In two previous articles (Sales Automation Playbook  and From Lead to Close), we explored how Sales teams get more done by using AI Large Language Models (“LLM”), Machine Learning (“ML”), and Robotic Process Automation (“RPA”) to automate tasks. They free up their time for high value activities – spending time with clients, closing deals.

What has got the Asteri team excited is the AI Sales Development Representative (“SDR”). AI SDR are hyper-specialized AI agents which do the heavy lifting of outbound motions with cold prospects. They search for prospects; find their contact information; reach out; follow up; secure meetings, and fill the sales team’s pipeline with new opportunities. 

It’s like – Where do I sign up? 

So, here is how you build an AI SDR with agentic AI workflows.

What Agentic AI Is (and Isn’t)

Let’s start with definitions for automation and agentic AI

Automation follows set rules: if X happens, do Y.  Automation is powerful for repetitive, predictable tasks that can be set up as workflows of tasks chained together: update the CRM, send an email, create a reminder. Automation is efficient, but rigid as it only works for scenarios that can be coded in advance.

Non-agentic workflows go a step further. They let an AI complete parts of that chain of tasks. For example, an LLM drafts a follow-up email, or an ML model scores a lead. Useful, but still locked into a pre-set sequence.

Agentic workflows plan, act, and adapt. Here, the AI decides which task matters next, based on context. They look at context, decide the next step, take action, and then adapt when the desired outcome is or is not achieved. In Sales, that might mean monitoring a lead’s engagement, choosing the right moment to re-engage, or routing a lead to a rep. This loop of Plan-Do-Check-Act makes the workflow dynamic instead of fixed.

What Agentic Workflows Do

In practice, agentic workflows are generally made up of multiple agents working together, each with its own role. Specifically, they can:

  • Plan: Set goals, break them into steps, and choose which actions to prioritize. This means agents adapt their strategy as priorities shift.
  • Do: Take concrete actions such as pulling data, updating systems, or triggering workflows. This allows them to push work forward.
  • Check: Pause and evaluate outcomes. If something doesn’t work to plan, the agent can adjust its approach instead of repeating the same mistake. 
  • Act: Pass tasks between agents, just like people in a sales pod. One might surface prospects, another enrich them, and a third draft outreach, all coordinated by a meta-agent.

These four capabilities are the foundation of autonomy. Together, they allow agentic workflows to adapt, respond to uncertainty, and coordinate at scale. For sales teams, this translates into workflows that run with fewer handoffs and less manual oversight.

How to Build an Agentic Workflow

An Agentic Workflow is built with three core components. These three components working together enable autonomy. 

  • Brain: The brain is usually a large language model. It does the reasoning, breaks down goals into steps, and generates the right language to communicate — whether to draft an email, summarize a meeting, or plan the next action.
  • Memory: Memory allows the agent to retain what happened before and apply it in the future. For example, remember what was promised to a customer, or know the last time a lead was contacted. Memory can also extend beyond simple recall: by accessing internal documentation such as product specs, pricing sheets, or playbooks, the agent can ground its decisions in the company’s actual knowledge. Memory is key. Without reliable memory, the agent starts from a blank sheet each time. Without reliable memory, an agent repeats mistakes or fabricates details. The memory challenge is typically solved using RAG, a topic we will discuss in a later article.
  • Tools: Tools are how the agent interacts with the outside world. Tools allow the agent to, for example, extract data from the CRM, query a database, send an email, or call an API.

Agentic Workflow in Sales

Consider an agentic workflow to find prospects, using a meta-agent that manages three specialized agents:

  • Agent 1: Company Finder. Agent 1, the Company Finder Agent, identifies companies that match your ICP. It plans the search, then scours online directories, databases and news sources to find new companies. It then proceeds to collate a list of company names and their website URL
  • Agent 2: Decision Maker Finder: Agent 2, the Decision Maker Finder Agent, gets handed a list of company names and URL by Agent 1. It plans the search, then scours online directories, LinkedIn, databases and news sources to find the decision makers in said companies. It then proceeds to create a prospect in the CRM, complete with Company details (name, URL) and Contact details (first name, last name, job title, LinkedIn URL)
  • Agent 3: Data Enricher. The Data Enricher Agent gets handed a list of decision makers from Agent 2. It checks various databases such as Apollo to find the decision makers’ email and mobile number. It checks other databases to identify contextual buying signals – new funding announcements, recent news, leadership changes, etc. It saves this structured information in the CRM record, ready for review.
  • Meta-agent. Oversees the workflow, decides when enrichment is needed, and ensures handoffs happen smoothly.

These workflows happen independently and with minimal oversight of your human SDR. Inevitably, though, your human SDR will need to QA the work of the agentic workflow. 

There are many use cases for agentic AI, such as::

  • Cold outreach agent: Sends an introductory email to the decision maker to introduce your company and its credentials
  • Nurture agent: Keeps leads warm. Detects inactivity, triggers the right nudge, and escalates when engagement signals reappear.
  • Engagement agent: Personalizes follow-ups at scale. Pauses when a reply comes in, adapts messaging to context, and routes the lead back to a human rep at the right time.

Pitfalls and Considerations

Agentic AI is simple but it is not easy. Many technical challenges need to be solved.

In our view, the main one is Memory, which can be unreliable, leading to hallucinations or contradictions if not grounded in the right data. Tools can misfire if not configured with proper guardrails, creating downstream errors. If built inadequately, agents may pursue the wrong goals or generate outputs that don’t align with brand or compliance requirements.

For sales teams, this means early pilots should include human review and clearly defined guardrails.

It also means being disciplined about data governance:  if the agent doesn’t have access to clean, up-to-date internal knowledge, its decisions will be flawed.

These challenges set the stage for our next article: how Retrieval-Augmented Generation (RAG) can provide the reliable memory that agents need to operate safely and effectively.

By Nicolas Mialaret and Maurits Van Roozendaal

cccc

Figure out AI for your business Book a call with Asteri!

Mores articles on Revenue

Follow us on Linkedin

Receive exclusive Benchmarks and Research directly in your inbox Subscribe to our ASTERI+

Monthly Summary of ASTERI Insights and more Subscribe on Linkedin