Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Sunday, October 4, 2026

Back to today

British American Tobacco: Harvest

British American Tobacco’s U.S. reset revives combustibles and rebuilds its nicotine-pouch portfolio after legacy-brand writedowns and Velo 1.0’s failure.

BTI $121B
Invariant 5 clicks

From X

@convequity 3 clicks

Convequity on X

What $MNDY can learn from Cursor’s breakout and $PLTR’s head start General productivity AI may still be waiting for the foundation-model capabilities that helped coding applications take off. That leaves a possible breakout ahead for MNDY. But waiting for the next model release would give its competitors the same opportunity. Our latest SaaS Survival Scorecard explores what MNDY could build in the meantime. During 2023 and early 2024, Cursor had already built its coding editor and an early harness around models such as GPT-4 and GPT-4 Turbo. That surrounding software supplied codebase context and helped developers apply and review changes. But the models still struggled to perform complex coding work reliably. Cursor had built a way to deliver AI assistance, while the intelligence available through it limited how much work developers could delegate. The turning point came with Claude 3.5 Sonnet in June 2024, which Cursor made its default model the following month. Its stronger coding capabilities made the application Cursor had already built much more valuable to developers. Subsequent releases, including Claude 3.7 Sonnet and Sonnet 4, extended what it could do. General productivity could experience a similar improvement. Coding, however, has an advantage that helps explain why it advanced earlier: verifiability. Code can be run. Tests can check whether it produces an expected result. A training system can try different approaches and reward the ones that pass. This gives model developers repeatable feedback about what worked. Work management combines verifiable actions with decisions that are much harder to judge. Did an agent send the document to the correct approver? That can be checked. Should the approval have been requested in the first place? Should one customer’s urgent request interrupt another team’s project? The answer may depend on commitments, exceptions and relationships that nobody has written down. The result may only become clear weeks later. Rewarding an agent for closing tasks quickly could even encourage it to close work before the customer’s problem is resolved. This helps explain why narrower AI-native applications have progressed faster. Harvey focuses on legal work. Sierra focuses on customer interactions. Their narrower scope makes it easier to define the job, connect the necessary systems and assess whether the customer received value. Legal interpretation and customer disputes still require judgment. The advantage is a more specific problem than “help everyone across this organisation work more effectively”. MNDY therefore has work to do before foundation models improve. It could capture more of the knowledge that managers use to coordinate people and departments, then turn that knowledge into rules and records an agent can check. For example, when a manager repeatedly overrides an assignment, the agent could ask why. When a project requires an extra approval, it could identify the condition that triggered it and ask the responsible person to confirm the rule. Over time, MNDY could connect those rules to named customers, documents, dependencies and permitted actions. It could then check whether the handoff was accepted, whether the work was reopened and whether the intended outcome occurred. That would improve the information available to today’s models and create better examples for evaluations and eventual post-training. $PLTR shows how valuable this preparation can become. Its forward-deployed engineers have spent years working directly with customers and the people who use its software. Those conversations reveal how work actually gets done, including exceptions that a standard process diagram leaves out. Consider a manufacturer waiting for a component. One customer can accept a substitute. Another requires a quality approval. A third faces a delivery penalty that makes the delay much more expensive. An engineer working with those users can discover these differences and build them into the software. This means the AI agent has explicit rules and relationships to follow instead of having to guess from a collection of documents. PLTR’s ontology also gives data from different systems a common vocabulary. The agent can recognise that an account in one application and a customer in another refer to the same organisation. This accumulated work helps explain why PLTR could make earlier, less capable models productive when it introduced AIP in 2023. It also gives PLTR an advantage that is difficult to reproduce quickly. MNDY’s off-the-shelf approach - as with typical SaaS - has historically placed more responsibility on customers to describe their processes through boards and automations. That makes the product easier to distribute, but leaves much of the unwritten knowledge in employees’ heads. PLTR’s customer-specific implementations have captured more of that knowledge through direct engineering work. MNDY’s opportunity is to find a cheaper, repeatable way to capture it across its installed base. Its existing harness work is a start. Context, memory, tool controls and evaluations take it beyond basic retrieval, which is why our report scores its model-layer integration at 2.5. The larger opportunity is to make more customer workflows explicit and verifiable. That could improve agent performance before the next model breakthrough arrives. Better foundation models would then operate with a maintained understanding of how each customer works. If customers depend on that understanding to complete work reliably, MNDY would have a stronger reason to retain their business than access to a model its competitors can also buy. The full MNDY report develops this analysis within our SaaS Survival Scorecard. #AgenticAI #SaaS

Investor Letters

All letters

Nike, Part Two: The Deep Dive

Nike’s second Latticework program examines the company and the global athletic footwear and apparel industry for professional investors.

NKE $51B
Latticework by MOI Global 1 click

Nike, Part One: The Story

Nike grew from a $1,000 shoe-distribution venture in 1964 into the largest sports brand, using athlete endorsements, product innovation and…

NKE $51B
Latticework by MOI Global 2 clicks
GPK · long

Graphic Packaging Holding CO

Graphic Packaging (GPK) is pitched as a long with $22 fair value, implying roughly 100% upside, based on its low-cost paper-based packaging…

GPK $2.5B
Value Investors Club 2 clicks

What I've Been Reading

Yesway, Dollar General and Expedia feature in a research roundup spanning convenience retail, discount stores and AI-driven travel search.

YSWY $1.4BDG $27B
Stock Thoughts

Everyone Bought the Shovels

Element Solutions sells consumable chipmaking chemistry, with a $75 target from $44 based on an AI-driven semiconductor-materials rerating.

ESI $8.9B
Mispriced Assets

Maker of Men

James B. Duke built American Tobacco into a monopoly, then extended his legacy through Duke Power and the Duke Endowment.

BTI $121B
Invariant

Under The Hood - Rollins ($ROL)

Rollins ($ROL) pairs automatically renewing pest-control contracts with route density and family-business acquisitions, producing 24…

ROL $16B
Bearhold Research

Publications we follow

Every publication on this site, hand-picked.

128 publications

Newsletters 106

YouTube 4

Blogs 17

Seeking Alpha 47 of 223

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

84 accounts
Nominate an account

Investor letters we follow

189 firms

Sectors

Ordered by how often our editor reaches for them.