Building an LLM from the ground up is a colossal task. As Aaron puts it, you need the equivalent amount of information of all of Bloomberg’s financial history for the whole globe for the last forty years for it to be worth it.

You can fine-tune LLM’s. But they didn’t do that either. Again, because of the amount of information needed for it to be worth it.

“[you need] roughly all medical articles and all medical journals for the last hundred years to tune an LLM effectively.”

So they’re using OpenAI.

Using OpenAI has meant they could build fast. They also have access to a sophisticated, continuously evolving AI model.

They’re customising it, of course. They’re prompt engineering it so that it works on their use case really well.

The impact is twofold:

• Deel benefits from cutting-edge technology without the extensive resource investment required for development, and
• Their platform gains enhanced capabilities in handling complex, global HR and payroll queries with improved efficiency and accuracy.

Having a strong bias to build vs buy steered Deel away from using other products like Glean for its use case. But also because of what they wanted to build on top of it in the future…

The Product

I was given a quick demo of the product. It looks like a chat interface with some pre-canned queries.

When submitting a query like “How many people did we hire this month?”, instead of returning just a number, it showed a graph of hires in the last 12 months with a trending line.

Not only is it telling you the answer but it’s putting it in context with what you want to see based on adjacent questions you’d likely want to ask. How does this compare to last month? Where in the org are those hires? Where in the world are they?

Other tools Deel has built internally:

• Workbenches for their CS folks so that when you get a ticket it gets routed to the person with the most expertise in that particular area based on recently closed tickets.
• Anomaly detection for people doing wonky things with payments so they can send to their fintech crimes to investigate.
• Expense management so that when you send a receipt, it reads it and inputs it correctly.
• Filling in applications across different governments.

The impact on the team

Building this platform has improved how Deel employees handle information retrieval. This kind of work may seem fairly trivial or standard in the age of AI, but this has massively improved their experience at work.

“Think about it from the perspective of an employee, especially when you're talking about something like parental leave - a super high anxiety conversation. As a manager, I want to be present. I want to be engaging with you. I want to be supportive and excited for you in all the ways that you are. Using this tool, instead of me having to go do a bunch of Slack messages on the side, I can literally pull up the thing and 30 seconds later, we have our answer and now we're on to ‘how are we talking about coverage? How are you thinking about it? How can I help? What can we do to support you etc? It's just life-changing.”

And from a business perspective, the whole Deel team has won back time - a lot of it.

Rolling it out to customers

Internally, the platform had been a success.

Now to release it to customers.

They started by only giving access to some of their closest customers - the ones that, if something in Beta went a little awry, they’d be able to deal with it the right way. Gradually, they’ve been able to roll it out to more as the platform improves and becomes more robust.

What Deel got right to make this worthwhile for customers:

Accuracy. Getting things wrong is not just bad - it’s terrible from both a compliance perspective and from an employee’s perspective. It’s a relationship killer.

They built their data structure in such a way that makes it very easy for the LLMs to do their job and are prompt engineering to ensure each response has a source to refer back to.

Deel structures its data into analysis models designed for high throughput read and write.

Privacy. Privacy is a top concern for Deel. They’ve implemented robust measures to ensure data security and confidentiality. Like single-tenant environments, for example. In Aaron’s words, with multi-tenant architecture, “you’re only one customer ID away from getting someone else's data”.

They focused on making the data segmented so no hallucination could allow for other data to be brought in.

Access. For small organisations, Deel AI enables them to operate the same way big companies do. Performance marketing teams the world over have analysts; finance teams have FP&A professionals. What’s the equivalent for HR teams? There aren’t any unless you’re at companies like Google, Apple, Unilever etc.

Speed. For big organisations, instead of asking someone something who then goes and asks someone else and someone else, Deel AI cuts down the time to the answer significantly.

“it's like essentially giving everyone their own data analyst.”

As an exec at a company like Deel, if you need some data about the organisation you’ve got a team that you can ask to get it. They’ll probably go talk to someone else etc. At a huge company, it may take weeks and you get a fancy PowerPoint, several meetings and a bunch of people probably did a bunch of work that you didn’t need them to do.

If you’re in a smaller organisation there’s not likely a person in charge of that. So how can you query the entire company’s data? If you wanted to know:

• How many employees did I hire last week?
• What’s my average tenure of salespeople?
• How many employees do I have in this country? What are their names?

Being able to expose that is pretty powerful.

If someone in your organisation doesn’t have the time, resources, or people to help or lacks the technical skillset, with Deel AI the answers are seconds away.

The customer experience

Head of People at GitPod uses Deel’s AI product internally

To see how Deel AI is impacting customers, I spoke with Eva Hyder, Head of People at Gitpod.

