1 of 35

GenAI, SaaS, & You

Jake Saper

@jakesaper

April 2023

Confidential — Do Not Distribute

1

2 of 35

We invest in people who change the way the world works.

Confidential — Do Not Distribute

2

3 of 35

Confidential — Do Not Distribute

3

4 of 35

Background
Where are we today?
What should you build/invest in? What’s defensible?
How should you build with GenAI? How can risks be mitigated?

Goals for Today

Confidential — Do Not Distribute

4

5 of 35

What is a Large Language Model?

Confidential — Do Not Distribute

5

A large language model (LLM) is a type of artificial intelligence (AI) program designed to understand and generate human language. It uses deep learning techniques to analyze vast amounts of text data and learn the patterns and structures of human language. It uses these patterns to predict next words/phrases.

GPT stands for "Generative Pretrained Transformer." It is a type of large language model (LLM) developed by OpenAI, which uses deep learning techniques to generate human-like language. The GPT model is based on the transformer architecture, which was first introduced in a 2017 paper by Vaswani et al. The transformer architecture is designed to process sequential data, such as language, in parallel and has proven to be highly effective for natural language processing tasks.

GPT models are behind products like ChatGPT, which has had amongst the fastest growth curves in tech history.

6 of 35

ChatGPT has Mainstream Pull

Confidential — Do Not Distribute

6

7 of 35

Emerging GenAI Landscape

Confidential — Do Not Distribute

7

3 layers

Layer 1: Foundational models like OpenAI’s GPT suite. Databricks released Dolly 2 as an opensource model. Emerging split here is between closed and open foundational models. Can get in to this in a moment
Layer 2: Picks and shovels to help companies make better use of these models.

Open source libraries like LangChain which help folks chain models together to add memory or state and accomplish tasks
Vector databases like Chroma and Pinecone to help companies integrate their own data into these models
Prompt management systems like HumanLoop which help companies engineer and track prompts

Layer 3: App layer

Horizontal companies integrating this tech into their platform. Zoom launched ZoomIQ which does smart meeting transcription
Specialized companies deploying this tech to accomplish more narrow jobs to be done.

GitHub Copilot in coding
Ironclad in legal contracting, which we’ll discuss later

Closed vs open foundational models is an emerging split in the ecosystem

Considerations:

Security/privacy

Samsung employees punished for uploaded proprietary data to ChatGPT

Ability to customize to your needs

High value, domain-specific workflows likely to rely more on specialized systems vs general AI models

Speed/latency
Cost

To build/maintain yourself

Some companies with existing ML teams are notably slow to adopt latest third party advances

To rent

Pricing models for foundational models still evolving. Will have implications on pricing models at app layer

Focus today’s convo largely at the app layer

8 of 35

Current State of Startup Market

Confidential — Do Not Distribute

8

9 of 35

New Investments:

Applying AI to Customer Problems

AI Value Creation in B2B

Enabling AI Adoption

Model Providers

Infrastructure & Tooling

Enhanced SaaS Products

Services Automation

Spotlight: Our Generative AI Theses

Years of AI investing experience enables us to cut through the noise and �identify differentiated opportunities

5

DONE

Confidential — Do Not Distribute

9

Lots of hype and confusion around the role AI will play in business. Let’s start with first principles:
There are two ways businesses can take advantage of AI [Click]:

Either by developing and selling products that enable AI adoption or
Directly applying this technology to solve customer problems

The most funded and hyped businesses developing products to enable AI adoption are, of course, foundational AI model providers like OpenAI [Click]. We have decided not to prioritize this segment both because we believe it will continue to commoditize in part thanks to improvements in open source models and because we believe the capital required to support compute requirements will result in significant dilution.
[Click] The category of AI enablement companies we’re more excited about are the picks and shovels that help businesses deploy these models
[Click] Of companies applying AI to solve customer problems, SaaS companies making use of this tech are the best known and are a prime focus for us.
[Click] We’re also exploring a much nichier but potentially alpha creating opportunity: services companies making use of AI to productize. We think there may be an underhyped opportunity to build terrific businesses with this strategy and will share more in a few slides.
[Click] Each investment we’ve made over the past year plays in to one of these areas which we’ll dive deeper into in our section on Fund VI.

