropped.
- Organizations adapt to optimization. Overly restrictive controls produce workarounds, not savings. Shadow AI increases wherever sanctioned tools are slower or rationed.
- Optimize where it matters. For smaller deployments, the optimization effort can exceed the savings. Get visibility first, then focus on the largest cost patterns.
- The optimal balance changes over time. METR's developers went from measurably slower on early-2025 tools to measurably faster on late-2025 agentic ones. Optimization must be an ongoing capability, not a one-time exercise.
Building AI cost discipline
Knowing where AI costs originate is necessary, but not sufficient. Organizations also need to understand what they gain from that investment and make deliberate tradeoffs between cost, quality and productivity. That requires looking across all three layers together. Developers influence how efficiently AI is used, leadership shapes the incentives and guardrails around that use, and systems determine whether efficient behavior can scale. Optimizing any one in isolation might just move costs elsewhere.
As AI capabilities and development workflows continue to change, the optimal balance will change with them. The goal isn't to minimize AI spending at all costs. It is to build the visibility and discipline to understand where costs are created, where value is generated and where optimization will have the greatest impact.
For IT and development professionals, the practical implication is direct: your daily choices about context quality, model selection and execution efficiency now carry financial weight that extends beyond your team. Teams that develop cost discipline alongside technical skill will be better positioned as AI budgets face tighter scrutiny. Those looking to build these skills systematically can explore an AI Learning Path for Software Developers or review AI for IT & Development resources to stay current with evolving practices.
Why this matters for IT and development professionals
The shift from per-seat to consumption-based pricing means every developer now directly influences enterprise AI spend through routine decisions. Understanding where costs accumulate-context quality, model-task matching, execution efficiency-is no longer a finance concern but a core engineering competency. Professionals who can articulate the cost implications of their AI usage and optimize accordingly will be better positioned as organizations tighten AI budgets and demand measurable ROI.
Your membership also unlocks:
