A mid-sized HVAC company in Finland was drowning in email. Their office manager spent about 3 hours every morning just reading, sorting, and replying to messages before any real work could start.
Quote requests from contractors. Complaints about delayed installations. Scheduling changes. Invoice questions. Photos of job sites. Floor plans that needed review before quoting.
Every email required a different response, and most of them followed the same pattern day after day. The office manager knew this was a waste of time. But the volume never slowed down.
The problem, in numbers
We tracked one week of email activity before building anything:
- 80-120 emails per day hitting the main inbox
- 3+ hours daily spent reading, sorting, and replying
- 35% of emails were quote requests with attachments (floor plans, photos, specifications)
- 25% were scheduling changes or confirmations
- 20% were routine questions that had the same answer every time
- 20% actually needed a human to think and decide
That means 80% of the inbox was predictable, repeatable work. Perfect for an AI agent.
What we built
The agent connects to the company's existing email system. No migration, no new software for the team to learn. It runs in the background and processes every incoming email through four steps.
Step 1: Classification. The agent reads each email and categorizes it: quote request, complaint, scheduling, invoice, general question, or unknown. This happens in under 2 seconds.
Step 2: Attachment analysis. This is where it gets interesting. When someone sends a floor plan PDF or a photo of a job site, the AI opens the file and extracts relevant information. Square meters from floor plans. Equipment models from photos. Specifications from technical documents. This data feeds directly into the quote or response.
Step 3: Draft response. Based on the classification and any extracted data, the agent drafts a reply using the company's actual tone and templates. For quote requests, it pulls pricing from their system and creates an initial estimate. For scheduling, it checks availability and proposes times. For routine questions, it sends the answer directly.
Step 4: Human review queue. Anything the agent isn't confident about gets flagged. Complaints, unusual requests, and high-value quotes land in a "Respond" folder where the office manager reviews them. But now instead of starting from zero, there's already a draft and a summary of the situation.
The results
The office manager went from spending half the morning on email to reviewing a short queue of flagged items. The rest of the inbox is handled before she even gets to the office.
Quote requests that used to wait 4 to 8 hours for a response now get an initial reply in under 5 minutes. The company estimates this alone brought in 3 to 5 extra jobs per month that would have gone to competitors who responded faster.
How the tech works (briefly)
For those curious about what's under the hood, without going too deep:
The classification layer uses AI language models to understand the intent of each email. It's trained on the company's historical email data, so it knows the difference between a quote request and a complaint, even when the customer doesn't use clear language.
Attachment analysis uses vision AI for images and document parsing for PDFs. It can read a floor plan and extract room dimensions, identify equipment in photos, and pull line items from specification sheets.
Response drafting uses a separate AI model that's been tuned to write in the company's voice. It has access to pricing data, scheduling availability, and a knowledge base of common answers. Every draft follows the company's actual email style, not generic AI-sounding text.
The whole system plugs into their existing email through standard protocols. No one on the team had to change how they work. The agent just handles the part that was eating their time.
What surprised us
The biggest win wasn't time savings. It was the quote requests.
Before the agent, quote requests with attachments sat in the inbox for hours because the office manager needed time to open each floor plan, measure it, and prepare a response. Some slipped through entirely during busy weeks.
Now every quote request gets acknowledged and processed within minutes. The company went from losing an estimated 15 quote opportunities per month to zero.
At their average job value, that's significant revenue recovered from what was essentially an inbox problem.
Would this work for your business?
If your team spends more than an hour a day on email, and most of those emails fall into predictable categories, the answer is probably yes.
The setup takes about two weeks. We study your email patterns, build the classification rules, train the response model on your company's data, and test thoroughly before going live. After that, the agent improves over time as it learns from corrections.
You don't need to change your email system. You don't need to train your team on new software. The agent just works in the background, doing the part of email that no one should be doing manually.