Tl;dr
Glean used the conference to reposition itself. It is no longer pitching “enterprise search with a chat box on top” — it is pitching a full agent platform, and it is now competing head-on with the horizontal assistant products such as Claude Cowork and Slack.
Four things worth paying attention to:
1. Glean is competing on unit economics, not just capability. They put a cost-per-task comparison (81% cheaper and 78% preferred responses) against Claude Cowork on the main stage. Their argument is architectural: a pre-indexed corpus and flexible LLM routing is cheaper to query than live federated MCP search in Cowork.
2. “Unified front door” is not one door. The strongest session of the two days was the Expedia IT panel, which landed on the opposite of the usual UFD pitch — different user populations need different entry points, unified by shared context rather than a shared UI. Notably, Glean’s own AI Gateway is described as serving “multiple front doors,” so they appear to agree.
3. The agent lifecycle session was the most immediately usable content. A concrete 7-step framework plus a hard-nosed diagnostic for why most agent projects die. We could adopt this framing more or less as-is.
4. The partner ecosystem is where the real gaps show. Seven Glean partners presented, and most exist to fill something Glean doesn’t do well natively — security, data masking, action execution, document generation.
Competitive Read: Glean vs. Claude Cowork
Glean made an explicit cost-and-quality argument against Claude Cowork in the keynote.
Figure 1. Conference slide comparing the reported cost per task: $0.58 for Glean versus $2.98 for Claude Cowork.
The architectural argument, stated plainly: Glean pre-indexes everything, so retrieval is a search over a warm index. Cowork federates live via MCP at query time. Glean claims that makes them both cheaper and faster.
Figure 2. Conference slide showing the reported response preference: 78% for Glean versus 22% for Claude Cowork.
Take this with a pinch of salt because:
- Pre-indexing trades freshness for speed. A warm index is cheap to query but is only as current as its last crawl. For agreement data and anything transactional, that trade may go the wrong way.
- The benchmark is “cost per task” on their task mix. Task selection drives this number almost entirely. So does the 78% preference figure — we don’t know the grader design or the task distribution.
- The cost comparison almost certainly excludes indexing infrastructure. That cost is real and ongoing; it’s just amortized somewhere the slide doesn’t show.
Net: the architectural insight is legitimate and worth internalizing — pre-indexed vs. live-federated is a genuine fork in the road with real cost implications. The specific numbers are marketing until someone runs them on our workloads.
Product Launches
Glean Tau — the headline launch. A general-purpose agentic app, closely comparable to Claude Cowork, that can execute tasks on your machine. Glean’s stated wedge: it arrives already loaded with enterprise context, so there’s no MCP-connection setup tax before it’s useful. That onboarding-friction argument is probably their strongest one, and it’s more durable than the cost claim.
Team Chat — multi-participant AI conversations. A direct shot at Slack: humans and agents in the same thread, with Glean owning the surface.
Glean Slack Agent — the hedge against the above. A native Slack interface, for the customers who aren’t leaving Slack.
Glean Transform — an agent-opportunity dashboard that helps enterprises identify which workflows are candidates for agent automation, including “skill mining.” Notable as a pre-sales and adoption instrument as much as a product: it manufactures the business case for buying more agents.
Glean Intelligence — the model-management and cost-control layer, and the piece most relevant to how this category is maturing. Four components:
| Component | What it does |
|---|---|
| Model hub | 40+ open and frontier models available under one roof |
| Auto Routing | Matches each task to the cheapest model that can do it (demos: “summarize last three QBRs” → GPT-5.6 Luna; “make a first call deck” → Claude Opus 5; “explain this diagram” → Gemini 3.6 Flash) |
| AI Usage Controls | Usage limits, spend forecasting, model access controls, and usage visibility — applied across users, departments, and models |
| AI Gateway | Models, context, and memory delivered into multiple front doors, with MCP tools, observability, mobile device management, and consistent security |
Figure 3. Glean Intelligence brings together model access, auto routing, usage controls, forecasting, and an AI gateway.
Two features that stood out to me:
- Graceful quota degradation — when a user exceeds their allocation, Glean silently downgrades them to cheaper models rather than cutting them off. Meaningfully better UX than a hard wall.
- Spend forecasting — forward-looking budget visibility, not just a retrospective usage dashboard.
The pattern: AI spend predictability is maturing from an ops afterthought into a differentiated, sellable capability.
The Full Platform Stack
Worth understanding as a whole, because it shows where they intend to compete. Four layers, bottom to top:
| Layer | Contents |
|---|---|
| Glean Protect (foundation) | Context-aware threat detection · agent guardrails · skills & memory governance |
| Glean Enterprise Context | Enterprise graph — infrastructure, products, processes, content, customers, people. Personal graph — goals, tasks, work habits, writing style, response style |
| Glean Intelligence (new) | Auto routing · AI usage controls · model hub |
| Surfaces (top) | Glean Assistant (Tau, Team Chat, Proactive AI) · Glean Agents (independent agents, sales, on-call management, voice of the customer) · Glean Transform (skill mining) · Open Platform (AI Gateway, APIs) |
Figure 4. Glean’s platform framing connects assistant experiences, agents, Transform, Intelligence, enterprise context, and Glean Protect.
