73 million users. 37 support reps. A 60% autonomous resolution rate. What we learned working with Vivino on AI in customer service
At Dreamforce, Vivino shared a number that’s hard to ignore: 73 million users, a support team of just 37 people, and an AI achieving a 60% autonomous resolution rate within 30 days.
This example is special for us for a few reasons. We’d been working with Vivino at the very early stages of this project, helping them figure out the strategy and build the first prototype. So behind that statistic, we see something you don’t usually get from a conference stage: how much work goes in before AI actually starts doing its job. And here’s what we learned from it, put simply: the real value of AI in customer service isn’t in answering customers, it’s in actually solving their problems.
From chatbot to part of the process
Normal customer service works like this. A customer writes in, an agent picks up the request, looks for the information they need, does a few things across a few systems, and gets back to the customer. That’s fine when you don’t have many requests coming in. The trouble starts once the business grows and the number of people writing in grows with it.
Up to a point, the only real option was to hire more agents as you got more customers. You could tighten up the process, make agents more efficient, add bits of automation here and there, but the basic setup stayed the same: a person handles every single ticket.
AI agents break that pattern. Instead of just writing a reply, an agent can work out what the customer actually wants, access the relevant information from the CRM and knowledge sources, check it, take an action, and close out the job on its own. If something’s too tricky, it hands it to a person. That sounds small when you put it that way, but in practice it’s a big shift. AI stops being an add-on and becomes part of how support actually runs, which is why you can’t really treat a project like this as “just adding some software.”
The hard part happens way before anything goes live
When a company says “let’s add an AI agent,” it’s tempting to jump straight to the tech: look at what the platform can do, pick a few use cases, build a demo, show it off. But the real challenge usually has nothing to do with the technology itself. It’s about which tasks actually make sense to hand to AI, what it needs to see to do its job, what it can handle on its own, and where a person absolutely needs to stay involved.
That’s what we spent most of our early time on with Vivino. Our job wasn’t just to prove Agentforce could talk to customers. We had to check whether the idea actually held up inside a real support process. What does the agent need to know to make good calls? What systems does it need access to? What happens when something unusual comes up? Where does automation genuinely help, and where does it just add complexity?
That’s a real shift in how you approach the problem. We didn’t build the AI first and go looking for a use for it afterward. We started with the actual problem and worked toward the technology, not the other way around.
Get Angelina Losik’s stories in your inbox
Join Medium for free to get updates from this writer.
What we were really testing with the prototype
At this stage, we built a prototype together with Vivino, and it’s worth saying plainly why a prototype even matters here. Not to make a nice demo. Not even to build a small version of the final product. A prototype exists to test whether an idea actually holds up. We wanted to know whether an AI agent could really work inside this process, and whether the process itself was even a good fit for AI to handle on its own.
That’s a very different mindset. Build around the technology and you end up with a bunch of cool features pretty fast. Build around the actual process, and you have to cut everything that isn’t essential and ask the much less comfortable question: does this actually work? That’s when you start to see where the data is missing, where the agent doesn’t have enough context, what can be automated, and what should stay with a person for now. For us, that was one of the most useful parts of working with Vivino.
AI without context doesn’t fix anything
There’s another thing that becomes obvious in a project like this. An AI agent is only as useful as the context, data, and actions behind it. If it doesn’t know who it’s talking to, can’t see what happened before, doesn’t have the right information, and doesn’t know the business rules, it runs out of steam fast. You can have a great model, but if you don’t give it enough context, all you get is something that sounds convincing without actually helping. That’s not good enough for customer service, because the goal isn’t a nice-sounding reply, it’s the customer actually getting their problem fixed.
So when we build these systems, we look at the whole picture: data, CRM, the knowledge base, business rules, what actions the agent is allowed to take, and how it hands things off when it needs to. That combination is what turns AI from a chatbox into something that actually gets work done. There’s a bonus here too: an AI agent doesn’t just do the work, it also creates a much richer source of data about what’s really going on. In traditional customer service, nobody can review every single ticket by hand, so many interactions simply don’t get that level of attention. Once AI handles a big chunk of the volume, you suddenly get a much clearer view of what problems come up the most, where customers get frustrated, which situations keep repeating, and where the process itself needs fixing. So AI isn’t just answering people or doing an agent’s job, it starts telling the business a lot more about how support actually works, and that’s one of the more interesting side effects of switching to an agent-based setup, for me.
And this is where Vivino is a really good example
Vivino works at a scale where you can’t just treat AI as a fun experiment. 73 million users means a massive number of possible conversations, and support still has to stay good and consistent while the team still needs room to handle the genuinely hard cases. That’s exactly where it makes sense to move past small bits of automation and go for a real agent-based setup. If AI can handle a big share of the routine stuff by itself, people get to spend their time where it actually counts, and from there it’s just a matter of scaling it up.
It’s worth being clear here about the line between what we did and what Vivino did next. We were there early, doing the Salesforce consulting, shaping the approach, building the prototype, and Vivino took it from there and scaled it up themselves. We can see the result today: a 60% autonomous resolution rate achieved within 30 days with Agentforce Help Agent. Vivino putting this case on stage at Dreamforce feels, to me, less like the end of the story and more like proof the whole direction was right.
What matters most to me here
There’s a lot of noise right now about what AI agents will be able to do next, but I think the more interesting question is a different one. How well does a company actually understand its own processes? Is its data in decent shape? Does it have clear rules for making decisions? Does the team even know which tasks make sense for AI, and which don’t?
Because AI doesn’t let you skip having a well-run business. If anything, it does the opposite: it shows you fast where the business isn’t set up well. If the process is clear, the data’s there, the rules are set, and everyone knows who’s responsible for what, AI can genuinely take on real work. If none of that’s in place, adding AI just becomes a way of trying to fix an organizational problem with a piece of software.
That’s probably the biggest lesson we took from Vivino. Don’t start with “what can we do with AI.” Start with “what part of our work do we actually want to change, and why.” That way the tech stays a tool, not the point, and that’s when AI starts changing not just how you talk to customers, but how the whole support operation runs underneath.
With Vivino, that path from an idea to something running at real scale has already happened, and now it’s genuinely interesting to see what’s next. Because in the end, 60% isn’t really the point. What matters is what’s behind it: AI went from being an experiment to actually being part of how the business runs. That’s a very different way of looking at technology.
