← Video Library · Case Studies
Why Most Teams Fail with AI (And How to Fix It)
Published May 21, 2025 · 24 views on YouTube
How does Howdy.com use AI in recruitment?
Howdy.com uses large language models to evaluate CVs and interview transcripts against job opportunities, matching candidates on both technical and soft skills. Instead of building a custom machine-learning matching algorithm from scratch, the team found that LLMs could interpret vague role requirements and behavioral interview evidence directly, without needing millions of data points.
Why didn’t traditional machine learning work for this matching problem?
According to the interview, traditional machine-learning matching (the kind used by dating-algorithm services) requires thousands or millions of data points to work reliably, and Howdy simply didn’t have that volume of data. Large language models, by contrast, could reason about matches between a handful of opportunities and CVs without needing that scale of training data.
How do LLMs interpret a vague job requirement?
When a customer asks for something like “an e-commerce engineer,” different humans on a hiring team (a recruiter, a technical veteran, a software engineer) can interpret that differently, leading to mismatched candidates. Fine-tuned AI systems, given the right data, can unpack what the term actually means in context — for example, experience with WordPress, PHP, Shopify, or payment APIs — and match candidates accordingly.
What productivity gains has Howdy seen from this approach?
The interview describes a process that used to take about 45 minutes of manual work: pulling a CV and interview transcripts and writing a synopsis for a partner to review before deciding whether to interview a candidate. With AI systems doing that work automatically and instantaneously as soon as data is available, Howdy describes the change as a drastic productivity difference — while still keeping a human in the loop for the final decision and quality control.
Full video transcript
How is Howdy actually integrating the AI into your own recruitment and management of that workforce? And maybe what are some efficiencies that you’ve seen doing this from the recruiting standpoint and your own workforce management to find the right talent that’s actually going to be using these tools?
Yeah, the number one place where we’re starting to apply these things in our day-to-day processes is to streamline talent recruitment, assessment, and matching. It’s funny — when we started Howdy, one of the big things we have to do is vet talented people for their soft skills, their technical skills, their experience, and then somehow match them to the opportunity where they’re going to succeed long term, because we want to provide great value to our partners in the United States. We have another customer, which is our teammates and our colleagues in Latin America. So both sides of that equation need to really work.
At the beginning we had so many amazing things that we wanted to do in terms of machine learning and data mining to really make that match automatic — a mathematical figure-it-out. We were thinking about the OkCupid algorithm that solved love, I remember the marketing term for it — that 80 or 90% of their couples didn’t divorce after five or ten years or something like that. They had some amazing retention metric.
But I don’t think there was ever enough data to make this work, because you’d need thousands and thousands of data points, if not millions, to make something with traditional machine-learning algorithms actually work. What we’ve found in our operations is that if you have a bunch of opportunities and a bunch of CVs, these large language models are actually really masters at evaluating and considering the match on soft skills and technical skills.
I’ll give you an example. If a customer is looking for, let’s say, an e-commerce engineer — what does an e-commerce engineer really mean? If you have a recruiter, a technical veteran, and a software engineer all in the pipeline trying to find an e-commerce engineer, that can lead to misses. But these AI systems are actually really good, especially if you fine-tune them and give them the right kind of data. They can tell you what a customer actually means by an e-commerce engineer — usually experience with WordPress, PHP, Shopify, or payment APIs, and so on. These automated tools help us unpack that, find people that match those opportunities, and crucially also pick up on the soft skills our partners are asking for.
They might be asking for somebody entrepreneurial who’s okay with chaos, and in the interview transcripts, these LLMs can actually pick up evidence for those kinds of behaviors, which makes a behavioral interview such a rich source of real data that an AI system can reason about. For us that’s just super-human throughput. In places we still do this manually — we’re really good at it, but we still make mistakes because we’re human. As we offload more of this to large language models, under the supervision of people who have experience doing this, we’re finding that things are more consistent, things are much faster, and the experience is better for our partners and for the teammates who agree to be hired on the Howdy.com platform. We’re seeing those kinds of productivity boosts.
For example, it might take 45 minutes to get the CV, the transcripts, and all that, and build a little synopsis for a partner to review and determine whether or not they want to interview this person — just have a chat with them. That chat could be a 20-minute chat, and it might take an hour of reading through all those things and writing things up. In the old world, our new systems can do this automatically, instantaneously, as soon as that data is available. So it’s a drastic difference in productivity for us. And again, it’s the kind of stuff that takes some skill, but it’s mostly grunt work — reading a bunch of stuff and collating it and putting it together. An AI system is perfect for that, so long as you keep a human in the loop for the final product and the quality.