Reducing support overhead by enabling agents to modify backend records
Move AI agents from read-only search engines to read-write assistants that update Salesforce and backend records through API tool calls to reduce manual work.
A customer calls a business line at 2 AM to update a shipping address. The AI agent answers immediately, correctly identifies the user, and recites the policy for address changes. It then tells the customer to wait until 9 AM for a human representative to manually type the new address into the database. This interaction is a technical success: however, it is a functional failure. The business paid for the AI compute and the phone minutes only to schedule a manual task for their staff the next morning.
Most deployments fail because they treat agents as interactive search engines. The agent reads a website or a vector store of PDF files and provides answers based on that static knowledge. While this deflects simple FAQ queries, it does nothing for transactional requests. True automation requires the agent to move from a read-only state to a read-write state. This prevents static knowledge bases from creating support bottlenecks that frustrate users who need action instead of just information.
Duvi handles this transition by allowing agents to call external tools. These tools include HTTP APIs or direct integrations with Salesforce. When an agent has access to these endpoints, it can verify a customer identity and push updates to the business records. Instead of telling the customer how to change their address, the agent performs the update during the call or the WhatsApp session. This eliminates the manual workload increase usually seen with documentation only agents.
Setting this up requires more than just indexing a website. The business defines the capabilities of the agent within the system prompt. Facts stay in the knowledge base, but the rules for using tools are defined in the prompt. Each rule should include a reason so the model understands the logic. For a voice agent on a Twilio number, the prompt is tuned to handle the brevity and interruptions typical of spoken conversation. This allows businesses to replace the rigid phone tree with contextual intent and execution.
The consistency of these actions across channels is where many systems break down. A customer might start a conversation on a website chat widget and then follow up via WhatsApp. If the agent logic is siloed, the WhatsApp agent might not have the same API permissions as the web agent. Duvi uses the same agent and the same knowledge base across every surface. This unified approach helps in handling context fragmentation when customers switch from phone to WhatsApp.
Using a single agent definition prevents data fragmentation. If the agent modifies a record in Salesforce via a phone call, that change is immediately reflected in any subsequent interaction on WhatsApp. The agent knows what happened because it references the same live data sources and the same conversation history. The underlying logic and available tools remain identical whether the customer uses a phone call or a digital message.
A risk in automating actions is the potential for an agent to get stuck or misinterpret a complex request. Duvi manages this through human handover. If the agent encounters a situation it cannot resolve with its available tools, it can transfer the live chat or the call to a human staff member. Automation handles the high volume of repetitive tasks while humans handle the edge cases that require subjective judgment. This logic is critical because the cost of an internal ticket transfer remains high, and resolving issues at the first point of contact saves significant resources.
The implementation involves connecting the business's own numbers or accounts. By using providers like Meta or Twilio, the business maintains ownership of the communication channel. Integration with the website is done with a single script tag. There is no engineering work required to launch the initial agent. The complexity lives in how the business chooses to connect its APIs to the agent toolset.
When an agent can actually do the work instead of just talking about it, the cost per conversation drops. The business no longer pays for an AI agent to tell a customer to wait for a human. It pays for a completed transaction. This shifts the role of the support team from data entry to oversight. They spend their time reviewing conversation analytics and insights provided by Duvi instead of manually updating shipping addresses.