Three Agent Personas
Next Trillion Dollar Opportunity?
Venkata Pingali (Scribble Data) & Rajesh Parikh (Cynepia Technologies)
There has been quite a bit of discussion around why AI SDR (Sales Development Representative). They are in our opinion failing (for now) because they are ignoring what we learnt about the sales organizations, process, and what makes sales teams successful. We are expecting magic. It is not the job of one agent but a structured system of agents that have different roles. Using sales agents as an example application area, we define a Three Agent Personas (TAP) Framework, explain the rationale behind them, and discuss why this presents a trillion dollar opportunity.
Understanding Early AI SDR Failures
In the real world, an SDR is onboarded in any enterprise by first taking him through a very structured training process which include organisation specific knowledge and testing him through the understanding before putting him/her to the real work. How do we then expect a generic AI SDR to come and take up the role of an individual who has been cultured and supported by all the other processes within an enterprise? SG&A are typically 20-30% of the organizational revenue indicating the kind of heavy lift that must be done. At one large insurer a new sales executive operates in a shadow mode, watching the actual sales process for at least 6 months before being put to a real job. This should highlight the need for nuanced training both on onboarding and initial phase and the support structure required while the person is doing the real job.
The picture below looks trivial until we realize the implications on the AI agent economy.
The reason AI SDRs missed this is that agents are new, and we are keen to launch a product, and hope to learn on the fly. Agents, unlike traditional deterministic software, have to operate in a human context. It didn't take very long to figure out that this doesn't work.
We need a better approach that reflects the skill and thought that goes into jobs.
Big Idea
We propose a system of Three Agent Personas (TAP) - Knowledge, Training, and Functional. Each persona could have one or more agents, or even sub-personas. These systems of interacting agents will accomplish end to end digital workforce goals with high trust.
Knowledge Agent - Aggregates external and internal knowledge and organizes them into a knowledge database. A lot of this knowledge is unstructured, multimodal, near real time, and human-generated. This agent should assess the content for relevance, accuracy, timeliness, and actionableness, and prepare a high trust source for use by training agents. We see this knowledge agent working with other similar agents within and across organizations (e.g., with a partner organization)
Training Agent - Takes the knowledge about the company, products, competitors, processes, partners, delivery model etc and further processes this information to generate training and test datasets for calibrating/fine-tuning/personalizing/onboarding generic functional agents. This agent is focused on high accuracy, repeatability, consistency, style, and quality of functional agents. Evaluation of functional agents is a key outcome of the training agent. A successful evaluation marks a successful onboarding of the functional agent
Functional Agent - Takes action in the real world such as making customer calls. They are goal driven, and have appropriate memory, interfaces, controls, and reporting capabilities. The functional agent might itself be a system of agents similar to what we find in the real world. Given the cost and risk associated with action in the real world, this functional agent will have strong policy and other controls, monitoring support, and circuit breakers. This agent will by definition be incomplete. So we should expect strong UX with support for graceful failover and support functions when it is unable to complete the job. There will be industry benchmarks for this space to give minimum guarantees of performance on the long tail of situations.
Trillion $$ Opportunity?
The system of three agent personas that we are proposing is not new. It is what we implement in real organizations today. We deconstructed the existing structure, and translated it into the agentic world. This is not just applicable to Sales but to all other functions, organizations and business processes including marketing, engineering, product development, and customer success. We are now looking not at the size of the software industry but a percentage of the economy. If this delivers 0.3% productivity improvement each year on a base of 30T$ US GDP, we will reach 1T$ easily over 10 years. We think it will be shorterbecause we underestimate the productivity improvements possible.
We think that structured approaches to agent will increase adoption and drive a change in the landscape of opportunities:
Growth due to democratization - As the agentization decreases the overall cost of software and services, more products will receive software upgrades and advanced features, and the old people-heavy processes will give way to a combination of agent+human process. Companies that couldn't afford sales teams will now be able to. As economic Nobel Laureate Ronald Coase suggested, as the transaction cost decreases, there will be more transactions (See Erik Brynjolfsson at AI Native 2024). People are not going away. We expect the human touch - which will be scarce in the agent economy - will see revival and will be a serious value add/differentiator.
Agent development and deployment (across organizations)- The opportunity arises from the fact that these agents need to be customized and optimized for every industry, organization, and business unit, and they are continuously upgraded. We expect different companies to produce each of the three persona because their nature and goals are different. There is enough diversity, complexity, value addition, and longevity to the personas to allow for an industry to grow around them.
Agent Deployment and Operations (within organizations) - Agents are more complex than traditional rule-based software systems. They are of higher value but also bring uncertainty with them. There will be new roles involving managing the data inputs and knowledge base, policies, evaluation criteria, agent networks, and in some cases deep customization. They will have titles like Forward Deployed Engineers (from the vendor side) and AI Engineers (from the customer side).
Sales TAP
Training the AI SDR Agent requires us to start with a generic Sales Agent and fine-tune its behavior to meet cultural, accuracy, style, legal and other requirements. This is similar to new employee onboarding or apprenticeship. We take a well qualified sales individual, and give the person enough organizational, market, and product knowledge and sales process context to apply the general skills they have. Then they become specialized sales staff who can sell niche products like medical devices.
The training may be for sub-roles within sales such as sales manager, and pre-sales. A specialized agent is required to do this - an agent that starts with a mass of knowledge and develops the context-sensitive training and evaluation data, process, and mechanisms. This agent again has to specialize in sales agent training. A devops training agent is very different from the sales training agent. We know this to be true from our broader experience.
This is still not enough. The world is a dynamic place - new competitors might have entered, regulation may have changed, and product strategy may have changed. As a result, the training given to meet a given set of requirements has a short shelf life. The training has to be continuous and proactive. For this a new knowledge-focused agent is required whose job it is to scan the environment to gather the required knowledge and turn that into actional sales intelligence. If the customer community’s priorities have changed, and they identify a new set of requirements, then that information should be captured realtime, and linked to all other knowledge, and passed on to the training agents. This is a fuzzy role today - distributed across the organization. Often it falls to the organizational leadership. But this will be developed as a focused activity because of the changing nature of competition.
In real sales organizations, there are multiple specialized roles. Each individual is training
Other Example TAPs
The Three Agent Persona (TAP) framework can be applied to many areas. Here are some examples:
Conclusion
Early versions of agents (AI SDRs) are missing two other crucial agent personas to complete the picture, and flow. We believe that this will become conventional wisdom in a few months, and accelerate the development of effective systems of agents. We should expect to be specialized agent companies emerging doing full TAP implementation or one of the three personas. It will make for an exciting future.