What I Learned from Running an AI Startup for Three Years

The moment I knew it was over, I was sitting in a WeWork in Dublin staring at a spreadsheet that said we had the bones of four months of runway left. It was a Tuesday. I remember because Tuesdays were the day I used to pretend everything was fine on the team standup.

AutoPlan was an AI-powered planning platform for the construction industry. We built scheduling optimisation tools that could cut project timelines by 15-20%. The technology worked. The product worked. The company didn’t, at least not in the way I needed it to.

Here’s what I learned.

What worked

The problem was real. Construction projects overrun by an average of 20 months and 80% over budget. That’s not a made-up stat from a pitch deck; it’s from the McKinsey Global Institute report that basically became my bible for three years. When you’re solving a problem that costs the industry $1.6 trillion annually, you don’t have to convince anyone the pain exists.

The law diploma was invaluable. This sounds like a tangent, but stick with me. I did a diploma in technology law alongside the PhD. When we started negotiating enterprise contracts with construction firms, I could actually read the terms. I caught liability clauses that would have sunk us. Founders almost never talk about legal literacy as a competitive advantage, but it saved us real money, probably €30,000 in legal fees we didn’t have to spend.

The ESA partnership opened doors. We got into the European Space Agency Business Incubation Centre. The credibility of having ESA’s logo on your slide deck when you’re pitching to enterprise clients is hard to overstate. It also connected us to satellite data sources that became genuinely useful for site monitoring. Sometimes the best thing an accelerator gives you isn’t the money, it’s the letterhead.

What didn’t work

Market timing. We were building AI for construction in 2018-2020. The industry wasn’t ready. I’d walk into meetings with project managers who were still running everything on printouts and WhatsApp groups. The gap between “this technology could save you millions” and “we’d need to change how we work” was a canyon. I underestimated how wide that canyon was by about three years.

Fundraising in Europe. I’ll be blunt: raising a pre-seed round for a B2B AI startup in Dublin in 2019 was brutal. The cheque sizes were small, the process was slow, and the appetite for deep tech was limited compared to what I was hearing from founders in San Francisco. We raised €150,000 through Entrepreneur First and angels. In the Valley, that’s a rounding error. Here, it was supposed to get us to product-market fit.

Solo founder challenges. I had a team, brilliant people, but I was the sole founder. That meant every hard decision (who to let go when money got tight, whether to pivot, when to shut down) landed on me alone. There’s a specific loneliness to being the person who has to look someone in the eye and tell them you can’t make payroll next month. I don’t recommend it.

(Quick tangent: the Entrepreneur First programme is designed to help you find a co-founder. I went through it twice, EF6 and then EF’s company-building programme. Both times I ended up going solo. Make of that what you will about my personality.)

What I’d do differently

Start with the workflow, not the algorithm. We built the optimisation engine first because that’s what excited us. We should have built a simple scheduling tool that construction managers would actually use daily, then added the AI layer on top. You can’t optimise a process that isn’t digital yet.

Pick one segment and go narrow. We tried to serve general contractors, subcontractors, and project owners simultaneously. Each has completely different workflows, incentive structures, and buying processes. If I did it again, I’d pick residential subcontractors under 50 employees and build only for them until we owned that niche.

Move to London sooner. The construction tech ecosystem in London was two years ahead of Dublin when we started. By the time I figured that out, we’d already burned through most of our credibility-building time.

So what now?

I shut AutoPlan down in late 2021. It wasn’t a dramatic failure. We returned what was left to investors, wound down properly, and I made sure everyone on the team landed somewhere good.

The thing about running a startup for three years is that you learn more about product development, team dynamics, and your own limits than you would in a decade of working for someone else. I just wish the tuition hadn’t been quite so expensive.

I’m writing this from my first month at ZF Group, a company with 150,000 employees. The contrast is… something. But that’s a story for another post.