The strategic read: the moat they’re claiming is the Enterprise Context layer, not the assistant on top. The personal graph is the more interesting half — modeling an individual’s goals, work habits, and writing style is a stickiness play that a horizontal assistant with MCP connectors can’t easily replicate on day one. Everything above that layer is increasingly commoditized; everything below it is table stakes.
Session: Inside the Agent Development Lifecycle
The subtitle was “How to develop agents that drive ROI.” The most immediately usable content of the conference.
The framing
“Agents are software. And should be treated like it.”
Figure 5. A conference slide capturing the core principle: agents should be treated like software.
Simple, and correct. The whole session followed from it: if agents are software, they need a development lifecycle, an owner, tests, telemetry, and a governance review — not a prototype and a demo.
The 7-step lifecycle
Figure 6. The lifecycle presented for building agents that drive ROI: opportunity, design, performance, context, develop, launch, and monitor and improve.
Note that “Performance” comes third — metrics and the golden test set are defined before any context wiring or building happens. That ordering is the whole point, and it’s the step most teams skip.
Avoiding the zombie agent trap
The sharpest slide of the conference: “Without a lifecycle, many enter and few emerge.” A funnel showing agent attrition:
Figure 7. Without a lifecycle, many agents enter experimentation but few become adopted, measurable business capabilities.
The five questions that gate the funnel:
- Why did we build this agent?
- What exactly is it supposed to do?
- How do we know if it is working?
- What is it allowed to do?
- Who owns it after launch?
“If those answers are missing, agent activity does not become a strategy.”
That line is worth repeating internally. It’s a clean test for distinguishing agent activity from agent strategy, and it’s directly applicable to how we evaluate our own agent proposals.
Session: Choosing Your AI Front Door (Expedia IT Panel)
Directly relevant to a lot of enterprises defining their AI strategy
The grocery store analogy (worth stealing for our internal framing):
A grocery store has one front door — but the delivery driver doesn’t use it. They use the stocking door. And yet if that same driver comes in to pick up a paycheck, they walk through the front door like everyone else.
The point: “unified front door” means different things to different user populations, and the same person needs different doors depending on what they’re trying to do. The unification that matters is context, not UI. Once a user is in — through whatever door — they should be able to reach context across every connected service via MCP.
Worth noting that Glean’s own architecture concedes this: the AI Gateway is explicitly described as delivering models, context, and memory into multiple front doors.
Other takeaways from the panel:
- The adoption curve is consistently slow-then-fast. Early-stage adoption metrics will look worse than the eventual trajectory. Worth setting expectations accordingly.
- Expedia switched their QBR decks over to Glean Interactive Artifacts (new HTML based deck and dashboard feature). Action item: we should test this — deck production is a real recurring cost for us.
- An audience member asked the panel, “what they were doing with Copilot since Expedia is a Microsoft shop and whether it was set against Glean?”. The moderator cut the question off. Coexistence with Microsoft is clearly a live and uncomfortable question for this category.
Evening Keynote: Jeetu Patel (Cisco CPO) on the Human Bottleneck
The sharpest framing of the conference:
If coding is automated, code review becomes the bottleneck. If review gets automated, then taste — knowing what to build — becomes the bottleneck.
The bottleneck doesn’t disappear as you automate. It moves upstream, and it keeps moving toward judgment. For a PM org, that’s the whole argument for why this wave raises the value of product judgment rather than commoditizing it.
His two supporting points:
- Learning to learn is a durable skill.
- Learning to unlearn matters just as much — and is much harder. Hard-won expertise in a workflow that’s about to be automated is a liability if you hold it too tightly.
Day 2: Partner Ecosystem
Seven Glean partners presented. Read as a group, they’re a map of what Glean doesn’t do natively.
| Partner | What they do | Gap they fill |
|---|---|---|
| Fireworks | Inference engine over open-source models, tuned per domain | Domain-specific performance at a cheaper cost |
| BaseTen | Optimized inference on open-source models (similar to Fireworks) | Inference cost / performance |
| Skyflow | Security layer over Glean (marketed as Glean Protect) | Security posture |
| Seclore | Data-level security — context-aware masking of PII and sensitive data per query | Query-time data protection |
| Gamma | Interactive documents generated from Glean data | Document / presentation output |
| StackOne | Action execution layer, with Glean as control plane. Demos: new-hire provisioning, access & IT ops, month-end close | Doing things, not just answering |
| Enterpret | Customer issue detection from social channels | Voice-of-customer signal |
Two observations:
Security is the loudest gap. Two of seven partners are security plays, and Seclore’s framing was the most interesting thing on the Day 2 stage: the industry has moved network security → endpoint security → cloud identity security, and the AI era forces the next move to data security — protecting the data itself at query time, because the retrieval layer will surface anything it’s allowed to see. That reframing is directly relevant to how we think about agreement data in AI contexts.
The StackOne pattern is the one to watch. Glean as the context/control plane, a separate layer executing the actions. That’s a plausible shape for how agent platforms and systems of record divide responsibility — and we sit on the system-of-record side of that line.
Best demo of the day: Enterpret detected from social chatter that a bathroom latch was broken in one retail store, and the system automatically emailed that store’s manager to fix it. Small, concrete, and a clean illustration of unstructured signal → specific action.
Get freshly brewed hot takes on Product and Investing directly to your inbox!
Leave a Reply