Arvind Jain on Enterprise AI, Open Models and the Future of Work

Glean founder and CEO Arvind Jain joins Harry Stebbings to debate enterprise AI, model competition and the future of work. Jain's views on enterprise search and AI platforms come from the perspective of a founder whose company competes in that market.

Asked whether he is motivated more by winning or fear of losing, Jain chooses the latter. He says each company starts in a new environment despite the lessons from previous ventures, and the pace of AI disruption makes it dangerous to rely only on past success.

Jain describes Glean as having begun with workplace search across many internal systems and evolved into an enterprise AI platform. Its product combines access to models with company context, he says; this is his description of Glean rather than an independent product assessment.

Stebbings asks whether enterprises should fear dependence on frontier-model providers. Jain says leaders worry about control of intellectual property, information and working methods. If agents perform core operations, he argues, dependence may extend beyond technology to the way a business actually operates.

Jain points to institutional learning: people repeatedly adjust a process, often without documenting every change. He predicts that agents could accumulate analogous learning and asks who will control it. Whether and how a particular agent retains such learning depends on its design and deployment.

Jain says customers increasingly want open models as their quality improves and AI bills rise. He identifies cost as the main current driver in his conversations, while noting that some organizations also require inference in private data centers.

He says some customers have become more comfortable sending data to model providers under appropriate contracts. This is his observation about customer attitudes, not a guarantee about any vendor's data practices or a substitute for reviewing a contract.

Asked whether Anthropic could displace Glean or other application companies, Jain calls some vertical offerings relatively shallow and says he has not seen widespread wholesale replacement of existing tools. He suggests they can expand access to work previously done only by specialists; these are his competitive observations, not independently verified market results.

Jain acknowledges that Glean competes with model providers for enterprise customers, including when Claude connects to internal systems. He argues that building useful enterprise context is more complicated than attaching systems through connectors, and that Glean must explain this distinction to customers.

Being early can create brand recognition and an opportunity to compete, Jain says, but is neither necessary nor sufficient for lasting success. He advises founders to solve customer problems while watching model capabilities, and sees advances by OpenAI, Anthropic, Google and open-model developers as useful inputs to application builders.

For enterprise workloads, Jain estimates that many models, including open models, can handle 90% or more of use cases. He calls this a form of commoditization and says Glean selects models for tasks and cost. The percentage is his broad estimate, not a published benchmark or assurance for every organization.

Jain says his team recently became comfortable running much of its workload on an open model, but that it is still early to know how customers will respond. He expects some concerns about Chinese models, including hypothetical backdoors and reputational risks; he does not present evidence of a specific backdoor.

Jain expects competition among closed labs and open models to pressure pricing and questions whether the model business alone will be as lucrative as some expect. He mentions hearing a pricing rumor without substantiating it; it is not evidence that a provider will change prices.

He distinguishes OpenAI's consumer product from Anthropic's developing ecosystem of automations, skills and connections, and says both are more than model suppliers. His prediction is that a majority of enterprise workloads will use open models in three years, not a reported adoption statistic.

Microsoft is one of Glean's most significant competitors, Jain says. Bundling can be effective, but he argues a specialized product can win on quality and that consumption-based pricing may weaken the advantage of offering software as part of a package.

Stebbings counters that a large enterprise may prefer an already approved vendor, avoiding the compliance and management burden of many suppliers. Jain agrees that the burden exists but says pricing has historically been a major challenge for companies competing against bundled products.

Jain says prospects more often tell Glean that they already have Microsoft Copilot than that they have chosen a lab's product and need nothing else. He calls Microsoft formidable but says it is too early to determine which type of competitor will be tougher.

The discussion turns to returns on AI investment. Jain sees pockets of measurable value, such as support teams resolving more cases after assistance with reading and summarization. His example is illustrative, not a controlled measurement of a specific deployment.

Coding looks different: Jain says developers are using AI extensively, but companies tell him faster code generation does not necessarily translate into faster product releases, because coding is only one part of delivery. These are his conversations and observations, not a universal finding.

At Glean, he says features are shipping faster, but a larger and more experienced team makes it hard to isolate AI's effect. Engineering productivity is difficult to measure; more lines of code alone do not establish more value.

Jain estimates that almost all initial code at Glean is now written with AI, while humans still review it. The team considered removing reviews because they can become a bottleneck, but he says the company remains relatively conservative about submitting generated code.

AI can produce a great deal of code that must later be understood and maintained. Jain says current AI tools are not perfect at that work, so Glean accepts the review cost while benefiting from faster drafting; Stebbings presses him on whether review erodes the speed advantage.

Jain calls enterprise AI ROI partly a throughput and context problem. In his view, a rudimentary connection to many systems may force an agent to spend time and tokens assembling information before doing useful work. He argues that providing better context can make applications faster and less costly.

Should leaders ask employees to replace themselves with AI? Jain thinks that is the wrong goal and says current systems cannot fully replace many jobs. Stebbings argues that automating most of a role could still reduce demand for it; they disagree about where the tipping point lies.

