Agentic Research & Implementation Case Study

An overview of the steps taken and methods used to gather, synthesize, and then find ways to illustrate implementation of findings.

Medium and Message

Bland AI, an AI Voice Telemarketing code-free tool for designing and deploying campaigns through complex workflows. The focus is on home services industry.

Bland AI Platform Exploration

Some time was spend exploring the platform to identify what it required to run a campaign. It was clear that the platform took plain text prompts, often scripts looking for AI generative completion. There were many voices and you could create ‘Personas’ to use in the campaign that represented an AI that has been crafted to speak in a certain way and with a certain voice.

It wasn’t necessary to go much deeper to see that the platform settings weren’t grouped according to KPI or conversion goals. This wasn’t surprising but did confirm the direction of the research.

1. Deep Research Agents

Identifying what we needed to know first was relatively simple: Data-backed research that was current and defined the top strategies for AI voice marketing in the home services industry.

The scope of the research was defined for the agents. The context of our task was also defined so that they could make decisions as to how to craft additional queries based on initial results.

  • Query the Perplexity AI and assess the results
  • Compose refined queries, drilling down and asking for a matrix of results wherever possible
  • Continue for six iterations, refining the queries and drilling down on the most relevant strategies

2. Research Synthesis Agents

This phase was also left to the agents. They were to identify what commonalities existed between the strategies across the research and compose overview documents that could be used to inform what type of deeper research overviews could be created, as well as identifying where more research was needed.

This resulted in a handful of overviews of strategies that were paired with various demographic groups. They followed some particularly interesting patters such as how to handle objections and how to use topical information like the current economic climate.

3. Targeted Research Agents: Voice Engineering & Bland AI Implementation

Having seen the Bland AI platform and reviewed the type of information provided in the strategy research, it became clear that there were two main questions that we would need to answer to make any finding into tangible information. The first question (1) was simple and conceptual, very much like most marketing. The second question (2) was rather technical and required deeper insights into the platform.

  1. What do we talk about?
    • What are the most effective strategies for the medium?
    • What are strategies for achieving specific goals?
  2. How do we talk about it?
    • What will the creative work look like?
    • What kind of technical knowledge is required?

The focus at this point was much more focused on that second question. It is one thing to have a “human sounding voice” but the strategies gathered so far would require more a nuanced control over the voice.

Armed with this information, the agents were tasked doing some more digging and then creating additional synthesis documents showing us the top level view of what kind of information could be gleaned from the research.

First, on the degree of voice control we’d need to achieve:

Second, on the technical implementation of the voice in the Bland AI platform:

4. Framework Defining Agents: Detailed Voice Marketing Strategy

Now that it seemed like we had what we’d need, we needed to break down the findings into a framework that could be used to create a persona. This required a few stages. Each stage we constructed in a way that would provide useful research findings documents. We wanted them to be more actionable and useful than the original deep research synthesis.

Top Five Conversion Strategies:

  1. Assertive-Urgent Voice Marketing
  2. The Curiosity Gap Approach in Voice Marketing
  3. Empathy-Based Objection Handling in Voice Marketing
  4. Demographic-Tailored, Topical Combinations Framework
  5. Emotional Trigger Optimization Framework

These were then synthesized into a framework that could be used to create a persona.

5. Voice Notation & Scripting Agents

We had frameworks for the strategies. In the research, each strategy defined the type of voice patterns and modulation that would be used. Now we needed to know how to write that information into a script. This would allow us to further define our frameworks into personas.

6. Persona Creation Agents

Time for some creative writing. This is where we would be able to finally make all of the research finding into something tangible and actionable. First we armed the agents with our guidelines for marketing copywriting.

Next the created archetypes that allowed them to define personas that included description of voice including things like tone, pitch, and modulation. We had them create two personas for each strategy.

7. Implementation Agents: Communicating the Strategy

We had some very actionable documents, but they could be pushed further. We wanted some kind of one-page summary that could be used to implement the notation writing strategies. And then we them to use the script writing guide and along with our crafted personas, and create actual scripts.

Finally we had taken the strategy, pulled out what very human elements, and gave them a framework. This was done through archetypes: a persona. A character that happened to speak a certain way. Including how to make the AI talk in that certain way.

  • Pauses: Short, Medium, Long, Paragraph Break
  • Emphasis: ALL CAPS, Italics, Bold
  • Tone Variations: Friendly, Serious, Excited, Concerned
  • Pace Variations: Fast, Slow
  • Pitch Control: Low or 130Hz, High or 180Hz
  • Inflection Patterns: Rising, Falling, Questioning

Indexing The Findings

All of the research and implementation findings were indexed to make as useful as possible at this stage.

Conclusions

Hella Cheap, Hella Fast

I basically worked from when we ended our call on Tuesday afternoon until now, Wednesday evening. So much of it is once and done. The rest is an enjoyable amount of creative work.

Cost

$ 4.61 for all research findings on:

  • Best-practice strategy for AI voice telemarketing for
  • Home services industry focused; electrical and plumbing
  • Persona archetypes for voice telephony
  • Script writing guide for voice telephony
  • Voice modulation techniques for Bland AI
  • Script examples for each persona in each strategy
  • Technical guide for using Bland AI

Maybe about $ 10 spent while doing testing and adjusting what kind of output I wanted from the agents.

Time Running the Workflow

~6 hours of time fully focused on copywriting for the prompts.

  • 1-2 hours prompt writing for research and synthesis
  • 3-4 hours prompt writing for implementation

Time Building Agents

~12 hours getting bugs out of new tools to handle file size.

  • No move development time until more tools are added
  • Planning on spending time planning logic for branching workflows

Quality

I still cannot get over the quality of the output from the agents. Insightful. AI is just so comprehensive and makes connections that humans literally can’t which makes them forever beyond as creative as a human could be.

The most time I spend on the prompts only increased the quality of the output. They’re like little actors and you just set a scene. You should read some of the prompts in the code. Claude was a Hollywood movie star “using what I’ve learned from my acting career to better the quality of the character in the AI Voice scripts.” No joke. Insane.

Overview

So there is a lot more value to be found and the bulk of the hard work is done. And then of course the ideas going forward are endless. I had one of the agents pull all the API information from Bland AI because imagine this:

  • Agents find topical conversation information
  • Agents keep all strategy up to date and adjust based on testing
  • Agents write endless scripts
  • The same agents are put in workflows that work with the features in the tool, making the calls, gathering the information on leads, etc.
  • Ever looking for next steps you just have one of them analyze everything and come up with best next steps for X goal