When software stocks sold off in early 2026, the market treated the whole category as a casualty of AI. That reading was too broad. Goldman Sachs Research said in February that the repricing had been applied broadly rather than selectively,¹ and the better question is which software keeps improving the longer a customer uses it. The answer comes down to one capability: capturing decisions, not just records.
In a recent episode of AI Radicals, host Satyen Sangani sat down with Ashu Garg, a partner at Foundation Capital, to unpack where value goes when AI models become interchangeable. Full disclosure: Garg and Foundation Capital are investors in Alation.
The real exposure sits in seat-priced, lightweight interfaces
Project management tools, Garg says, "seem like spreadsheets on steroids," and a sales compensation calculator "never deserved to be an application."²
Goldman's analysts draw the same line in their February 2026 framework. The most exposed products act as lightweight user interfaces and earn most of their revenue from seats. At the platform and infrastructure layer, agents generally increase demand for data management, orchestration, and security. The analysts also estimated that valuations had moved from implying 15-20% medium-term growth to implying 5-10%, a repricing they judged too broad.³
Garg says the real "house on fire" problem for enterprises is software that is hard to use, inflexible, and short on the business value they bought it for. A 10% saving on the bill, he says, is "a rounding error."² Sangani adds that applications earn their keep in the last 20 percent: the reliability and scale that an enterprise AI operating model depends on.²
Durable software records why a decision happened, as well as what happened
Garg's central argument springs from the decision trace. A CRM records that a customer received a discount, but the reasoning, the exception, and the approvals live in Slack threads, emails, and calls, spread across a dozen or more interfaces.² The infrastructure that captures those traces over time is what he and his Foundation Capital partner Jaya Gupta call the context graph, a close relative of the enterprise context layer that many data teams now leverage.
Foundation Capital's essay on context graphs argues that systems of record store the current state. Salesforce knows what an opportunity looks like now, the authors say, but not what it looked like when the decision was made. They define the context graph as "a living record of decision traces."⁴ Independent support comes from Goldman, whose durability test asks whether a platform governs approvals, compliance, and execution.⁵ Gartner's May 2026 guidance concerns semantics rather than decision traces, but it points the same way: it advises data and analytics leaders to establish a context layer because traditional schema-based data models alone lack business context and data meaning.¹⁰
Owning the record protects a vendor, and capturing the decision lets that record compound in its accuracy. This layer also differs from both older structures (a distinction explored in knowledge graphs versus data catalogs.) The risk for everyone else is spelled out by Garg: if context is captured in a layer above an application, that application becomes "nothing but a database and an API."² He adds that no one is swapping out Salesforce today.²
Sangani approaches the same problem from the catalog side. What a team documents always lags behind what is currently true, and when four versions of the truth are circulating, the hard part is deciding which one is right.²
Apps that learn take on more decisions as they prove themselves
Garg’s test for any enterprise application is simple: ”Is the application that I use a month from now better than the application I use today?”² In practice, better means the app needs a person less often.
His example is a small agentic workflow focused on discounts. At first, an application checks with the VP of sales before every discount. Over time it learns from those approvals, starts approving routine cases on its own, and tells the VP afterward: “Hey, I made this decision, FYI.” Eventually, *”the FYI may not be necessary.”*² The VP’s attention shifts to the exceptions.
That progression separates software that compounds from software that stays put. A calculator that does one defined job performs the same on day 500 as on day one, and Garg notes that narrow tasks like that can already be handled in a chat interface. An app that learns from each decision becomes more valuable, and harder to replace, the longer it runs.
Autonomy has to be earned, though. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously by 2028, up from 0% in 2024.⁶ The same release expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls.
Sangani describes the same pressure from the vendor’s side. Customers say capabilities ship faster than they can absorb, so the product has to metabolize new releases for them.² (The Alation Intelligence Operating System is built to keep agents, context, and data current, governed, and improving.⁷)
Whether to fix the model or the context depends on the shape of the work
When an AI application gets something wrong, a team can fix it in two places. It can fix the model by fine-tuning it on examples so the correction is built in. Or it can fix the context by giving the application business rules and customer-specific knowledge to look up each time it acts. Garg’s view is that the right choice depends on the shape of the work: uniform, narrow tasks suit the first, and varied, multi-role processes suit the second.²
Hippocratic AI shows the first case. It builds voice nurses that phone patients for narrow, high-volume tasks, such as follow-ups after cardiac surgery, the kind that suit specialized enterprise AI agents. Hospitals largely follow the same nationwide protocols, so one nurse’s job looks much like another’s. Hippocratic AI therefore runs entirely on open-source models and fine-tunes them on the corner cases that surface in large-scale human evaluations.²
A sales process is the second case. Reps, sales engineers, managers, and VPs all use the same application, and decisions vary from one customer to the next. Fine-tuning a model for every customer would make the cost of embedding those decisions “exponential,” Garg says, and every change in a decision would mean fine-tuning again. So customer-specific decisions belong in a context layer instead. He suspects even Hippocratic AI’s hospital-specific decisions will eventually move there.²
This is why the SaaSpocalypse is narrow. A uniform task can be absorbed into a model or a chat interface. A process that varies by customer and spans roles needs a layer of context that someone has to build and maintain, and that layer is where a durable application keeps its value.
