How a Groningen company processes requests 30 times faster with AI

At a Groningen-based home-sustainability company, every incoming request took about ten minutes of manual work: reading the email, looking up the address, gathering data about the home and typing it all into the customer system. With dozens of requests a day, the backlog grew and promising customers waited too long for a reply. Today an AI workflow does that same groundwork in about twenty seconds, and the team calls the most promising requests first. Below you can read how that works, what it delivered and which part of this approach is usable for your organization.

Table of contents

1. The problem: more requests than the team could handle
2. The solution: enrich before assessing
3. The result in practice
4. What is usable for your organization
5. How we can help
6. Frequently asked questions

The problem: more requests than the team could handle

The company advises homeowners on making their homes more sustainable and received a steady stream of requests through its website and by email. Good news, but every request called for the same ritual: read it, extract the address, look up what kind of home it is, judge whether it is a promising job, and copy the data into the customer system. Ten minutes per request, dozens of requests a day.

The consequence is easy to guess. Processing lagged behind the inflow, the order of handling was the order of arrival (not the order of promise), and the people who most wanted to sit at the table with customers spent part of their day retyping data.

The solution: enrich before assessing

The heart of the solution is a reversal: let AI collect all the facts first, and let people assess afterward. For each incoming request, the workflow we built does the following:

1. Extract the address. A language model reads the email and pulls the address out of the free text, even when the applicant writes it down sloppily or incompletely. The model is tuned for precision: better to hand a request back for a check than to push a wrong address through.
2. Retrieve home data. With the address, the workflow automatically retrieves the relevant data from public registers: construction year, home type, floor area and energy label.
3. Calculate a score. Every request gets a priority score on seven characteristics, including home type, size, construction year and energy label. The workflow also calculates the distance from the company to the applicant, and with it the travel time: a job around the corner weighs differently from a job on the other side of the province.
4. Prepare and match in the customer system. The requests appear ranked in the CRM and are matched to the advisor who fits them best, with a log of every step, so it is always possible to trace how a score came about.

Two design choices turned out to matter most in hindsight. The first: when in doubt, do not guess. If the language model cannot establish an address with certainty, the request goes back to the manual queue instead of entering the system with a half-correct address. Better ten seconds of extra human work than a technician sent to the wrong street. The second: every step is recorded. For each score you can read back which data went into it and where it came from. That makes the outcomes verifiable, and it makes adjusting easy when the company shifts its priorities.

The team then does what people do best: have the conversation. Judging whether a request truly fits remains human work; the groundwork has disappeared.

The result in practice

Processing time per request went from about ten minutes to about twenty seconds, thirty times faster. The workflow processes three to four requests per minute and can handle well over a hundred requests a day without anyone having to step in.

Just as important is what happened to the order. Where the oldest request used to be on top, the most promising one is now on top, with the right person. An old, draughty farmhouse with a low energy label lands with an experienced advisor; a small apartment with energy label A+ goes to a junior colleague. And because every request goes through the same steps, the handling is consistent: no more requests falling through the cracks because things got busy.

The time that was freed up did not vanish into the margin. The advisors now spend their day on the work they were hired for: looking at what is possible in people's homes, and making quotes for homes whose facts they already know before they ring the bell. The conversation also starts differently as a result; whoever calls already knows the construction year, the label and the home type, and can get straight to the question behind the request.

What is usable for your organization

The pattern behind this case applies more broadly than sustainability. It works anywhere requests or reports come in that you have to assess and follow up:

• Quote requests at construction, installation and service companies: enrich with project or object data, score on size and distance.
• Intakes and sign-ups in care and services: complete the data before the first contact.
• Damage reports or malfunctions: determine urgency on facts instead of on the tone of the report.

The conditions are always the same three. There has to be something to enrich: public registers, as here, or your own systems. There has to be a clear follow-up process, because a score without a call list changes nothing. And the final assessment stays with a person; the AI does the groundwork, not the decision.

You will find more examples of what is being built in the region in our overview AI in Groningen: what is happening and where you start as a business.

How we can help

If you want to know whether your request process lends itself to this, take the free AI scan: in a few minutes you see which stage of AI maturity your organization is in and what a logical next step is.

Feel free to get in touch for a no-obligation conversation. We first discuss whether this pattern fits your process; only after that does anything go on paper.

Frequently asked questions