Jain argues that an organization accepting a partial solution may lose to a competitor that combines the same AI tools with human judgment. Stebbings questions whether retaining more people necessarily produces a better product, especially when large teams can slow decisions.

Glean has more than 1,000 employees, Jain says, and he hopes it could reach 5,000 in five years. This is an aspiration, not a staffing commitment. He expects businesses to use productivity gains to do more work and improve products rather than automatically shrinking.

Jain agrees that overgrown teams can slow one another down and recalls companies cutting staff before the current AI discussion. His counterargument is that people are also an asset if assigned to the right projects; neither speaker establishes a universal staffing rule.

The two also disagree about the price of AI relative to labor. Stebbings sees valuable coding tools as potentially underpriced; Jain expects technology costs to fall and questions whether the current cost of some agent workflows is justified.

Jain describes an engineering triage agent at Glean that he says handled about 95% of issues for a 15-person on-call team. He says the agent cost roughly $1 million per month, more than the team's cost, and questions its efficiency. He does not say the entire team was replaced.

He accepts that a 3.8% comparison to developer salaries does not look high in isolation, but claims open models can sometimes do comparable work at a fraction of the cost. The result depends on the task and setup; he does not present a detailed comparative study.

Jain predicts inference prices will fall by orders of magnitude, despite observing that some token prices had recently risen. Stebbings offers his own speculation about providers' incentives; neither provides evidence establishing why individual prices changed.

Jain expects per-person productivity to increase alongside the competitive bar for products. He predicts companies will need to deliver substantially more value for the same revenue, an argument for using AI to expand output rather than assuming headcount must fall.

Glean initially let employees experiment without strict token budgets, Jain says. He describes a wide spread in spending, from $20 a month to $10,000 or $15,000 for some people, and says more advanced use cases were concentrated among a small fraction of employees.

Jain says common uses such as questions and summarization are widespread, while he estimates advanced uses are limited to around 5% of employees. He favors sharing useful examples in company town halls over rewarding people simply for consuming more tokens.

On recruiting, Jain says general technology hiring had become easier than at the SaaS peak as large employers slowed hiring or cut staff, while demand and pay for top AI and machine-learning talent intensified. His comments are based on his market experience and include his own uncertainty about individual companies' headcounts.

Both discuss expensive seed-stage hiring. Jain prefers raising a sufficiently large round, but recalls Glean's Series C as particularly expensive: he estimates revenue was only a few million dollars while valuation exceeded $1 billion, and says he does not remember the revenue precisely.

Jain says the financing also helped communicate ambition and credibility to prospective employees. Both speakers believe investor reputation can help attract candidates, though that is their experience rather than a rule for every hire.

Jain says his disciplined approach to spending is under pressure in an AI land grab. He still believes companies need products customers value and returns on marketing spend, but recognizes that excessive conservatism might cause a company to miss a fast-moving opportunity.

Asked about future jobs, Jain predicts more composite roles that combine product, engineering and design or sales and technical demonstration. Stebbings points out that combining specialties can mean smaller teams; Jain responds that higher productivity may also increase the amount of work companies attempt.

Jain expects some narrowly defined data-analysis and recruiting-sourcing tasks to be absorbed into broader roles, with business users obtaining more answers directly. These are his predictions about changing tasks and jobs, not evidence that any specific role has already disappeared.

On sovereign models, Jain says he hears competing signals and is not an expert on whether the trend is rising or falling. Stebbings believes recent events make sovereignty more urgent; the speakers do not establish a comprehensive account of any country's policy or restrictions.

Jain argues that large up-front model-training costs make the traditional unfunded open-source development model difficult, while maintaining that the US has a strong open-source community in other fields. He expects greater interest in developing open models domestically.

Stebbings raises Chinese models' prominence on a routing service. Jain says running models in a controlled environment can make some users more comfortable, while expecting US policymakers and technologists to care about domestic capacity. He does not establish that isolated deployment eliminates every security risk.

Stebbings speculates about a political quid pro quo involving OpenAI and restrictions on competitors. Jain says he hopes that will not happen and doubts regulation can resolve the competitive problem. The exchange is speculation; neither provides evidence of such an arrangement.

In a quick-fire round, Jain tells computer-science students not to be unduly discouraged. He names Google as an established company that has embraced AI, while acknowledging it is itself an AI company and therefore not a clean comparison with every legacy business.

Jain worries that abundant startup capital can encourage unsustainable costs, such as very high engineering salaries at seed-stage companies. At the same time, he thinks building a startup and achieving a good exit can be easier today than in some earlier periods, while emphasizing that entrepreneurship remains difficult.

He describes founding as stressful rather than glamorous and says a founder needs a strong sense of mission. Financial security may have helped him make decisions without worrying about feeding his family, he says, but he believes sustained effort and a desire to build something significant remain essential.