Outside evidence points the same way. Epoch AI finds the best open-weight models have trailed closed frontier models by about four months since January 2026.⁸ As models converge, as Garg expects, what differs between companies is their context. Gartner expects agents to progress from task-specific roles to collaborating agents and then to ecosystems that cross applications.⁹ That is where Garg says the context layer earns its keep, and where the gap between metadata and context starts to matter.
How do you know the context is right? Sangani’s answer is to run AI agent evaluations, because, in his words, "You can’t figure out your context unless you know what the application is on top.”²
What data leaders should take from this
Narrow, single-task products can be absorbed into a model or a chat interface. Products that span roles, customers, and applications keep a durable layer, and the context they capture is what separates software that learns from software that executes.
When you evaluate an application vendor, ask two questions. What does the product learn from each decision it touches? And who owns that accumulated context if you switch? Answers to both depend on AI governance that keeps the context trustworthy.
Curious to learn how Alation can support your context strategy? Book a demo with us today.
Sources & notes
Every external claim on this page is independently verifiable. The public sources are listed here.
1. Goldman Sachs Research said the repricing of software stocks was applied broadly rather than selectively. — Goldman Sachs, "Which Software Companies Will Benefit from AI?", 18 February 2026. ↗ https://www.goldmansachs.com/insights/articles/which-software-companies-will-benefit-from-ai
2. Quotations and paraphrases from Ashu Garg and Satyen Sangani are from the AI Radicals episode with Ashu Garg. — Alation, AI Radicals, "Open Source, Open Question," episode page. ↗ https://www.alation.com/podcast/episodes/open-source-open-question-ashu-garg/
3. Exposure concentrated in lightweight, seat-monetized interfaces; platform-layer demand for data management, orchestration and security; valuations implying 15-20% medium-term growth at their peak versus 5-10% after the selloff. — Goldman Sachs, 18 February 2026. ↗ https://www.goldmansachs.com/insights/articles/which-software-companies-will-benefit-from-ai
4. Salesforce described as built on current-state storage; context graph defined as "a living record of decision traces." — Jaya Gupta and Ashu Garg, Foundation Capital, "AI's trillion-dollar opportunity: Context graphs," 22 December 2025. Garg is a co-author, and Foundation Capital is an Alation investor. ↗ https://foundationcapital.com/ideas/context-graphs-ais-trillion-dollar-opportunity
5. Durability test: whether the platform governs approvals, compliance and execution (system-of-record ownership). — Goldman Sachs, 18 February 2026. ↗ https://www.goldmansachs.com/insights/articles/which-software-companies-will-benefit-from-ai
6. At least 15% of day-to-day work decisions autonomous by 2028, up from 0% in 2024; over 40% of agentic AI projects canceled by end of 2027; about 130 of thousands of agentic AI vendors are real. — Gartner, press release, 25 June 2025. ↗ https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
7. AIOS described as keeping agents, context, and data current, governed, and improving. — Alation, AIOS overview. ↗ https://www.alation.com/aios/
8. Open-weight models lagged closed frontier models by an average of four months since January 2026. — Jack Edwards and Luke Emberson, Epoch AI, 29 May 2026 (CC BY). ↗ https://epoch.ai/data-insights/open-closed-eci-gap
9. Progression from task-specific agents to collaborating agents to agent ecosystems across applications. — Gartner, press release, 26 August 2025 (updated 5 September 2025). ↗ https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
10. Gartner advises data and analytics leaders to establish a context layer as a core component of D&A infrastructure, because traditional schema-based data models alone lack business context and data meaning. — Gartner, press release,
ANALYST ATTRIBUTIONS & DISCLAIMERS
Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," press release, Anushree Verma, 25 June 2025. Gartner, "Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025," press release, Anushree Verma, 26 August 2025. Gartner, "Gartner Says Lack of Semantics Causes Inaccurate AI Agents and Wasted Spending," press release, Rita Sallam, 11 May 2026.
Gartner does not endorse any vendor, product or service depicted in its research publications and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose. GARTNER and Magic Quadrant are registered trademarks and service marks of Gartner, Inc. and/or its affiliates in the U.S. and internationally and are used herein with permission. All rights reserved.
Goldman Sachs: This article is being provided for educational purposes only. The information contained in this article does not constitute a recommendation from any Goldman Sachs entity to the recipient, and Goldman Sachs is not providing any financial, economic, legal, investment, accounting, or tax advice through this article or to its recipient. Neither Goldman Sachs nor any of its affiliates makes any representation or warranty, express or implied, as to the accuracy or completeness of the statements or any information contained in this article and any liability therefore (including in respect of direct, indirect, or consequential loss or damage) is expressly disclaimed.
- AI
- Data Catalog
- Data Governance
- Digital Transformation
- Enterprise Data Catalog
- Modern Data Stack
Keep reading
More from the data desk
