The Agent Stack Meets the ROI Wall

Within a 48-hour stretch, three large platform vendors each tried to define a different floor of the enterprise agent stack. AWS used its New York summit to introduce Context, a managed, self-learning knowledge graph designed to give agents governed access to enterprise data without manual curation, paired with Continuum for agent security and expansions to Bedrock AgentCore, Kiro and a DevOps agent. The framing across both AWS's own materials and outside coverage is consistent: agents are only as trustworthy as the data they can reach, and AWS wants that reachability to run through its services.

Databricks pushed at an adjacent layer. Lakehouse//RT, announced at its Data + AI Summit and covered alongside a new transactional/analytical processing approach, claims sub-100ms query latency at high concurrency and up to 16x performance gains over dedicated real-time serving systems, with named customers reporting multi-fold response-time improvements. The strategic argument is consolidation: collapse separate operational and analytical systems so agents and dashboards read from the same governed substrate. These are vendor- and customer-reported figures from a beta product, not independent benchmarks, and should be read as directional architecture signals rather than settled performance facts.

Adobe took the same logic into a very different surface. Its creative agent entered public beta across Premiere, Photoshop, Illustrator, InDesign and Frame.io, and was simultaneously wired into ChatGPT, Claude, Copilot, Gemini and Slack. Adobe's own survey of more than 16,000 creators, cited in its announcement, reports that 75% describe creative AI as integrated or essential to their work while 85% insist final creative decisions should remain human — a useful tell about how vendors are positioning agents as orchestrators of multi-step production rather than as replacements for judgment.

Read together, the three announcements describe a shared pitch: a context and governance layer (AWS), a unified data and serving layer (Databricks), and an embedded action layer inside the tools end users already open (Adobe). For technology and operations leaders, the practical question is less which vendor wins than how many of these layers a given organization wants under one roof — and what that implies for lock-in, integration cost, and the speed at which agents can be deployed against real workflows.

Sources: VentureBeat AI (https://venturebeat.com/data/aws-enters-the-context-layer-race-with-a-graph-that-learns-from-agents-not-manual-curation); cio-dive (https://ciodive.com/news/AWS-continuum-AI-security-claude-mythos/823184); aws.amazon.com (https://aws.amazon.com/blogs/machine-learning/context-intelligence-for-your-data-and-ai-agents-at-scale/); aboutamazon.com (https://www.aboutamazon.com/news/aws/aws-summit-nyc-2026-ai-agents); databricks.com (https://www.databricks.com/company/newsroom/press-releases/databricks-launches-lakehousert-bring-real-time-analytics-directly); siliconangle.com (https://siliconangle.com/2026/06/16/databricks-declares-end-pipelines-unified-platform-operational-analytical-data/); news.adobe.com (https://news.adobe.com/news/2026/06/adobe-unveils-major-expansion); blog.adobe.com (https://blog.adobe.com/en/publish/2026/06/18/more-time-spend-on-your-craft-adobe-brings-power-creative-agent-creative-cloud-apps); 9to5mac.com (https://9to5mac.com/2026/06/18/adobe-expands-firefly-capabilities-extends-agentic-tools-to-creative-cloud-apps/)

While vendors expanded the stack, the demand side tightened. Coverage of analyst data this week reports that more than half of generative AI proofs of concept have been abandoned, and projects that as many as 40% of agentic AI initiatives will be cancelled by the end of next year. The same coverage notes that AI risk disclosures among large US public companies have risen sharply over two years, and points to debt-financed AI infrastructure and per-employee tooling costs as pressure points that boards and audit committees are starting to notice.

Venture-side commentary echoes the shift. A widely cited investor interview describes named enterprises burning through annual AI budgets in months and trimming model licenses as initial enthusiasm meets cost reality. A separate practitioner piece argues that most stalled programs fail not on the technology but on the basics: unclear ownership, missing success criteria, and tools deployed before the business problem is defined. That diagnosis is vendor-adjacent — the author works for an automation platform — but the pattern it describes is consistent with the harder numbers elsewhere in the window.

A Genpact/HFS Research study of more than 2,000 executives, surfaced through both a vendor release and a bylined op-ed, puts a frame around the gap. It identifies four compounding 'enterprise debts' — data, process, technology and talent — and estimates roughly $18 trillion in recoverable value across the Global 2000 if they are resolved, with potential for 8% faster revenue growth and 16% lower operating costs. The same study reports that only about a third of enterprise data is AI-ready, around a third of the workforce is AI-ready, and only 6% of enterprises have funded debt-resolution programs at scale. These are modeled figures from a vendor-sponsored study, not audited results, but the direction is hard to miss.

The contrast with the buying climate is sharper still. A separate CFO survey reports that two-thirds of finance leaders plan to increase technology and AI spending over the next year, even as economic confidence sits at a multi-quarter low, and that nearly all responding organizations are now piloting, scaling or fully integrating AI. The story of this window is not retreat — it is a bifurcation. Budgets are still moving, but the tolerance for open-ended experimentation is visibly narrowing, and the questions attached to each renewal are getting harder.

Sources: media.genpact.com (https://media.genpact.com/2026-06-15-Enterprises-Are-Sitting-on-18-Trillion-in-Trapped-AI-Value-New-Research-Shows-How-to-Unlock-It); fastcompany.com (https://www.fastcompany.com/91559945/you-cant-build-your-ai-future-on-broken-foundations); TechCrunch AI (https://techcrunch.com/video/neas-tiffany-luck-says-enterprises-are-still-figuring-out-their-ai-roi); bitcoinworld.co.in (https://bitcoinworld.co.in/nea-tiffany-luck-enterprises-ai-roi/); CFO Dive (https://cfodive.com/news/cfos-boost-tech-spending-despite-economic-uncertainty-grant-thornton-ai/823210); em360tech.com (https://em360tech.com/podcasts/why-most-enterprise-ai-investments-fail-deliver-roi); techtarget.com (https://www.techtarget.com/searchcio/feature/AI-market-correction-What-IT-leaders-must-know)

Two policy signals this week are worth holding alongside the vendor and ROI stories. The US administration imposed export controls on a frontier model and its underlying base model, restricting access for foreign nationals — including, per the reporting, those employed by the developer inside the United States. That is a notable extension of national-security tooling into the day-to-day mechanics of who can touch which weights, and it lands directly on hiring, deployment geography and customer-access questions that AI-using enterprises had largely treated as commercial decisions.

In parallel, leaders of the largest frontier-model companies joined heads of state at the G7, where the agenda included frontier AI risks, infrastructure, sovereignty and child safety. The presence itself is the signal: AI governance is now being shaped at the level where standards tend to harden into binding rules, with model developers at the table.

For executives, the practical literacy point is narrow but important. Vendor selection, workforce composition and cross-border deployment can no longer be treated as purely technical or commercial choices. The same agent stack being assembled by AWS, Databricks and Adobe will be deployed into an environment where which model is allowed, by whom, and where, is increasingly a policy variable rather than a procurement one.

Sources: the-verge-ai-feed (https://theverge.com/podcast/951542/anthropic-claude-fable-5-mythos-ban-pentagon-ai-regulation-trump); cnbc.com (https://cnbc.com/2026/06/17/g7-trump-ai-tech-leaders-openai-anthropic-google.html)

Vendors are assembling an 'agent stack' — context/knowledge graphs, real-time data engines, security layers, embedded copilots — and selling it as the foundation for production AI. Buyers increasingly need a parallel 'evidence stack': defined business problems, named owners, measurable outcomes, and exit criteria. When the two are out of sync, spend compounds faster than value. This window is the clearest sign yet that the two stacks are being built at very different speeds.