An example she gave was about trying to find information about UK pensions - specifically her needing to understand a deduction inside someone’s pay slip.

Eva asked a support team member and Deel AI at the same time. As she was waiting for the response from support, AI gave an answer that was more in-depth than what support came back with.

Then after asking a follow-up question (to both support and the tool), the AI immediately gave a concise response, while support had to escalate it on their end which took a day and a half to get a response.

“I wear a lot of hats at our company, supporting many of our operations. My day has to be so rigid, so I usually like to work on one thing at a time. It’s very helpful for me when I can do a task all at once rather than context switching. Deel AI allows me to do that because I can find what I need in one go, rather than going back and forth.

We're in 14 countries, we're globally distributed. I'm on the West Coast in California, US. Our center of gravity is in Europe. So being able to get super quick answers and being able to compare it with support, who I actually really do trust a lot, made me feel like I could trust the response.”

The future of AI at Deel

Proactive insight

When thinking through the current state of how AI helps Deel, it’s answering questions—it’s giving people information.

Looking ahead, Deel envisions a future where their AI not only responds to queries but also proactively provides insights around that information.

This was the big thing they wanted to build on top of their system (and why they didn’t use off-the-shelf tools): a smart ‘analyst’ that infers what you’re asking and what you likely want to know next.

An example:

Currently, if you ask the AI tool “What is the tenure of our sales reps?" it can give you the answer and perhaps how it’s fluctuating over time, which is helpful.

In the future, Deel wants the tool to give more insight, e.g. “tenure is below the industry standard”, or “if you’re looking at expanding to this country—the market is just as good and you can hire 20% more people for the same budget”.

This shift towards a more proactive AI assistant aligns with Deel's vision of driving business growth, offering not just support but also guiding organisations in making informed decisions.

It’s a leap from reactive problem-solving to active engagement in business strategy.

Worker access

It’s in Deel’s imminent pipeline to open up the AI platform to their global roster of workers. So that’s the people who are hired by the clients who use Deel as a platform.

“We want to be the most worker-friendly HR platform on the planet. So it's a no-brainer for us. Knowledge is power. We want to give the workers knowledge so that everyone's just happier and better.”

Using AI internally

Deel is committed to enhancing internal processes by embedding AI into their operational fabric. Processes like:

• Streamlining support queries
• Improving decision-making processes
• Overall operational efficiency

Overall I’d say Deel is not just an advocate of integrating AI, but also an active user of the technology they develop.

When they use their AI tools, they test how good they are and keep improving them, ensuring it's practical and fit what businesses need.

What are the opportunities?

What can we learn from Deel and what are the opportunities?

1. Building AI applications:

Deel didn’t explore building an LLM or even fine-tuning one, given the data needed to make those effective. But getting an LLM to be accurate, provide sources and never hallucinate is a challenge that many companies are experiencing today.

Many companies have only scratched the surface of ‘what’s possible with AI’ and are (perhaps incorrectly) looking to fine-tune an LLM. You could create content to help educate folks on this topic; assess the options, how to implement an effective solution (with accurate data) etc. Format-wise, it could be a newsletter, course, consulting, or productised service.

2. Implementing AI at work:

Deel has a high build vs buy approach which is what made them lean into building their internal tool. Looking for what would make the team more efficient, by analysing what are slow, arduous tasks that many team members experienced, they thought an internal AI search tool would be valuable.

Think of all the small tasks that you have to do each day that may take you out of your deep work, that could feasibly be delivered by an AI solution.

When I was recently at an 800-person company, I was working on a course for new customers with my team. Months later we found out one of the marketing teams was planning content to help the types of customers we were looking at serving too. If we’d known that months earlier we could’ve collaborated and made a bigger impact on launch.

Given the resources, if Slack messages were fed into AI across all channels it could pull out similar work opportunities that could benefit from collaboration.

There are countless examples of this in larger companies happening every day.

If you’re adopting AI in your business, here are some things to consider:

• What models are good enough to use for our use case?
• How are we going to stress-test this and improve it?
• Is this something internal that could be extended externally?
• How is our data structured? Is that optimal for an LLM to understand?
• What is our relationship killer? What is essential for ensuring trust in the AI outputs? (sources, accuracy etc)
• If we buy vs build, what limitations does that have for future plans?
• Is buying off the shelf the best way to test a solution?
• What information is constantly being asked internally and externally?
• How much time does this solution save us? our customers?
• What is our quickest way to a valuable product experience?
• What are we doing now that is slow, arduous and takes people time to complete?
• Where are the gaps in sharing information and context in our organisation?