Orvia handles every customer inquiry for a B2B machinery dealer

The challenge: managing six inbound channels by hand

Before Sorrato deployed Orvia, the dealer handled customer inquiries across six separate channels: the website contact form, the Harvey Shop webshop, the Sandra chatbot on the website, Marktplaats listings, a showroom appointment form, and machine-specific inquiry forms on the product pages. Each channel produced emails that landed in different inboxes. A team member had to open, read and judge each one before any useful work could start.

That judgment was the real problem. Most incoming email was not a customer inquiry at all. Newsletters, supplier correspondence, cold outreach and internal messages arrived in the same inboxes. Someone had to read each one to decide whether it needed a response and, if so, whether it was a quote request, a showroom appointment, a service call or a spare-parts question. Only then could the right response be written.

There was no central record of open inquiries, no way to see whether a sender had contacted the dealer before, and no reminder when a quote went out without follow-up. The result was 0.5 to 1 hour per working day spent on email triage and switching between inboxes. That time created no value for customers and could not be won back.

The solution: Orvia as the first reader

Orvia reads every inbound email from all six channels before any team member sees it. She decides whether the email is a real customer inquiry, classifies its type, checks whether the sender has contacted the dealer before, and prepares a draft response. When the inquiry matches a known product category, the price list is attached automatically. All of this is done before a person opens the message.

Orvia is not a chatbot that sits on one channel and answers set questions. She works across the whole intake layer of the dealer's sales process: she reads emails from six sources, writes structured records to Supabase, shows every inquiry in one Lovable interface that looks like an email client, and stages draft responses in Outlook, where the team reviews and sends them. Orvia does the reading, judging and drafting. The team handles the relationship.

The 6-step process: how Orvia works

1. Capture. An email arrives through one of six channels: the website contact form, the Harvey Shop, the Sandra chatbot, Marktplaats, the showroom appointment form, or a machine-specific inquiry form on the dealer's website.
2. First-stage triage. Google Workspace Studio checks every inbound email against one yes/no question: is this a customer inquiry? Newsletters, supplier emails, cold outreach and internal messages drop out of the flow here.
3. Orchestration and storage. Make picks up every email that passes the filter, turns it into a normalized record and writes it to Supabase. Each record holds the source channel, timestamp, sender details and the raw email content.
4. Classification. OpenAI reads the record and classifies the inquiry type: product quote request, showroom appointment, service inquiry or spare-parts question. OpenAI also checks whether the sender has sent earlier inquiries and shows that history in the Lovable interface.
5. Draft response. OpenAI writes a draft response using the dealer's response templates. If the inquiry matches a product category with an available price list, Orvia attaches that price list to the draft automatically. If no price list applies, Orvia prepares a draft for the team to complete with a custom quote.
6. Human review and send. The draft lands in Outlook, ready for review. The Lovable interface shows the original email, Orvia's classification and the draft side by side. The team member reviews the draft, adjusts it where needed and sends it. Every lead is added to ActiveCampaign for automatic follow-up campaigns.

The AI tech stack: OpenAI, Supabase, Make and the orchestration layer

Sorrato deliberately avoided a single-vendor AI suite for this build. Each layer uses the tool best suited to that job, with Make connecting them.

Solving the two-stage AI classification problem

The decision that defined the build was to split classification into two separate stages instead of running one multi-class prompt. The first stage runs in Google Workspace Studio at inbox level and has one job: is this email a customer inquiry, yes or no? The second stage runs in OpenAI inside the Make flow and has a different job: for the inquiries that passed stage one, what type is it and how should we respond?

The two stages exist because their mistakes cost different things. At stage one, a false negative (a real inquiry filtered out by mistake) is commercially worse than a false positive, so the filter is deliberately permissive. At stage two it works the other way round: a misclassified inquiry produces a confused draft, so the prompt is precise and the evaluation set is tighter. Each stage has its own prompt, its own evaluation set and its own deployment cycle. This separation is what made the downstream work deterministic: the Make routing, the Supabase writes and the Outlook drafts.

