5 Things That Are Actually Working and 5 Things That Aren’t in B2B SaaS AI with Ironclad’s CEO and a16z
Ironclad CEO and co-founder Jason Boehmig joined Seema Amble, Partner at Andreessen Horowitz at SaaStr Annual to share their observations on what’s currently working – and what’s not quite there yet – for Artificial Intelligence (AI) in SaaS.
With Ironclad’s journey from an AI-first concept in 2014 to a Series E+ company and a16z’s extensive portfolio view, their insights offer a valuable perspective on the current state of AI in SaaS.
What’s Currently Working in AI for SaaS
1. AI-powered Customer Onboarding and Customer Support
We’re seeing a lot of companies using either AI tools they built themselves or through 3rd-party AI SaaS vendors have success helping users understand a new behavior or use a product for the first time.
Everything from the demo all the way through support, onboarding and implementation can benefit from AI assistance. AI is proving particularly effective at assisting customer onboarding and support in the following areas:
Guide users through first-time product experiences
Automate implementation
Reduce professional services costs
Discover user preferences automatically
Scale customer support efficiently
As Jason notes from Ironclad’s experience: “What drives me crazy as a founder is the effort required to get a customer set up and successfully using the platform. AI can help discover preferences in contracts, understand how legal teams like to negotiate and implement that automatically in the software.
I’m extraordinarily excited about this, because it can help expand the category and can help expand the number of customers that we can successfully serve and bring down our cost of service.”
Implementation and getting your customers live on your products as fast as possible is a journey that many SaaS companies struggle with. So if you can lessen the burden on yourselves and get the help of an AI that has learned from many of your other customers and that customer itself, why not use it?
2. Automating Basic Manual Work
The sweet spot for AI automation currently lies in streamlining routine tasks that previously required significant manual effort. Success stories include:
Transcription and note-taking (e.g., Zoom summaries)
Email copy generation (e.g., MailChimp)
Document search and analysis
Form filling and data entry
This is particularly impactful in vertical SaaS, where industry-specific manual processes can be automated to reduce staffing needs while maintaining quality.
3. Staying on the Main Branch for Core Model Capabilities
Rather than building custom AI models from scratch, companies are finding success by:
Using base models (like GPT-4) with fine-tuning
Adding RAG (Retrieval-Augmented Generation) capabilities
Maintaining access to rapid improvements in foundation models
“I think what’s interesting and what we found that maybe I didn’t expect is staying on the main branch, staying on the core capabilities, and supplementing it with fine tuning and additional techniques like RAG. For us, GPT-4.0, plus some fine-tuning, plus some RAG on a Delaware case law dataset performed much better than custom models,” shares Jason. “Staying on that main branch has a lot of benefits because you’re getting the rapid improvements to the foundational models as they’re delivered.”
4. Leveraging Vertical Data
This is where companies are successfully figuring out what their vertical concepts are that can benefit the most from automation and AI, and then working with AI to understand those core concepts. One example from Ironclad, but something we all experience as a SaaS company is contract negotiation. Oftentimes in the redlining process companies can get stuck if each one has a particular template they like and use as standard and it can be costly and take a lot of time to morph these two together into one agreed-upon final contract.
This is where AI can come in.
Instead of just comparing the two documents or templates, it can look for the root concepts and identify how different the concepts are on each document. So if AI understands the vertical concept, what it can do in this example is provide what’s called a conceptual red line, which instead compares the positions in two documents to better help attorneys negotiate contracts much faster.
Outside of this example, many SaaS companies are already finding success by deeply understanding and implementing vertical-specific concepts rather than trying to build custom models. This approach is delivering real business value:
Higher accuracy rates in specific domains
Better understanding of industry-specific concepts
More valuable outputs for specialized use cases
5. Founder-Driven AI Innovation
Seema Amble, Partner at Andreessen Horowitz shared that something they’re seeing across their portfolio of founders, especially the ones launching an AI product or a standalone product is boost their founder-level engagement on the human side of AI.
Seema explains, “You really do need founder-level engagement and drive to push the product forward and make AI a center of the strategy because every SaaS company today is an AI company. So if you’re going to differentiate yourself in terms of your AI native product, you need the founder-driven energy just as much.”
This focused, founder-led approach in AI helps:
Drive rapid innovation
Overcome organizational resistance
Bridge the gap between core product and AI capabilities
What’s Not Working in AI for SaaS (Yet)
1. Going Broad to Narrow
Companies starting with broad, horizontal AI applications are finding they need to narrow their focus.
At a16z Seema is seeing a lot of companies who are building, AI-first SaaS applications start to narrow down either on a use case or a market versus being more horizontal. One example is there’s been a boom of workflow automation tools coming to market in the last year since it’s a very tangible way to use LLMs to power SaaS. Workflow automation could tackle anything from filling out forms to documentation extraction, meeting summaries, creating copy, etc.
a16z saw several workflow automation companies that all started on a very horizontal basis and, over time narrowed it down to serving just specific industries. Why? Getting into the AI integrations within one specific vertical and understanding those users more completely and exactly how they interact with the product is more effective for going-to-market than broad strokes.
2. Relying on Self-Serve Sales Structure for Enterprise
Not everything needs to be a PLG product.
Despite the PLG (Product-Led Growth) trend, self-serve models for AI-powered enterprise products are proving challenging. Jason explains: “If I could give myself advice three years ago, it would be to go figure out a way to sell more million ARR repeatable deals. That’s your path to efficiency, as opposed to trying to tune the unit economics exactly right on PLG. And I think this is playing out in future AI products as well. The self-serve enterprise can be tough versus humans leaning into the bigger deals.”
3. Fully Automating Roles and Agents
Complete automation of complex roles isn’t yet feasible. We recently wrote a post using SaaStr survey data that reported how 83% Percent of You Haven’t Gotten AI SDRs to Work … Yet.
Instead, the winning approach is human-AI collaboration:
– Trained professionals assisted by AI tools
– Human oversight for complex operations
– Quality control for critical outputs
Let’s use the sales agent example. No matter what interactions you’ve had with an AI agent in sales, would you rather have a fully AI sales rep or would you rather have a trained sales leader who’s assisted by the best AI sales intelligence in the world? Cost aside, we think most companies would choose the latter. And this combination of human and machine will beat the machine only.
4. Realing on Quality with Complex Operations
For more complex operations, you still need a human in the loop driving the operation right now. For example, right now in AI-generated images, there are still cases where an image includes a third arm on the human or a sixth finger, or completely incorrect text. No brand manager would think of approving an AI-generated image. However, in some tools, those hallucinations around appendages and the quality of images has gotten a lot better already.
Specifically when it comes to more complex sequences of operations where you can’t afford to have the hallucinations or a wrong answer, like in legal cases or historical context – the AI just isn’t quite there yet.
5. Pricing and Monetization
The industry hasn’t yet settled on a dominant pricing model for AI capabilities, making this an area of ongoing experimentation. SaaS companies are still experimenting with various pricing models, including:
Core product integration
Premium upgrade tiers
Add-on modules
Outcome-based pricing
Key Takeaways
While AI is delivering real value in specific use cases, particularly in augmenting human capabilities and automating routine tasks, we’re still in the early days for many applications. Success currently lies in focused applications with clear use cases, strong vertical understanding, and thoughtful human-AI collaboration rather than full automation.
As Jason concluded, “We’re at the peak of inflated lofty expectations in AI… what we got to do to get out of the trough is find the repeatable value in AI.” Companies that can identify and execute on these specific value propositions while maintaining appropriate human oversight are seeing the strongest results.