Hi friends,

I am writing this from Fort Mason in San Francisco, where I have spent 2 days at Glean:GO 2026.

I saw most of what follows early, at the VIP analyst and media session on day 1, before the keynote. I have sat through a lot of briefings. It is rare that one changes how I frame an entire category. This one did.

Here is the thing that struck me first.

Every enterprise AI vendor right now is telling you their model is smarter. Glean walked on stage and argued the opposite. Models are converging and becoming interchangeable. What is not converging is whether the AI actually understands your company.

That single idea sits underneath every announcement below.

I have written this up in full detail across 2 pieces on the site:

This newsletter is the condensed version, with everything linked.

Botsitting.

Glean gave the industry a name for something everyone has been doing quietly. It is the time you spend babysitting an AI before it can do anything useful. Finding the right files. Pasting in background. Connecting the tool. Correcting it. Steering it through every step.

You did not delegate the work. You added a coworker who needs onboarding every single morning.

The argument Glean made in the analyst session is that botsitting is not a model problem. It is a context problem. A model that does not know your customers, your renewal cycle, your naming conventions, or who owns what will need hand holding no matter how good its reasoning is.

Glean framed 3 bottlenecks in the official release: botsitting, rising token costs, and AI sprawl.

The headline launch is Glean Tau, a new AI desktop workspace built on an open source harness.

What makes it different from the assistant you know: Tau reaches past the browser. It combines Glean's enterprise context with the files, applications, and code sitting locally on your machine, plus the latest models.

Glean says Tau can plan, execute, review, recover, and act on your behalf.

That word "recover" is the one I would circle. Planning and executing is table stakes in agent demos. Recovering from a failed step without a human tapping it on the shoulder is where most agentic demos quietly break. It is the difference between a tool you supervise and a teammate you trust.

Coding and engineering workflows are an early use case, but Glean was clear this is not a developer tool. It carries governance, security, and cost controls in from the platform.

Why it matters: the desktop is the last unindexed frontier. We spent 3 years perfecting SaaS connectors while the actual work sat in a folder on someone's laptop. Whoever closes that gap safely picks up an enormous amount of real context. It also expands the surface a CISO has to sign off on, which is the tension I would watch over the next 2 quarters.

This is the part data leaders should read twice.

Glean expanded Glean Intelligence with centralized usage visibility and AI usage controls, plus enhanced auto routing that lets administrators prioritize efficient, balanced, or frontier models based on price and performance tradeoffs.

An administrator can now set a policy that routes commodity work to a cheap capable model and reserves the frontier model for the steps that actually change the outcome. The Model Hub carries 40 plus open and frontier models. Usage controls cover limits, forecasting, model access controls, and spend visibility across users, departments, and models.

Why it matters: for 2 years we benchmarked intelligence. In 2026 we are benchmarking harnesses and routing. Model selection has moved from an engineering preference to a line item a CFO can see. Most enterprises are not organizationally ready for that.

My favourite demo of the event, and I do not think it will get the attention it deserves.

Glean Transform uses real activity signals to understand how work actually happens inside your company. Not how the process document says it happens. How it actually happens. It then identifies the split between people and AI, recommends where AI should take on work, connects those opportunities to specific skills and agents, and measures business impact.

The example on screen was a sales discovery and use case scoping workflow. 9 steps. 30 sub steps. 1,502 instances per month at 5.5 hours each, which is 8,261 hours a month of human time. The recommendation: 3 agents covering 88 percent of the workflow steps, with an estimated 4 hours saved per instance and 6,008 hours saved per month.

Why it matters: nearly every stalled AI program I have covered in 18 months died at the same 2 questions. Which process do we automate first, and how do we prove it worked. Transform points a product at both. If the activity signals are accurate, that is a genuinely different way to build a roadmap. If they are not, it is a very confident spreadsheet. That is what to pressure test in your evaluation.

New team chat lets multiple teammates work with Glean together in the same shared thread or interactive artifact.

The detail I care about most: Glean responds using each participant's permissions. Access is controlled, reviewable, and revocable. A finance person and an engineer in the same thread do not start seeing each other's restricted data through the AI.

Why it matters: enterprise work is not 1 person prompting in a private window. It is 4 people arguing in a thread. The reason most tools do not offer shared AI threads is that multiplayer collapses the permission model. Solving that is harder than the feature looks.

Glean expanded its AI Gateway to support more AI entry points, enforce restricted topic policies through Glean Protect, and extend governed MCP access beyond enterprise context to include organizational skills and personal memory.

The announcement that made the analyst room go quiet was context aware AI threat detection. Glean uses signals from its Enterprise Graph, including who is acting, what data is involved, and the sequence of actions, to separate legitimate agentic activity from data harvesting, exfiltration, or destructive behavior.