Orvia vs. a traditional chatbot

Old flow vs. new flow, side by side

Old flow (before Orvia)

• Team members open each inbox (website, Harvey Shop, Sandra, Marktplaats, showroom form, machine inquiry forms) in separate windows.
• They read each email to decide whether it is a real customer inquiry.
• They look up by hand (or skip) whether the sender has contacted the dealer before.
• They write a response from scratch, decide whether to attach a price list, and send from Outlook.
• No follow-up reminder is created, so quotes can go unanswered.

New flow (with Orvia)

• Orvia reads every inbound email from all six channels through Google Workspace Studio.
• Orvia filters out the noise at the first stage. Newsletters, cold emails and supplier correspondence never reach the team.
• Make routes confirmed inquiries into Supabase with the source details and sender history attached.
• Orvia classifies the inquiry type through OpenAI and flags whether a price list applies.
• Orvia stages a complete draft response in Outlook, with the price list attached automatically where it applies.
• The team opens the Lovable interface, reviews the original email, Orvia's assessment and the draft, then sends from Outlook in one step. ActiveCampaign logs the follow-up reminder automatically.

The business impact: zero lost inquiries and automated quote drafts

The most concrete change Orvia brought to the dealer is that the triage step is gone. Before Orvia, a team member had to read every email from six channels to decide whether it needed action. That judgment, repeated dozens of times a day, took 0.5 to 1 hour per working day. Orvia now makes that call automatically, before anyone opens the message.

The change goes deeper than the time saved. The team now opens one interface instead of six. Every inquiry in the Lovable view arrives already read, classified and drafted. The question a team member faces is no longer "is this a real inquiry?", because that has been answered. The question is "is this draft good enough to send?". That is a much faster and more valuable use of a salesperson's attention.

Every outbound message still passes through a person before it leaves the dealer. Orvia speeds up the work. She does not replace the relationship.

• 0.5 to 1 hour per working day no longer spent on manual email triage and switching between inboxes.
• Six previously separate inbound channels brought together in one Lovable interface.
• 100% of draft responses staged in Outlook for human review, with no auto-send.
• Price list attached automatically for every inquiry that matches a known product category.
• Full inquiry history shown per contact, with no manual lookup.
• Zero missed follow-ups: ActiveCampaign creates a reminder for every quote inquiry automatically.

"The hardest part of the build was not the integration plumbing, Make and Supabase handled that cleanly. The hard part was a two-stage classification problem. First, in Google Workspace Studio, we needed a reliable yes/no filter at the inbox level: is this email a customer inquiry at all? That filter has to be conservative enough not to lose real leads, but strict enough to keep newsletters and internal mail out of the system. A false negative, a real inquiry that gets dropped, is commercially worse than a false positive. So the first-stage filter is deliberately permissive, and the second stage handles the refinement.

Once an email enters the system, OpenAI does the second stage: what kind of inquiry is it, which machine does it reference, and does the inquiry match a product category with a known price list? Each stage needed its own prompt design and its own evaluation set before the view in Lovable could be trusted. Once both stages were stable, the downstream work, draft generation, follow-up tasks in ActiveCampaign, and drafts staged in Outlook, became deterministic.

And critically, every outbound email is still reviewed by a human at the dealer. Orvia accelerates the work, she does not replace the relationship."

Vito Kloots
Lead Developer at Sorrato
Published 8 May 2026 · Updated 21 September 2026
LinkedIn

Frequently asked questions

About Sorrato

Sorrato is a staffing agency for digital interns: it places AI-powered interns at B2B clients. Where traditional AI vendors sell software licences, Sorrato delivers a managed digital intern that is built, hosted and continuously improved for a fixed monthly fee. No internal IT project, no implementation sprint, no per-seat pricing.