10 of 35

Training

Fine-Tuning

Open Source

Proprietary

Closed Models via API

Compute

FOUNDRY

Hosting / Inference

Closed models like OpenAI’s GPT-4 are best used by enterprises for use cases including rapid experimentation and high volume, generic use cases

DONE

Open models can be fine-tuned with industry specific or proprietary data and augmented with RAG to solve the majority of use cases, primarily targeting higher value

Emergence is focused on the enabling development infrastructure and tooling that companies will need to leverage open and proprietary models

Proprietary IP

Accuracy

Customization

Compliance

Security

Models

Compute

Development

Expertise Required

Time to Value

Cost

Tooling

AI-Development Infrastructure and Tooling

Confidential — Do Not Distribute

10

11 of 35

GenAI-Enhanced SaaS Applications

Key Question:

How much value is the app adding beyond the LLM?

Our Approach: Leverage our Unique Assets

SaaS Workflow: Companies building robust workflow on top of LLMs to solve complete job-to-be done best positioned

Coaching Networks (Co-pilot): Capturing outcomes data will be critical to build moats over time

Industry Cloud: Deep dive on specific verticals, including healthcare and fin serv

DONE

Confidential — Do Not Distribute

11

Underlying AI models are available for any app provider to integrate, so the AI itself is, generally speaking, not a meaningful differentiation in apps at this point.
Over past year, have seen lots of companies that integrate these models (often leaning more prosumer) grow quickly providing things like marketing copy generation. But they’re having retention issues as users tire kick. We’re looking closely at retention this year.
Creating applications has become easier but understanding and solving customer problems is still hard. Differentiation comes from doing the hard work of successfully doing the job your customer needs done, which most often requires much more than just an LLM.
[Click] Thus, the key question we’re asking when evaluating GenAI-enabled apps is: how much of the “job to be done” can be done by going straight to LLM? Or, put differently, how much value is the app adding beyond the LLM?
To be defensible over time, we think there needs to be lots of value on top. Will often be in the form of good old fashioned workflow. Fortunately, we have a lot of expertise in building these types of businesses.
[Click] Workflow also has the benefit of keeping people in the app and gathering proprietary business outcomes data. A number of our more mature companies are doing this well: Ironclad, Doximity. And newer investments like Unify in the sales enablement space are also doing this. While there are many startups using LLMs to draft sales emails. Unify provides this commodity service but is focused on building deep data connectors and email cadencing workflow on top of the LLMs.
As you may well recall from our Coaching Networks thesis, we believe that the data these workflow tools acquire will be the critical angle to build proprietary AI over time.
We’re seeing a number of exciting opportunities in particular around applications of AI in industry cloud, including financial services and healthcare.
[Click] We’ve published some deep research on what GenAI opportunities might be particularly exciting in those sectors. There’s too much to discuss here but feel free to go to our website to read the details or pull any of us aside.

12 of 35

GenAI-Enabled Services

Massive, uncrowded opportunity

Why now is GenAI

Focus on high repeatability, execution-focused jobs to be done

Mechanical Orchard

DONE

Confidential — Do Not Distribute

12

While the vast majority of founder and VC energy continues to focus on the AI software market, we think there’s a compelling opportunity that AI is unlocking to disrupt services businesses.
[Click] These are massive markets that have seen relatively little disruption over the years. The Big 4 accounting and professional firms do $200B in annual revenue. BPOs do $250B. Traditionally, it’s been hard to build high margin, recurring revenue businesses in these markets.
But it has been done. In fact, Palantir is amongst the top 5 most value B2B companies on a multiples basis. They leveraged a services wedge to sell software over time.
[Click] We think AI could enable a Cambrian explosion of Palantirs.
[Click] We’re focused on services with 2 key characteristics:

1) They are highly repeatable, and therefore productizable by AI
2) They lean more heavily on execution vs advisory, and thus trade less on squishy qualitative advice and more on objective outcomes.

