Most business websites have a contact form that nobody trusts. It fills up with sales pitches and bots, the one real enquiry gets buried, and the person who should reply is busy running the business.
Ours was no different. When we rebuilt our site in the summer we treated our own front door as a small AI project, and this is the write-up: what we built, how it works end to end, and the choices that would apply to any queue of things arriving in your business that a person has to read before real work can start.
This is one of many mundane day-to-day processes we have improved through this kind of integrated AI use.
The job
Before this, a real inbound enquiry cost me ten to fifteen minutes before I had done anything useful with it.
I had to: check the person is real. Look up the company. See whether we already know them. Draft a reply. Send a link to find a time. Chase when they do not book.
Multiply that by every message that turns out to be an SEO agency, and the form becomes something you check less and less often, which is how the real one gets missed.
So the job had three parts: never miss the real one, never spend a person’s time on the rest, and never let a bot learn anything from how we respond.
The AI in this system judges each message against a rubric. To explain that jargon, a rubric is what a teacher marks an exam against: a written list of what good looks like, what counts for more and what counts for less, and what disqualifies an answer outright.
Ours is a page of plain English that we wrote. It says, for example, that someone offering us their services is a seller and scores low no matter how polished the message, and that a vague message from a plausible company is “unclear” rather than junk.
The model does the reading. The rubric is where our judgement lives, and we can change it in an afternoon.
What happens when someone presses send
1. The form runs some checks before anything is stored
A handful of unglamorous defences come first: a hidden field only bots fill in, a check that the form was open long enough for a person to have typed into it, limits on how many submissions can come from one place, and a check that the email domain can receive mail at all.
A message that fails these is filed as junk, and the sender sees the normal “thanks, we got it” screen. It learns nothing. Only two checks show an error, and both exist because a real person may have mistyped their address.
2. An AI reads the message and gathers evidence
For anything that survives, a model (we use Claude) is handed the message plus two things it fetches itself. If the company field looks like a website, it reads that site’s title and description. And it looks the sender up in our CRM, Pipedrive, by email and by company.
When an existing client came in through the front door recently, the system flagged the relationship before I had read a word, and the message was handled as a client conversation rather than a new lead.
3. It gives a verdict, with reasons
The model scores the message from 0 to 100 against the rubric: is this a buyer or a seller, is the need specific, does the company fit the kind of organisation we work with, do the email and company make sense together. It puts the message in one of four boxes: prospect, unclear, vendor pitch, or spam.
It has to return its answer in a fixed structure, so there is no free-form guesswork to interpret, and it has to give its reasons in a sentence.
4. A human decides
The AI never emails a visitor and never approves anything. A prospect or an unclear message lands in a private Slack channel with the verdict, the sender’s details, their need in their own words, any CRM match, and three buttons: approve, decline, view. Vendor pitches and spam are held in a separate one-line message with a view link, in case we want to overrule.
Each morning a short digest lists what was held overnight.
5. Approve, and the visitor gets a link to pick a time
When I press approve, I see the exact email that will go out, can add one personal sentence, and confirm. The visitor receives a single-use Calendly link for a free 30-minute call. When they choose a slot, the call goes straight into my calendar, Calendly tells our system, the CRM record is updated with the call, and Slack gets a “booked” line.
If they cancel, the record steps back and a fresh approval is one tap away.
6. The CRM is kept honest throughout
Only genuine new prospects create records. If the person or company already exists, nothing is duplicated. The lead carries the message verbatim, where it came from, and the AI’s verdict and score, so anyone on the team can see why it was treated that way.
What changed for me
The ten to fifteen minutes per enquiry is now one click, on my phone or my laptop, with the research already done and the reply already written. The rubbish never reaches me. What does reach me arrives with a recommendation and the evidence behind it, and I can accept or decline in the time it takes to read a text message. The meeting puts itself in my diary.
The measure for a project like this is minutes handed back per event, multiplied by how often the event happens.
On its own, the number looks small. The return is in the multiplier. An event like this happens a handful of times a day in our business and, in many businesses, dozens. At ten a day and ten minutes each, that is more than 400 hours a year, or ten working weeks of a senior person’s time.
The design decisions that transfer to any business
These are the choices that made this one work. None of them is specific to our industry.
The AI recommends. A person decides. A human confirms anything that reaches a customer, and sees exactly what will be sent before it goes. We chose this rule deliberately, and it is why we were comfortable switching it on.
Give the model your own knowledge before it judges. Reading the CRM and the company website before scoring is what turns a generic classifier into something that knows your business. We wrote the rubric in plain language, and we can change it in an afternoon.
Design for the day the AI is unavailable. AI services have off days like any other supplier: timeouts, outages, changes on their end. We decided early that a message must never be lost because the model was unreachable, and that nothing must ever be sent to a visitor on a judgement nobody made.
So if the model cannot be reached, the system tries again a little later, and if it still cannot, the message goes to Slack anyway, marked “review by hand”, with no CRM record created and no reply sent. We see it, we deal with it, and the visitor experiences a normal reply from a person.
A slow pipeline is fine. A silent one loses a customer.
Keep a record of every step. Every message carries a history of what happened to it and who did it. When something looks odd, we can see why in a minute instead of guessing.
Do not make the visitor wait for the AI. The form responds instantly, and the thinking happens after.
What it did not need
No new database service, no data science team, no six-figure platform. The whole thing runs on the same hosting as the website, one AI service, and the tools we already used every day: Slack, Pipedrive, Calendly. None of those is special. If your business runs on Teams, HubSpot and Microsoft Bookings, or on email, a spreadsheet and a shared calendar, the same shape works.
The AI sits between the tools you have and does the reading, checking and drafting in the gaps where a person used to.
Design for what the system gives back to the person.
Sometimes it is time. Sometimes it is better information at the moment of decision, or a decision made more consistently because the same rules are applied to every case. Here it was all three, and the model itself was the least interesting part. The rules we gave it and the decision about where the person sits did the work.
If you have a version of this problem
Every business has one of these jobs, and most have several: a queue of things arriving that a person has to read, sort and route before any real work starts. Inbound enquiries, supplier invoices, support tickets, job applications, tender alerts, insurance claims, referrals. Each one looks too small to bother with on its own. Over a year, they take up a senior person’s time.
If you would like to talk through where one might sit in your business, I offer a free 30-minute call with no obligation attached. Fill in the form on our contact page. You already know what happens next.
