What is the missing ingredient for AI Agent adoption?
LLM Agents are still very early in the technology adoption curve
Businesses are looking to onboard generative AI to solve specific pain points but they often don’t know where to start or which use cases to prioritize (more on that here). In light of the recent GPT-4o announcements, agents and agentic experiences are becoming increasingly popularised in the public sphere. Even so, businesses looking to onboard generative AI into their business just can’t seem to find the right match for their needs, what gives?
Although there have been some large leaps in the progression of agent technology and foundational LLMs, there’s still a wide gap in the space of applications that are relevant to businesses today.
A study by Salesforce showed that 62% of workers felt like they lacked the skills to “use [generative AI] technologies accurately and safely”.
This means there is still a significant user experience gap for agent applications to overcome before they can be considered a mainstream application. To us, it seems clear that the shortfall is more of a user experience (UX) challenge, rather than a technological one given that AI startups have raised a collective $18 billion this year alone; mostly on good traction in the B2B sector.
It seems like AI is going through growing pains as it slowly begins to work its way into the early adopter band of the technology adoption curve. However, to cross the famous gap that exists between early adopters and the early majority, as a community we have to address the user experience challenges that are plaguing the current suite of tools. Using Marc Andreessen’s analogy of AI as a type of computer, at the moment it still feels like customers have to interface with these tools using punch cards, which is unthinkable when compared to the seamless experience we take for granted with our laptops today.
Distribution for AI Agents is a big problem for builders
Even though the UX for many of today’s AI Agent applications leaves a lot to be desired, it presents a massive opportunity for innovation and ideation for agent builders to discover. While builders can focus on developing agents that can attend to a specific need, they are faced with another problem: Once you build your agent, how do you get it into the hands of your customers?
On the surface the solution seems obvious; build a SaaS and market that to your target customer profile. Unlike a traditional software business, however, some gaps need to be filled in first.
Do your customers know what they’re looking for?
The businesses that we’ve spoken to have all indicated that they understand the potential generative AI offers their business, but they’re not quite sure where or how to leverage it. This indicates a customer that is problem-unaware. This means despite your best SEO efforts, it’s unlikely that the customer will find your solution by direct search online.
If this rings a bell then you aren’t alone; the agent builders we’ve spoken to all listed distribution and customer acquisition as their main problems. Often data locality is an afterthought as they are very much focused on getting the business off the ground for the first agent use case that they’ve built out.
Can your LLM agent integrate with the customer’s existing software?
To provide the most value to the customer, your agent needs to be able to integrate nicely with their current software ecosystem. Every business uses its software stack (SAP vs Microsoft Dynamics for example). Even if your agent is generally applicable to a particular task, you might have to change your prompt or add new tool integrations to support different kinds of customers. Given the software diversity amongst businesses, there’s a high chance you have to do this often to support all the different permutations of the stack used to complete the same process (think of all the ways and different software you can use to reconcile invoices).
Can your agent work in the customer’s locality?
If the customer is in the EU or UK, then they will likely require that all their data is processed and stored in their data locality. Depending on your tech stack, this can be a tall order! Especially since many LLM API providers do not explicitly make any data residency guarantees, especially around data processing.
Businesses want generative AI but don’t know where to use it
B2B buyers rely mostly on reviews to make purchases, but right now, there isn’t anywhere they can go to compare AI solutions for their business.
A recent Gartner study shows that between 37–41% of buyers rely solely on reviews before buying new software products, and these “review-reliant” buyers are 70% more likely to make a software update as compared to their counterparts.
It’s important to note that agents present an upgrade to the way the customer currently accomplishes a task, and the main blocker of automation is simply resistance to change. Arming your potential customer’s champion (the warm lead on your site) with all the data they need to make a purchase decision is critical.
If you’ve managed to generate interest in your product and attract qualified leads who understand what they’re buying, the next step of the process is evaluation. As part of the B2B buying process, the customer’s next step will be to compare your solution against others on the market. It is important in the B2B buying process to be able to conduct a full and thorough evaluation of a purchase and its alternatives. As a business customer, they must carefully consider the reputational risk should a software purchase not pan out. More generally, we are more careful in the workplace regarding purchases that might affect the business because our credibility with our peers is at stake; this is one of the key differentiators between consumer and business buying habits.
