Logistics Provider Automates 60% of Customer Intake

Context

A logistics provider handling approximately 800 shipment requests per week was bottlenecked at customer intake. Their process worked like this: a customer would call, email, or submit a web form with shipment details. An intake coordinator would read the request, extract the relevant information (origin, destination, cargo type, weight, dimensions, timeline, special handling requirements), enter it into Salesforce, classify the shipment type, and route it to the appropriate operations team for quoting.

The company employed six intake coordinators working in shifts to cover business hours across time zones. During peak periods—typically Monday mornings and end-of-quarter rushes—the backlog would grow to 50+ requests, and intake-to-quote time would stretch past eight hours. Customers waiting for quotes would call to follow up, creating additional work that further slowed processing.

The intake coordinators were skilled at interpreting ambiguous requests and asking the right clarifying questions, but roughly 60% of incoming requests were straightforward—standard shipment types with complete information that didn’t require judgment calls. These routine requests consumed coordinator time that could be better spent on complex, high-value intake situations.

Constraints

• Multi-channel intake. Requests arrive via email (45%), phone (30%), and web forms (25%). Any solution needed to handle all three channels, not just the easiest one.
• Variable request quality. Emails range from well-structured RFQs with spreadsheet attachments to one-line messages like “need to move 3 pallets from Dallas to Miami, what’s the rate?” The system needed to handle this full spectrum.
• Salesforce as system of record. All intake data ultimately needs to be in Salesforce with correct field mapping, opportunity staging, and team routing. The Salesforce instance has extensive customization that the solution needed to respect.
• No disruption to phone intake. The company’s phone-based customers value the personal touch. The automation couldn’t replace the phone experience but could help coordinators handle calls faster.

Approach

We structured the project in two phases over six weeks.

Phase 1: Email and web intake automation (weeks 1-4). We started with the two channels where automation could work without changing the customer experience. The system reads incoming emails and web form submissions, extracts shipment details, classifies the request by complexity, and either processes it automatically (for routine requests) or queues it with pre-extracted data for a coordinator (for complex requests).

Phase 2: Phone intake assistance (weeks 5-6). For phone calls, we built a real-time assistant for coordinators. When a call comes in via Twilio, the system transcribes the conversation, extracts shipment details in real time, and pre-populates a Salesforce draft that the coordinator reviews during or immediately after the call.

Architecture

Email processing pipeline. Incoming emails hit a monitored inbox. An AWS Lambda function triggers on new messages, passes the email content to Claude API for extraction and classification, and routes the result based on complexity scoring:

• Routine requests (complete information, standard shipment type, no special handling): Extracted data is validated against business rules, a Salesforce opportunity is created automatically, and the request is routed to the quoting team. The customer receives an acknowledgment email with their reference number and expected quote timeline.
• Complex requests (incomplete information, unusual cargo, special handling, multi-leg shipments): Extracted data is pre-populated in a coordinator review queue. The coordinator reviews, completes missing information, and approves the Salesforce entry. Processing time is still reduced because extraction and initial classification are done.
• Ambiguous requests (insufficient information to classify or extract): Routed to a coordinator with the AI’s best interpretation and suggested clarifying questions.

Web form processing. Simpler than email because forms provide structured data, but the classification and routing logic is shared. Web forms skip the extraction step and go straight to validation and classification.

Phone assistant. Twilio handles call routing and recording. Real-time transcription feeds into a lightweight extraction service that updates a live Salesforce draft as the coordinator talks with the customer. After the call, the coordinator reviews the pre-populated record—typically needing only minor corrections—and submits it.

Salesforce integration. We built a dedicated integration layer using Salesforce’s REST API that handles opportunity creation, field mapping, team routing, and duplicate detection. The integration respects the company’s existing Salesforce workflows, validation rules, and assignment logic.

Implementation Details

Extraction accuracy calibration. We trained the extraction prompts against 2,000 historical intake records. The initial accuracy was 78% for full-record extraction (all fields correct). After two rounds of prompt refinement and the addition of logistics-specific context (common trade lanes, cargo type taxonomies, industry abbreviations), accuracy reached 94% for routine requests.

Confidence-based routing. The complexity classifier uses a scoring model that considers: completeness of extracted information, whether the shipment type matches known patterns, presence of special handling keywords, and cargo value indicators. Requests scoring above the routine threshold are processed automatically. Everything else gets human review with pre-extracted data.

Duplicate and follow-up detection. A common problem in the old process was duplicate entries from customers who emailed and then called to follow up. The system checks for likely duplicates based on sender, shipment details, and timing before creating new records.

Results

After 60 days in production:

• 60% of intake requests are now processed without human intervention. These are the routine requests with complete information and standard shipment types. Coordinators don’t touch them unless the automated quality check flags an issue.
• Intake-to-quote time dropped from 4.2 hours to 47 minutes. For automated requests, it’s under 15 minutes (the time for the quoting team to respond, not the intake step). For coordinator-assisted requests, pre-extraction cuts processing time roughly in half.
• Data entry errors in Salesforce dropped from 12% to 2.3%. Most errors in the old process were transcription mistakes and missed fields. Structured extraction and validation eliminated the majority of these.
• Coordinator workload shifted. Coordinators now spend most of their time on complex, high-value intake—multi-leg international shipments, hazmat cargo, unusual dimensions—where their expertise matters most. Job satisfaction scores improved in the quarter following deployment.

Lessons

Channel matters more than you’d think. The same information comes in very different forms across email, phone, and web. Building one extraction model for all channels performed worse than building channel-specific approaches. Email extraction, in particular, needed to handle the enormous variety of how people structure requests in free text.

Salesforce integration is the hard part. The AI extraction and classification were relatively straightforward. The Salesforce integration—respecting custom fields, validation rules, workflow triggers, and assignment logic—took more development time than the AI components. Organizations with heavily customized CRMs should plan for this.

Start with the auto-processable requests. The temptation is to tackle the hardest cases first. We got the most value fastest by automating the easy 60% first, which immediately freed coordinator time and reduced backlog. The complex cases still benefit from pre-extraction, but the ROI case was proven on the routine volume.

Next Steps

The company is expanding the system to handle rate confirmations and booking updates—post-quote interactions that currently require coordinator involvement but are often routine. They’re also exploring extending the phone assistant to suggest upsell opportunities based on shipment patterns.

Security and Data Handling

Customer communications and shipment data are processed through the company’s AWS environment. The Claude API integration uses the company’s enterprise agreement with data processing terms that prohibit training on customer data. Personally identifiable information in emails is flagged and handled according to the company’s data classification policy. All Salesforce data access respects existing role-based permissions.

This case study represents a real engagement with details modified to protect client confidentiality. Specific metrics are representative of actual results.