A 290% ROI on Agent Infrastructure? What the New Denodo Study Actually Tells You About Agent Sprawl

TL;DR: A new vendor-funded study claims a unified AI data layer delivers 290% ROI over three years, and it highlights a governance problem every agent builder already knows: sprawl is expensive, on the data side and the skills side alike.

If you've deployed even a handful of AI agents at work, you've probably felt the friction the numbers describe. Every agent wants its own connection to the data it needs, its own permissions, its own little pipeline. Multiply that by hundreds of agents and dozens of systems, and the integration bill quietly becomes the biggest line item in your AI budget. This week, data-management vendor Denodo put a price tag on that friction. On October 20, the Palo Alto company announced the findings of a study by research firm Veqtor8 titled "The ROI of an AI Data Layer for Agentic AI," and the headline numbers are hard to ignore: 290% ROI over three years, 68% lower AI operating costs, and agents shipping three times faster.

What the study claims

Veqtor8, led by Chief Analyst Andrew Milroy, built its analysis on interviews with enterprise practitioners, hyperscalers, and technology partners across global banks, telecoms, chemical manufacturers, and logistics firms. From those interviews, the firm constructed a composite financial model representing the average enterprise in the sample: 100 agents in Year 1, growing to 900 agents by Year 3, drawing on 100 distributed data sources.

The three-year model produced these vendor-reported results:

A 290% return on investment. Three times faster agent deployment velocity. 68% reduction in AI operating costs. 2.9x more business value from AI initiatives. Over 75% of integration and governance costs avoided, including maintenance costs down 79% and policy administration effort down 76%.

One figure deserves particular attention if you pay for tokens. The study found that filtering data before it reaches the agent, rather than letting the agent wade through raw sources, cut context payloads per task from roughly 18,000 tokens to about 5,760. That's a 68% reduction in what you're feeding the model on every single task. At fleet scale, that's real money.

The report is available free at denodo.com.

The problem behind the numbers: MCP sprawl

The study's core argument is that connecting an agent to one system is easy. Connecting dozens of agents to dozens of systems, under consistent security and governance, is where enterprises bleed time and budget. Denodo calls this the MCP server sprawl problem. MCP, the Model Context Protocol, has become the standard way agents connect to external tools and data. But if every agent needs its own bespoke MCP server, its own retrieval pipeline, and its own security controls for every source it touches, you're rebuilding the same plumbing hundreds of times. In regulated industries, the study notes, a single production agent may touch 30 to 50 disparate systems spanning cloud platforms, SaaS apps, and legacy infrastructure.

Denodo's answer is to profile data once, enforce policy centrally, and let every entitled agent reuse the same governed data products without moving the data anywhere. Alberto Pan, Denodo's CTO, framed the pitch as removing the trade-off between speed, governance, and cost.

It's a coherent argument, and the token-reduction figure in particular has a clear mechanism behind it: less irrelevant context means cheaper, faster, more focused agents.

Read the fine print before you build a plan around it

Here's the caveat you should apply to any vendor-announced study, including this one. Denodo commissioned and announced the research. The "independent" label comes from Denodo's own press release. The composite model is an abstraction built from interviews, not a measured audit of a real deployment with receipts attached. The 290% ROI describes a modeled average enterprise, not your enterprise, and your mileage will depend on how many agents you run, how messy your data estate is, and how much integration work you're currently duplicating.

None of that makes the study useless. The direction of the findings matches what practitioners report anecdotally: integration and governance, not model quality, are the dominant costs of production agent fleets. But treat the specific percentages as directional claims from a vendor with an obvious stake, and pressure-test them against your own numbers before signing anything.

Data is half the problem. Skills are the other half.

Here's the part the study doesn't cover, and it matters just as much. A governed data layer gets your agents the right information. What they do with it depends on their skills, and right now most organizations manage agent skills the way they managed data access before tools like this existed: ad hoc, unvetted, and copied from wherever.

Agent skills, the SKILL.md instruction packages used by coding agents like Claude Code, Codex, Cursor, and Gemini CLI, are proliferating across GitHub and beyond. Published research has found that roughly 26% of them carry at least one vulnerability. That's the skills-side equivalent of the MCP sprawl problem: thousands of packages, no consistent verification, no way to know which ones actually improve an agent's performance and which ones quietly introduce risk.

This is the gap SkillzBench is being built to close. SkillzBench, a free directory and testing resource from AIQ Labs, is designed around the same governance principles the Denodo study argues for on the data side: every skill crawled and deduplicated from public sources, pinned to a specific commit so you know exactly what version you're running, license-checked, security-scanned, and benchmarked in a sandbox with the same model both with and without the skill, so you can see the measured Lift before you deploy anything. A trust score from 0 to 100 combines security, quality, license, maintenance, and community signals, and the Lift measurement stays separate from ratings so performance data never gets tangled with opinion. A free MCP connector will let you equip any agent with exactly the skills and tools you choose, and only those, so each agent sees just what you've approved for it. Upstream changes get tracked, re-scanned, and re-benchmarked automatically.

The parallel is worth spelling out. Denodo's pitch is that you shouldn't hand-roll data access per agent. SkillzBench's pitch is that you shouldn't hand-roll skills per agent either. Both are versions of the same principle: govern once, reuse everywhere, and test before you trust.

What a business manager should actually take away

If you're budgeting for agents in 2027, the study offers a useful template even if you discount its headline figures. Sit down and estimate the costs it quantifies: duplicated integration work per agent, per-source governance effort, token consumption per task, and maintenance drag as your fleet grows. Then ask three questions. How many systems does each agent touch? How much of that connection work are you repeating across agents? And what's your process for verifying the skills and tools each agent is allowed to use?

If your answers are "lots," "most of it," and "we don't have one," you've found your ROI. The data layer question is one vendors like Denodo will happily help you answer. The skills question is the one SkillzBench is building free tooling for, and it's the one most enterprises haven't started asking yet.

Bottom line

The Veqtor8 study, whatever you think of its sponsor, puts a useful frame on a real shift: agent economics are dominated by plumbing, not prompts. Enterprises that govern data access centrally are reporting substantial savings and faster deployment, and the same logic extends naturally to the instruction packages your agents run on. Whether you solve the data side with a commercial layer or the skills side with free, independently verified tooling, the principle is identical. Sprawl is a choice, and so is governance. If you want to get ahead of the skills half of the equation, SkillzBench's waitlist is open ahead of its public launch, and everything on the platform, the directory, the MCP connector, the benchmarks, and the alerts, will be free.

Key takeaways

• Denodo announced a Veqtor8 study reporting 290% ROI over three years for an AI data layer, with 68% lower AI operating costs and 3x faster agent deployment
• The study's composite enterprise scaled from 100 agents in Year 1 to 900 in Year 3, across 100 distributed data sources
• Pre-ingestion data filtering cut context payloads per task from roughly 18,000 tokens to about 5,760 tokens
• The study is vendor-commissioned, so treat the figures as reported claims rather than audited results
• The same governance logic applies to agent skills: published research found about 26% of skills carry at least one vulnerability, which is the problem SkillzBench is being built to solve