The big question is, where can your customers go to compare and evaluate different agents / generative AI solutions for their use case? At the moment, short of reviewing the highly technical metrics from academic benchmarks like AgentBench and ChatArena, there is no straightforward way for the customer to make an apples-to-apples comparison of different agentic solutions to determine the fit for their business. As of right now, most business customers would have to either hire or sub-contract the talent that would be able to make the distinction between different agent products. Just think, if you were to google (or search on perplexity) for an agent that could summarise emails, book appointments, and enter data in a CRM, how would you compare agents from different vendors?
The AI sales agent I just described is a popular use case and is already on offer by multiple startups (regie.ai, taskade.com, and b2brocket.ai to name a few). All have different pricing and overlapping features. They likely utilize different LLMs which might be a mix of closed-source, open-source, and fine-tuned variants. As a customer, testing all these different solutions could take months, and I might fall out of the funnel entirely as my organization is already resistant to change.
How to distribute LLM-based AI Agents
Our research shows that businesses are not searching for “AI agents”, nor should they!
As builders ourselves, we’ve been trying many different methods to understand what works and what doesn’t for agent distribution and distilled them into the points below:
- Market your agent as an automation solution for a specific, niche problem within a business. Businesses have pains that they want to solve today, and generative AI is one of many possible solutions; they’re more likely to be searching for questions like “How do I automate email scraping?” or “Automatically reconcile invoices in SAP”. Businesses are struggling with manual, repetitive tasks that take up more than 20 hours per week. These tasks have been notoriously difficult to automate because most of them involve unstructured data like emails and PDFs, a perfect fit for an agent! Positioning your agent as a solution to specific problems significantly increases your chance of your solution getting found and resonating with your target customer.
- Onboard agents into a customer’s ecosystem is by providing native software integrations with their current technology stack. As part of your agent’s toolkit, it should be able to reach out and interact with the customer’s daily drivers like Excel, Quickbooks, SAP, Outlook, and many more. By integrating with their existing tools, you can piggyback off the adoption of these more widely adopted platforms to help onboard your agent into these businesses.
- If you’re using a closed-source LLM, research their data policies closely to understand how and where they process data. Businesses, in particular enterprises, will be very sensitive to this and will ask for guarantees on these policies. For instance, together.ai and OpenAI mostly process and store all data in the USA, which is fine for US-based customers but not for the EU or the UK. On the other hand, VertexAI (which supports the Google Gemini models), Azure OpenAI, and Mistral.ai provide guarantees for EU data residency and GDPR, which will let you serve customers in these localities. For those building on self-hosted, open-source LLMs, you have the much more difficult engineering challenge of ensuring that your model is deployed into the right regions and the data is processed right next to it. This can create complexities around routing, serving, and storing data from an infrastructure perspective.
How Solitude manages distribution and infrastructure for LLM-based AI Agents
The guidance above is easier said than done. On the Solitude platform, we wanted to make the lives of agent builders way easier by handling all the above for you. For agents built and posted on the Solitude marketplace, we’re developing the following feature-set:
- Managed infrastructure. Unique to Solitude, we manage all the agent infrastructure on behalf of both you and your customers to facilitate a smooth onboarding experience. Solitude hosts the agents on behalf of businesses either on our SaaS platform or in Bring-Your-Own-Cloud style deployments if the business has an on-prem requirement.
- Lead generation. We also carry out direct sales outreach to attract businesses with $10m+ in revenue who have specific pains around automation and manual data entry work and connect them directly with your agent. We will SEO optimize all the agent postings on the marketplace, and then cross-post them on sites like Upwork & Fiverr in response to job listings for workflows you indicate the agent could fulfill.
- Managed integrations. Solitude provides pre-made software integrations with common business applications like SAP, Outlook, Microsoft Office, Salesforce, Hubspot, and much more. We let you develop agents to automate specific workflows using these tools and test different planning algorithms like Tree-of-thought, React, and Graph of Thought to compare which ones work the best for different workflows.
- Seamless data residency. While you can select a preferred LLM for a particular agent/workflow, we also let you select alternatives in case LLM is not available in the customer’s data locality. We also manage all files and data uploaded by customers to our platform to ensure that they’re stored in a way that is compliant with the customer’s local data regulations.
Summary
Have you built an agent for a problem that you or your team faced and think that others would find it useful? We think you’re right! We’ve come across many amazing builders in the space who are creating immensely useful applications, but don’t have a way of getting them into the hands of customers. We built the Solitude platform for you.
If you’ve found this insightful and want to get in early on a platform that lets you get paid for your agents, signing up for our waitlist below or on our website is a great place to get an invite: https://solitude.ai
