How to Integrate an AI Agent into the Sales Process

TL;DR: Identifying areas in the process where AI can be useful, selecting the right tool, implementing and training the agent, measuring and optimizing the agent’s performance are the key points to consider.

Agents using LLMs (Large Language Models) are on the rise: OpenAI’s ChatGPT, Anthropic’s Claude, and Meta’s LLaMA are just a few examples.

However, adapting them to the daily Sales dynamics might seem a bit complex. If you haven’t yet implemented generative artificial intelligence in your sales process, this article is for you.

Let’s get started 😁

Where Could an AI Agent Be Useful?

Before answering this question, you should go through the following filter:

1. What problem do you want to solve? If you don’t know what you’re facing, it will be hard to move forward with a proposal. You need to take the time to understand the problem and know exactly where you could use artificial intelligence. A good strategy to better understand the problem is to use first principles thinking.
2. Is an AI agent the best solution to that problem? Many times, we try to kill a fly with a bazooka. Of course, artificial intelligence is very powerful, but if you’re looking to solve something that can achieve similar results at a lower cost, then it stops making sense.
3. Do you have the necessary data? When training the agent, you need various data, such as documentation about your company, catalogs, or any other information you’d like it to respond with.
4. How should the agent be integrated into the sales process? If you have a CRM, how should it interact with it? The same goes for any other tool. What would the agent’s role be in your workflow?
5. What cost would it have, and what would be the return? It’s VITAL to understand this part. You want to solve a problem but not bankrupt yourself in the process. It’s important that the agent’s gains exceed its costs, or at least that it breaks even.
6. How will you measure success? If the AI agent succeeded in the sales process, how would that be materialized?

Now, let’s answer each question with a hypothetical case:

• (Problem) Today, we’re receiving about 5,000 leads per month on WhatsApp. Our Sales team only manages to talk to 15% (750), and of those, only 20% are qualified (150). This is because we don’t have time to talk to everyone.
• (Is AI the solution?) We believe an AI agent could help us handle the total demand and free up our salespeople to meet with prospects and close deals.
• (Data) For this, it’s necessary to train the agent with our entire sales catalog, information about our company, and website. Additionally, we need to provide it with the most important questions to understand if a lead is qualified or not.
• (Integrations) We use HubSpot, so the agent needs to connect to the CRM to create a contact, a company, and, if the lead is qualified, a deal in the sales pipeline.
• (Cost and return) We have 4 salespeople, each earning about USD 1,500 per month, not counting commissions. If we want to hire enough staff to handle all that demand, we’d need to bring in about 20 more people (at least an additional USD 30,000 per month). With an AI tool, we could handle everything at 10% of the monthly cost (USD 3,000), and if we maintain conversion rates, generate about 1,000 qualified leads. Our conversion rate to sales is 10% (100 of the 1,000 leads), and our net ACV (annual contract value) is USD 3,000, so we’d be closing USD 300,000 net annually per month. With an annual cost of USD 36,000 for AI and USD 72,000 for the 4 salespeople, we’d have an ROI of ~278%.
• (Success) We’ll consider the test successful if the AI can maintain the rate of qualified leads over the total demand.

Not sure where to start? Don’t worry, here are some examples:

How to Create an AI Agent?

Patagon AI

With our solution, you could deploy a lead qualification agent on WhatsApp and webchat that:

1. You can train with any documentation you have or integrate it with an API to extract information. This is known as RAG (Retrieval Augmented Generation), which basically equips the agent with all the necessary information about your company to respond to potential customers.
2. You can integrate with the CRM of your choice. We currently support plug-and-play tools like HubSpot, Salesforce, Zoho, Microsoft Dynamics, Zendesk Sell, Pipedrive, and a few more. However, if you have another CRM not on this list, we can integrate it too.
3. You connect it to WhatsApp or webchat quickly.
4. You can see the agent’s performance on our dashboard. For example, how many conversations entered, how many of those conversations provided data, how many leads scheduled a meeting or appointment with your sales team, etc.

Here’s a small demo:

If you’re interested, you can contact Mariano Rey or visit our website.

Zapier

With the Chatbots functionality, you can create your own AI agent for different sales use cases. However, it comes with its own pros and cons:

Pros

• You can easily create agents and integrate them with other tools.
• You can do the prompt engineering and fine-tuning of the agent.
• No coding is required.

Cons

• You must pay for the Professional plan, at a minimum.
• The space and context for training the agent are very limited (which has a direct impact on the agent’s performance).
• It does not connect to WhatsApp; it’s webchat only.

If You Have Access to a Development Team

If you have the available resources, you can directly use the APIs available in the market, such as OpenAI’s Assistants API, Anthropic’s Anthropic API, or Meta’s LLaMA 3.

It’s important to note that deploying an agent requires time, as there are many important processes to consider, such as how RAG will be implemented, which model to use, how prompt engineering and fine-tuning will be done, etc.

How to Measure an Agent’s Results?

This is simple enough if the analysis explained above is done correctly. If we know what success means, then we can compare.

Ideally, you should run an A/B test considering:

1. 50% of the traffic goes to the current experience, and the remaining 50% goes to the agent.
2. In both cases, you can measure the main KPI and attribute it to each experience. For example: the human attention rate was 15%, and of those, 20% were qualified. The agent’s attention rate was 100%, and of those, 20% were qualified.
3. To conclude the experiment, you should have statistically significant data. This means that the likelihood of making a wrong decision at the end of an experiment is relatively low. This is especially true if you’re looking for differences in conversion variables (for example, improving the conversion rate by X% against the current benchmark).
4. Know how to proceed if the agent’s performance exceeds the current experience. You might ask yourself, for example, how should the agent be scaled? How should the Sales teams be trained to work alongside the agent?

And that’s all, folks!

Implementing a generative AI agent in the sales process can bring benefits not only in terms of costs but also in performance.

If you understand the problem your company is facing and think AI is the best way to solve it, then you should get to work choosing the right tools, testing, and evaluating. This technology is here to stay, and those who don’t adapt will see how competitors who are using it improve over time, extending their competitive advantage and accelerating their growth.

‍