The critical phrase: this includes cases where each individual action may be permitted, but the combined pattern signals risk.

Why it matters: this is the most underrated announcement of the event. Traditional access control asks whether a user is allowed to do a thing. Agentic security has to ask whether 40 individually permitted actions add up to something no human would have approved. That is composite risk, and it is a graph problem rather than a permissions problem. It also explains why Glean treats its index as a security asset and not just a search asset.

Glean is extending its assistant to act on context without waiting for a prompt. Proactive task management identifies what needs attention and moves work forward. Email triage brings company context into the inbox. A meeting coach is on the roadmap.

Interactive dashboards combine structured data with unstructured business context, tickets, conversations and documents, and refresh as the underlying information changes.

Why it matters: proactive is the hardest bet on this list. Get it right and it feels like a chief of staff. Get it slightly wrong and it feels like an assistant that will not stop talking. The difference is entirely context quality, which is exactly Glean's argument.

Be honest with yourself about that third group when you plan. Tau, Transform, threat detection, and email triage are the 4 most strategically interesting items announced, and all 4 are unreleased. That is normal for a flagship conference. It is also the right question for your account team.

Glean also published a benchmark, which I have written up in full here. The original is here.

$0.58 per task on Glean. $2.98 per task on Claude Cowork. An 81 percent reduction in token cost, with the Glean response preferred 78 percent of the time.

Run that at 100,000 tasks a month and the spread is close to $2.9 million a year on the same volume of work. That calculation is mine, so run it with your own numbers.

But do not stop at the headline. Cost per task decomposes cleanly:

tokens consumed x blended rate per million tokens

Variable 1 is set by your context layer. A prebuilt index means the agent is not rediscovering your company on every request, so prompts carry less dead weight and trajectories converge in fewer turns. That is retrieval precision showing up on the meter, not a less verbose model.

Variable 2 is set by your routing. Spread work across model tiers instead of running everything through 1 family. The counterintuitive detail Glean's engineers highlighted: the cheaper system used the frontier model more often, not less. It just used it surgically.

When you find a cost gap in your own stack, decompose it the same way. Token volume is a context and retrieval problem. Blended rate is a routing problem. Different fixes, different vendors, and you cannot tell which one you have from the invoice alone.

The caveat, because you deserve it. This is a vendor run benchmark against a named competitor. Glean designed the tasks, ran both sides, and graded the results. Glean has published more methodology than most of its peers, which I credit. It is still evidence, not a verdict, and Anthropic has not responded publicly as I write this.

Worth reading alongside it: Glean's earlier MCP evaluation from May, which held the harness constant and swapped only the context layer, and the enterprise search evaluation from February.

Instrument cost per completed task. Most enterprises cannot answer that for a single agent in production today. Until you can, every vendor conversation is theoretical.

Run the benchmark on your own workload. Take 50 real queries your teams actually run. Not demo queries. Run them across your candidate stacks, score preference blind, log tokens. It is 2 weeks of work and it will be the most useful 2 weeks of your year.

Stop evaluating assistants on demo quality. Evaluate them on 3 questions. What context can it reach. What does 1 completed task cost. Who can see and revoke what it did.

Do not let cheaper beat correct. A cheap wrong answer that a human redoes is the most expensive output in the building.

The industry spent 2 years benchmarking models. We are now benchmarking systems. That is a sign of a category growing up.

Emrecan Dogan, Glean's Chief Product Officer, put the thesis in 12 words: "Models are getting better and more interchangeable. Understanding the enterprise is not."

Whether Glean holds that layer, or Microsoft, or Anthropic, or someone not yet in the conversation, is the open question of the next 18 months. But the question itself has shifted from whose AI is smarter to whose system costs less to run correctly. That is a much healthier question for enterprise buyers.

Glean

The Ravit Show

That is the wrap from Fort Mason. If you are building your enterprise AI architecture this year, I would love to know which of these you think actually matters. Reply and tell me.

See you in the next one.

The next Apple may already exist. Insider sources say Elon has spent two years building a secret device inside Tesla's facilities — one he claims will be "10x bigger than the largest product in history."

There's just one problem: the company is private, and unless you know Elon personally, you can't buy a single share. That was true until Guardian's research team found three public ticker symbols sitting in the launch supply chain.

Click here to see all 3 tickers, free of charge.

You won't hear these names on CNBC — Wall Street hasn't published a word on the connection. But when the launch hits September 21, that quiet ends.

Some are already calling this the biggest opportunity since AI. For anyone who missed Apple before the iPhone, this may be a second look at that kind of setup.

Get the 3 ticker symbols here.

Unlock the free briefing now.

Ravit Jain
Founder & Host of The Ravit Show