[Click] One such opportunity is in the mainframe migration space. Helping big enterprises migrate core workloads from mainframes into the cloud is a large market with sleepy, tech backwards incumbents. AI can be used to help map legacy code and then pair-program to recreate it in a cloud native format.
That’s exactly what one of our most recent investments, Mechanical Orchard, is doing. Excited to share more in our Fund VI update.
Next, excited to bring up our COO Julie Bell to share an update on the firm.

13 of 35

What should you build with GenAI?

13

14 of 35

Something people desperately need

What’s your unique take?

14

15 of 35

If everyone can access these APIs, what is defensible?

15

16 of 35

Core Defensibility Question:

What portion of the “Job to be Done” can be done mostly/entirely with off-the-shelf LLMs?

16

17 of 35

Example JTBD: Legal Contracting

Confidential — Do Not Distribute

17

18 of 35

In addition to AI, JTBD requires complex workflow, permissioning/approvals, data integrations, etc

These good old fashioned building blocks of SaaS will likely prove to be the scaffolding around which defensible generative AI-enabled businesses are built

Confidential — Do Not Distribute

18

19 of 35

What can/will incumbents do?

What is the opportunity for startups?

19

20 of 35

Contextualizing GenAI

Confidential — Do Not Distribute

20

21 of 35

Identifying Areas Startups May Have an

Advantage Relative to Incumbents

21

Some thoughts on another potential innovator’s dilemma: pricing/revenue model

What new JTBD will be created?

03

Infrastructure for AI
Vanta for AI

What JTBD couldn’t have been done with previous tech?

02

New category creation opportunity
What was the domain of consultants/ professional services (e.g., pricing strategy)?

Where are incumbents not focused?

01

Narrower/more specialized data sets more likely to be proprietary
Verticals (e.g., Veeva)
Horizontal niches (e.g., Salesloft)

What JTBD benefit from a different UX/UI?

04

New UX/UI paradigms

Chat vs Point and Click
Coaching/Co-pilot

Innovator’s dilemma (e.g., Salesforce)

22 of 35

Building with GenAI:

Risks and Guardrails

22

23 of 35

Is Generative AI in danger of having its own FTX moment?

23

24 of 35

Growth of Auto-GPT: 🤯

Play with Auto-GPT here. But be careful…

Confidential — Do Not Distribute

24

25 of 35

How Can You Mitigate the Risk of a Bad AI Trip?

Confidential — Do Not Distribute

25

26 of 35

More on why most GenAI today isn’t tuned for B2B

Confidential — Do Not Distribute

26

Start by understanding your JTBD
Can think about them along two dimensions: how important is accuracy. And how important is being tied to real-world outcomes
Many consumer use cases require little accuracy or tied to the real-world. Chatting with a dead celebrity, enabled by companies like Character.ai, doesn’t have a “right” answer”. And by definition, there’s no tie to outcomes in the real world.
Some of the earlier B2B use cases require less accuracy, though need to be tied to the real world. Copy editing companies like Jasper don’t need to generate a “right” answer. There isn’t one. But they do need to tie to outcomes- e.g., did the target audience by the product when this copy was used
But most B2B use cases require both accuracy and outcomes-tie.
Given the tendency for GenAI to hallucinate, how can it be used responsibly in a high accuracy context?

27 of 35

Coaching Networks use AI to coach workers on how to do their jobs better, as they do them.

27

28 of 35

How Coaching Networks Work

28

Moving from Generative AI to Iterative AI

29 of 35

How Can We Ensure Worker Stays Engaged with Coach?

29

“It’s a growth opportunity, but that’s what’s keeping me awake at night—how to use them as a co-pilot, not an autopilot”

Melissa Werneck, Global Chief People Officer of Kraft Heinz Co

WSJ: April 13th, 2023

30 of 35

30

31 of 35

Outcomes data will be the foundation for longer-term moats

Confidential — Do Not Distribute

31

32 of 35

Takeaways:

How to Build with GenAI

32

Identify relative importance of accuracy for your use case

If high importance, consider UX that actively engages human “manager”

Capture outcomes data and build feedback loop in to model

33 of 35

33

P.S. we’re hiring

Thank you!

34 of 35

34

How to Hack Your Way into a Proprietary Data Set

35 of 35

Confidential — Do Not Distribute

35