Why Most AI Agents Lose Money — And the Scoping Discipline That Flips the ROI

To improve AI agent ROI, stop measuring activity and start scoping for outcomes: pick workflows where autonomous completion creates measurable business value, upgrade the tech stack agents depend on, and budget for the 5-to-30x compute cost upfront. McKinsey’s 2026 data shows only 37% of companies see any EBIT impact — the gap is build quality and scope, not the technology.

The headline number everyone cites from the McKinsey Technology Trends Outlook 2026 is grim: 89% of companies now use AI regularly, yet only 37% attribute any positive EBIT impact to it. Most commentary stops there, treating the gap as proof the technology is overhyped. That reading is lazy. The 37% figure isn’t a verdict on AI agents — it’s a verdict on how companies scope, cost, and build them. The buyers who close that gap do three specific things differently, and none of them require waiting for the technology to mature.

Why do 89% of companies use AI but only 37% see profit?

The gap between adoption and profit comes down to confusing usage with value. According to McKinsey, most companies are at least experimenting with agents, but experimentation generates cost without discipline. Agents add complexity to IT, demand tech-stack upgrades to run reliably, and burn far more compute than the chatbots they replace. When a company bolts an agent onto an unprepared stack and measures success by “are people using it,” the EBIT line never moves.

The software development data makes this concrete. A McKinsey study found that in one case, AI tools increased programming activity by 180% — but published releases rose only 30%. More code is not more software. Worse, productivity declined in 30% of companies after teams adopted coding agents. If your yardstick is “lines generated” or “number of AI interactions,” you will congratulate yourself while margins erode. The economic benefit lives in the outcome of the whole process, not the volume of activity inside it.

Why agent workflows cost 5-30x more than a chatbot

The single most underestimated line item in agent budgets is compute, and it is the quiet killer of ROI. A chatbot answers a query with one inference. An agent runs repeated inference loops, calls tools, retrieves documents, and coordinates across systems to finish a task. McKinsey calculates that a single agent workflow requires five to 30 times more computing resources than a typical chatbot query. In a recent McKinsey survey, 93% of participants said this pushed them past their AI budgets.

That is not a rounding error — it is the difference between a project that pays back and one that quietly bleeds. If you’re a finance leader approving an agent pilot, assume the per-task cost is an order of magnitude higher than your chatbot experience suggests, and insist the business case survives the 30x worst case, not the 5x best case. A pilot that only works at the cheap end of that range isn’t a business case; it’s a gamble.

What has to be upgraded before agents work reliably?

Agents fail silently when the plumbing underneath them isn’t ready. McKinsey is explicit that existing tech stacks need upgrades for agents to function reliably, because an agent is only as good as the systems it reaches into. It needs clean, accessible data; stable APIs to act on; and orchestration with human-in-the-loop checkpoints to catch mistakes before they compound across a long workflow.

Newer tooling helps close that gap. Per the McKinsey report, agent development tools from Anthropic and OpenAI have improved how agents use tools and interact with computers, while frameworks like LangGraph from LangChain added the orchestration and human-in-the-loop controls needed to guide agents through longer tasks. But frameworks don’t fix a brittle data layer. Before scoping an agent, most companies need the unglamorous groundwork first — reliable custom API and data pipeline integration so the agent has something trustworthy to act on. Skip that, and you’re automating on quicksand.

A scoping framework for agent workflows that actually pay back

The ROI pattern in the data is unambiguous: concentrated value, not evenly spread. According to McKinsey’s study, around 80% of developers using AI tools saw productivity gains of only about 3%, while the top 20% saw 55%. Only a quarter of companies using agents for software development saw real gains in cycle time. The winners weren’t luckier — they scoped tighter.

A workflow is worth building an agent for when three things are true. The task is high-volume or high-value enough that a 5-30x compute cost still nets out positive. The outcome is measurable — a closed ticket, a reconciled invoice, a shipped release — not a vague “assistance.” And the surrounding systems are stable enough for the agent to act without constant breakage. If you’re a support operation drowning in repetitive tier-one tickets with clean CRM data, that is a payback candidate. If your process changes weekly and lives in three disconnected spreadsheets, automate the simple path first before reaching for an agent.

Custom-built vs. off-the-shelf agents: where the ROI gap comes from

Off-the-shelf agents are the fastest way to see something work and the fastest way to plateau at that 3% gain. A generic tool can’t be scoped to your highest-value workflow, can’t integrate deeply with your proprietary systems, and gives you no control over the compute cost that determines whether the thing pays back. You get the demo, not the margin.

A custom-built agent earns its ROI by being scoped to one outcome you can measure, wired into your real data and tools, and governed by guardrails you own. That is the difference between the top-20% performers and everyone else. But building is not always right. If the workflow is generic, low-volume, or already well-served by a SaaS feature you pay for, building a custom AI agent is over-engineering. Buy the tool, measure it honestly, and only commission a custom build when the off-the-shelf ceiling is visibly costing you money. Within two years, the 37% figure will likely split sharply: companies that scoped for outcomes will post real EBIT gains, and the rest will keep funding activity.

How to budget for agentic AI without blowing past projections

Start the budget from the compute reality, not the pilot invoice. Model the per-task cost at the high end of McKinsey’s 5-30x range, multiply by expected volume, and add the stack upgrades the agent depends on. Then set the success metric before launch — releases shipped, tickets resolved, hours returned — so you’re never tempted to celebrate activity. With 93% of surveyed companies overrunning their budgets, treat compute as a variable cost that scales with usage, not a fixed license fee — that discipline keeps projections honest.

FAQ

Q: How do you improve AI agent ROI? A: Scope each agent to a single measurable business outcome, upgrade the data and API layer it depends on before launch, and budget for compute at 5-30x chatbot cost. According to McKinsey, only 37% of companies see EBIT impact — the winners concentrate on high-value workflows rather than spreading agents thinly.

Q: Why are AI agents so expensive to run? A: Unlike chatbots that answer in a single inference, agents run repeated inference loops, tool calls, retrieval steps, and cross-system coordination. McKinsey calculates a single agent workflow uses five to 30 times more compute than a typical chatbot query — which is why 93% of surveyed companies exceeded their AI budgets.

Q: Should you build a custom agent or buy an off-the-shelf one? A: Buy when the workflow is generic, low-volume, or well-served by existing SaaS. Build custom when a specific, high-value workflow needs deep integration with your systems and tight control over compute cost — that scope discipline is where the real ROI gap comes from.

Key Takeaways

• Measure agent success by business outcomes — releases shipped, tickets resolved — never by code volume or interaction counts; one McKinsey case saw 180% more code but only 30% more releases.
• Budget compute at the 30x worst case, not the 5x best case; if the business case only works at the cheap end, it isn’t a business case.
• Fix the data and API layer before deploying agents — unreliable plumbing is why agents fail silently and productivity dropped in 30% of adopting companies.
• Concentrate on a few high-value workflows rather than spreading agents thin; the top 20% of developers saw 55% gains while the rest saw 3%.
• Default to buying off-the-shelf, and commission a custom build only when that ceiling is measurably costing you money.
• Book an AI ROI assessment before your next agent investment, so scope and compute cost are modeled before budget is committed — not after it’s overrun.