Against expectations: How Freepik built better AI support on Salesforce with Fin

Freepik serves more than 100 million users worldwide as one of the world's leading AI-powered creative suites. Pablo M. Nicolini joined the company as Customer Service Manager a year ago. Working alongside Chema Ruiz, the Technical Lead responsible for Freepik's Salesforce infrastructure, he had a clear mandate: use AI to decouple support quality from headcount growth. As an AI company itself, the pressure to get this right was felt at every level.

The logical first choice for the team was Agentforce. They were already using Salesforce for their CRM and helpdesk, custom workflows were in place, and Agentforce appeared to offer a native path to AI-powered support. Over six months, three engineers dedicated sustained time to the implementation. They escalated blockers to Salesforce's engineering team in the US, attended technical meetings, and kept rebuilding. They never reached production.

Within days of testing Fin, things changed.

Following a quick and easy set up, Fin is now involved in 68% of all customer conversations at Freepik and fully resolves 40% of all support volume (around 18,000 conversations every month) without any human involvement. The team's human agents still use the same Salesforce case routing and workflows as before. Nothing was ripped out. Fin just integrated seamlessly with their existing setup.

With Fin working so smoothly, the next phase is already in motion: expanding action-based flows, widening coverage, and putting Fin's procedural capabilities to work so that any team member, not just engineers, can configure and adapt the system in real time.

Freepik is not a traditional SaaS company adding AI as a feature. The company has spent years building generative AI tools (image generators, video tools, upscalers, character systems) at a scale that now serves more than 100 million users and nearly one million paying subscribers. AI is the product.

That context shaped everything about how the team approached customer support. As the product suite expanded from stock content into AI creative tools, support became more complex alongside it. Billing questions became more nuanced. AI credit consumption needed context. Subscription tiers multiplied. And the volume of these queries continued rising year-over-year.

When Pablo joined, his mandate was to professionalize the function and scale it, not by adding headcount but by doing more with what was already there. “The goal was never to adopt AI for the sake of AI. The goal was to provide a better service with the resources we have,” says Pablo.

For a company with AI at its core, there was an expectation set by the COO and investors that the support operation reflected the product’s ambitions.

The existing model was heavily email-based. It worked for structured queries, but it introduced friction: back-and-forth threads to clarify issues, slow resolutions, and no way to gather context quickly. Live human chat at scale wasn't feasible without a much larger team. The only realistic path was deeper automation, leveraging a system that could answer questions and take action.

The team turned to Salesforce first.

Salesforce was already central to how Freepik ran support. Cases came in through email-to-case and web-to-case. Billing integrations with Chargebee and Stripe were live. Refund automations were already built. When Salesforce launched Agentforce, it looked like a natural extension of work already done: the team was already in the ecosystem, security approvals were in place, and adding an AI feature felt like it wouldn't require a new vendor cycle.

“We already had Salesforce and many things integrated there. They had a solution for the virtual agent. The assumption was that it would be consistent with the environment we were already working in,” says Pablo.

Chema's starting point was the same. As a Salesforce developer who had spent years building out the company's infrastructure, adding a native AI feature appeared straightforward. And technically, it should have been.

It wasn't.

Three engineers spent months on the Agentforce implementation. They configured Data Cloud, built retrievers and indices, ran test cycles, and tuned configuration. When blockers emerged, as they did consistently, the team escalated directly to Salesforce's engineering team.

“We had two meetings with the top team at Agentforce. I was explaining the issues and they were saying: ‘we know, but we don't have a solution.’ They only resolved one issue out of everything we raised,” says Chema.

The responses were always the same: features were in development, they'd be available “soon”. But “soon” was a moving target.

“We didn't see limitations,” says Pablo. “We just saw blockers.” Basic visibility was missing: no dashboard to monitor conversations, no performance data, no way to supervise what the Agent was doing. That was supposedly coming. So was email support. So were the fixes to the most common errors they'd surfaced.

When the team pressed for deeper technical support to get past persistent blockers, Salesforce's answer was to bring in an external consultancy. “They said if we needed further help, we should hire a service to support us,” recalls Pablo. “At that point I thought this is becoming extra effort and extra budget on top of something that isn't working.”

