Case study · Industrial supplier
An AI agent that turns quote-request emails into Salesforce quotes
Parquet is building an AI agent for an industrial supplier of precision measuring instruments. The agent reads the shared info@ mailbox, creates the Opportunity and Quote with line items in Salesforce for each quote request, and notifies the account owner. Anything uncertain goes to a person. A sales rep reviews each quote and sends it with one click as a reply in the original email thread. The agent runs on Anthropic's Claude Platform and starts in observation mode before it creates anything.
At a glance
- Client
- Industrial supplier
- Industry
- Industrial supply (precision measuring instruments)
- Status
- In progress
- Platforms
- Claude, Claudeforce, Agentforce, Salesforce, Sales Cloud
- Services
- AI Implementation, Sales Operations (SalesOps), Tech Stack Integration
The challenge
- Quote requests arrive by email, often with long lists of part numbers, and each one was keyed into Salesforce by hand before a quote could go out.
- Orders (purchase orders and "please process the attached" emails) follow a separate process tied to QuickBooks and must not be disturbed.
- Product codes in emails don't always match the catalog exactly, and some customers aren't in Salesforce yet.
What we built
Email triage
The agent reads info@ and classifies each message as a quote request, an order, a follow-up on an existing deal, or something to skip, using a supplier and newsletter domain list the client controls.
Opportunity and Quote creation
For each quote request it finds or proposes the account and contact, matches part numbers to products, and creates the Opportunity, Quote, and quote lines in Salesforce.
Human review by design
Unmatched products, new customers, personal email addresses, and very long requests are set aside for a person. Reps review every quote before it is sent.
One-click reply
The rep sends the approved quote as a reply in the same email thread, using the client's own quote email as the model.
Rules in plain text
The client edits the processing rules themselves, as plain text in Salesforce, without a developer.
Safe rollout
Observation mode first: the agent records what it would have created so results can be compared with quotes built by hand, and the real cost per email is measured before launch.
Where it stands
- Phase 1 Salesforce cleanup shipped alongside the agent: Vertex tax recalculation on quote line edits, Quote Builder fixes, duplicate account and contact cleanup, record transfer from former reps, and a refreshed quote template.
- The agent starts on the info@ mailbox only. Reps' own mailboxes follow once it has proven itself.
- We will publish measured results (time per quote, accuracy, cost per email) after the observation period.
Technology
Claude Platform (Anthropic API), Salesforce Sales Cloud and Quotes, Microsoft 365 / Exchange via Microsoft Graph with an application access policy limited to the info@ mailbox, QuickBooks integration left untouched.
Platforms
Services
Questions we hear
Can an AI agent create Salesforce quotes from emails?
Yes. In this project a Claude-based agent reads quote-request emails, matches the requested parts to Salesforce products, and creates the Opportunity, Quote, and quote lines. A sales rep reviews each quote before it is sent.
How do you keep an AI email agent from making mistakes with customers?
Three ways: it starts in observation mode and creates nothing until results are checked, anything uncertain is routed to a person, and no quote reaches a customer without a rep approving it.
Does the agent have access to every mailbox?
No. A Microsoft Exchange application access policy restricts it to the single shared info@ mailbox, with read and send permission only.
More case studies
- Specialty retail (pianos)
Piano retailer
Running a multi-showroom piano business on one Salesforce org
- Commercial finance (factoring)
Commercial lender
From website form to signed credit application, without re-keying
- Logistics and transportation
Logistics company
Moving a logistics sales team to a new Salesforce instance
Have a process like this?
A 30-minute call is usually enough to tell whether the fix is process, data, configuration, or AI.