This is part 5 of our Built with AI series. Part 1 covered the personalized prospect videos that started it all, and in part 4 AI built, tested, and prepped a feature for deployment by itself. This time the project was bigger: replacing an entire commercial product over one weekend. I demoed the finished system live at our September AI workshop, and the video of that demo is coming soon.
Why build a signature system at all?
Because Microsoft 365 has no built-in way to give everyone in a company the same email signature. One of our clients uses a product called Exclaimer for this. It manages global signatures centrally, since without it you are sharing templates between people and relying on each of them to apply the template correctly, and some devices cannot apply them at all. Your signature on your phone ends up different from the one on your computer.
That is a solved problem, but it is solved by a subscription. I wanted to see whether Claude and I could solve it ourselves, so on Friday of last week I started talking to Claude about how to build it.
How does a centralized signature actually get onto an email?
The short version: your outbound mail takes a brief detour through a server we control, which stamps the signature on and sends the message along. When you send an email in Microsoft 365, it goes to the transport service inside the 365 mail server, which routes the message where it needs to go. Transport rules can act on mail as it passes through; we already use them for things like anti-spam. Our rule diverts outbound email to a virtual machine we host in Azure, which speaks SMTP, appends the correct signature, and sends the email on its way.
The sender does nothing special. The signature is applied after the message leaves the mailbox, so it is identical whether the email came from Outlook on a desktop or a phone.
What did the Friday proof of concept involve?
By Friday morning I had a working test: an email sent from a test account arrived in my business inbox with a signature stamped on. Getting there meant standing up a test Microsoft tenant on a spare domain I had, btechs.com, creating a mailbox, and assigning a license. Then Claude built out the basics: the Azure virtual machine, the SMTP handling, and the transport rule setup.
There was plenty of supporting configuration too. We added DNS records so the relayed mail would not show up as spam. We set up a certificate so Exchange would trust our server. Claude figured out each of those pieces as we went. The first signature was deliberately basic, nothing complex, and at that point the system was not usable for much else. But the concept was proven in a morning.
How did it become a real product over the weekend?
Through phases, worked in spare moments on Saturday and Sunday. Phase two built the database and the web application. Claude’s first pass at signature editing was raw HTML, which is gibberish unless you write HTML for a living, so I had it build a visual designer instead. Now I drag fields where I want them, adjust alignment, swap a logo between round, cornered, and square, make elements hyperlinks, and lay out up to six columns. I pulled in the Font Awesome icon library, plus emoji and brand icons for social media links, with control over icon color and size. Everything saves automatically.
The fields themselves are driven by Microsoft 365. Name, title, department, phone, address, all pulled from each user’s 365 account, so one signature template renders correctly for every employee. Change your cell number in 365 and your signature updates with it. As the IT provider we can manage those fields for a client, or grant users access to maintain their own.
There is more in there than I expected to build: templates for new mail, replies, and internal mail, signature rules that assign templates to groups or individuals, a live preview so you are not emailing yourself to test, and scheduled campaign banners that attach a promotion to a group’s signatures between two dates and then disappear.
What went wrong, and how did AI help fix it?
Two real problems came up in testing, and both got solved by describing them to Claude in plain English. First, test emails kept landing in junk folders. I told Claude, this is going to junk mail, and it asked for the email headers. I did not know how to get headers out of Outlook, so it walked me through that too, then told me which DNS records to add. I made the changes in my GoDaddy account and Claude verified I had entered them correctly.
Second, Outlook blocked all the images in the signature by default. I noticed other companies’ signatures displayed fine, asked why ours did not, and Claude adjusted the system to embed the images directly in the message rather than linking to them. The images have displayed cleanly ever since.
This is the pattern from the whole series: I never debugged alone. I described symptoms, Claude asked for evidence, and it told me exactly where to click to collect it.
What is left before other companies can use it?
Phase four is multi-tenant onboarding, and I want it to be nearly effortless. The system was designed from the start to serve multiple companies, though right now only the test tenant is attached. The onboarding flow I am building will list the DNS records a tenant needs, validate them once they are added, then log into the tenant with the permissions we hold as their Microsoft 365 provider and create the transport rules automatically.
Two known gotchas remain. Signatures are not yet copied into the sender’s own sent items, so a forwarded message does not carry it. And if a user already has a personal Outlook signature, we need to decide whether ours replaces it or joins it. Working through gotchas like these is most of what building with AI actually is: talking through how I want things handled, catching cases I did not think about, and setting ground rules as I go.
What does a weekend build mean for your software budget?
It means the math on buy versus build has changed. Exclaimer charges per mailbox, every month, forever. My total hands-on time to reach a working product, with a designer, groups, campaigns, and deliverability sorted, was maybe four to five hours across a weekend, and a couple of those hours were the Friday proof of concept. I am not a full-time developer. I described what I wanted, and Claude handled the code, the server setup, and the troubleshooting.
Not every SaaS subscription deserves replacing. But if your business pays monthly for a tool that does one specific job, it is now reasonable to ask what custom software would cost instead. Sometimes the answer is a weekend.