After more than six months, no Agent in production, and a growing realization that the timeline to something usable kept shifting, the team made a decision. They would look for another solution.

Freepik ran a structured comparison between three options: the lightweight AI chatbot they'd already been running, Agentforce, and Fin. The evaluation criteria wasn’t about features in isolation. The team wanted to know which tool they could configure and maintain without sustained engineering effort to handle their full operational reality that consisted of informational queries, backend actions, and reporting visibility.

The setup process clearly illustrated the gap. With Agentforce, connecting to the knowledge base required building retrievers, creating indices, and configuring metadata layers. With Fin, the team pointed it at their help center URL and it scraped the knowledge base directly. From there, adjustments could be made by anyone.

“In just a few days of integration, we had the same quality we’d been working toward for six months with Agentforce. It was super fast to implement and way easier to maintain,” says Pablo. “I was even able to change things myself. With Agentforce, I always needed Chema.”

For Chema, who had spent years inside Salesforce’s architecture, the contrast was sharper still.

“You assume the Salesforce-native option will be easier. But it was the opposite. We were three engineers spending months on Agentforce, and Fin is an external system, so we expected the integration to add complexity. But it felt like Fin was inside Salesforce and Agentforce was the external tool,” says Chema.

Fin also didn’t require Freepik to rebuild their operating model. Human agents stayed in Salesforce. Case routing continued to work as before. Handoffs from Fin landed in the existing Salesforce queue. The Chargebee and Stripe integrations the team had spent years building remained in place. There was no rip-and-replace. Fin connected on top of the existing helpdesk rather than competing with it.

That mattered. The team had already made one painful decision to start over. They weren’t going to build on another uncertain foundation.

Fin is now embedded across Freepik's customer-facing surfaces. It handles informational queries, subscription and billing questions, and an expanding range of action-based flows (like refund requests, payment status checks, and account management queries) by connecting to Freepik's own APIs.

Fin is involved in 68% of all incoming conversations, and resolves 59% of those. Overall, that means 40% of all support queries are now fully resolved without a human agent, closed from start to finish by Fin alone. That's around 18,000 conversations every month.

But the numbers only tell part of what changed.

More productive interactions with customers through chat

Freepik's support operation had been email-heavy because live chat at scale wasn't feasible without a larger team. That constraint is gone. With Fin handling repeatable volume, the team has capacity to offer real-time support on more pressing issues. And the quality of the exchange with the customer using chat has improved significantly.

“With a chat, you can ask for the screenshot, the error, the details. In just a few minutes, you have everything. Over email, that same process could take days,” says Pablo.

Cases arrive better-formed. Handoffs are faster. During incidents, Fin can notify users of platform issues in real time, a step that previously required manual intervention.

Data insights to unblock progress

Performance visibility was another unexpected improvement for the team. During the Agentforce phase, basic performance data had been missing entirely. With Fin, conversation-level reporting was available from day one.

“Having easy access to reporting was a key thing for me. That was something completely new that unblocked many things,” says Pablo.

Support became iterative rather than reactive. The team could see what wasn't performing, fix it, and move on.

A mindset shift

“Support teams are always afraid that these tools will take over their jobs,” says Pablo. “But then they realized Fin was there to help. What changed was where their time went.”

Routine interactions, like templated email responses, information gathering, and repetitive FAQ queries, were handed over to Fin, and human time focused on issues that required judgment, context, and experience. The team didn't shrink. Its work changed.

With stable automation in place and a team comfortable working alongside Fin, the focus has shifted to depth and control.

Action-based flows are expanding beyond billing and subscription management. Coverage is increasing across more surfaces. Guardrails for more complex financial and account scenarios are being refined. And the team is preparing to put Fin's procedural capabilities to work in a way that opens configuration up to the whole team.

“When I saw the Procedures feature, I thought now I can just describe what I need and Fin can do it. If we can get to that, it will be a game changer,” says Pablo.

Freepik's support operation runs around the clock. But the people who know it best, who understand the customer questions, the edge cases, the moments where guidance needs updating, aren't always the ones who can change the system. That gap is what the next phase is designed to close.

“If more people can safely manage and adjust Fin, that changes how quickly we can improve,” says Pablo.

That's where the trajectory points: a support operation where the team that knows customers best is also the team that can shape the system that serves them.