Our marketing team used to spend more than two weeks on account research and personalized content creation for a single campaign. Now it takes under a day, and we run as many as 15 account-based campaigns a quarter, a significant jump up from our previously standard three.
The AI model didn’t do that alone. What did was the unglamorous work underneath it. That is the unsexy part, and it’s the part most companies skip.
As Twilio’s CEO Khozema Shipchandler put it in August during our Q2 earnings, we’re not in early innings, we’re still at batting practice.
Our results are real, but they’re the exception. Most companies aren’t there yet.
When vision keynotes and customer stories promise an autonomous future, but lived experiences don’t match up, trust fades. Surveys point the same way: 86% of companies believe customers are satisfied with their AI-powered service, but only 51% of customers agree. And much of what’s promised isn’t running at scale yet. Deloitte found only 25% of organizations have moved 40% or more of their AI pilots into production. The distance between pilot and production is where a lot of companies are losing footing.
Where we actually are
I think of enterprise AI adoption in three phases. This is my own framework, not an industry standard.
Phase one is personal productivity, where individuals use AI to accomplish work faster. This is where most teams sit and will likely remain for now.
Phase two goes beyond single users to a full team moving work forward in partnership with AI through shared workflows, custom agents, and reengineered processes. Some teams are here, including ours.
Phase three is where workflows transform: processes run autonomously at scale, end-to-end. Many vendors are selling it. Few customers are running it.
The problem is that all three phases have been condensed into a single story and sold as though phase three is already here when it’s not.
What the two weeks were actually made of
Getting from two weeks to under a day wasn’t mostly about the model. Before we automated anything, we had to start with humans first. This helps us account for taste and ensure AI isn’t replacing real human ingenuity and creativity where it matters most. Without this, the output would’ve been misaligned to our brand overall. A good example is our attempt to utilize our AI tools to localize a French campaign and it added berets onto every person, which looked a bit silly and lacked cohesiveness with our overall creative standard and wasn’t something we could push live.
Another good example is data hygiene. For agents to take action on a team or worker’s behalf, they have to source clean, connected, and easily-accessible data. Few enterprises have this. One survey of data leaders found only 8.4% say their data is ready for production AI. AI doesn’t fix broken ways of working. It magnifies the cracks in the foundation.
Plus, most enterprises aren’t building on one platform alone. One team is working with Claude, while another is paired up with Copilot, and a third is using ChatGPT. On our team, it’s much the same – teams have access to ChatGPT and Gemini for content creation, and we’re working to open up Claude to our broader organization for marketing-specific work. The connective tissue between different models remains immature, and real transformation depends on it.
Brand cohesion, data hygiene, governance and integration are the unsexy infrastructure that rarely makes headlines. Neglect them and you can end up in the headlines for the wrong reasons: take the 2024 CrowdStrike outage, triggered by a faulty software update that disrupted millions of Windows devices worldwide. This cost Fortune 500 companies an estimated $5.4 billion and is a stark reminder that the quality controls underpinning enterprise technology are anything but unsexy.
Leading through the moment
What should enterprises do with this? Start by resisting the urge to overpromise, and give customers credit. The demo isn’t the shippable product, and customers know the difference. The moment there’s a stake in the ground that says there’s a consistent stitch across workflows, the obvious follow up question becomes “show the customers running this in production.”
Then, be honest. Recognize the phase you’re actually in, and embrace it. At Twilio, we’re in phase two, and building deliberately toward phase three without pretending we’re already there (or that we have all the answers).
There’s no shame in being in the middle. The companies that do the unglamorous work first, the data, the governance, the integration, are the ones that will get to phase three with results they can show.
Twilio sells infrastructure for customer engagement in the AI era. The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.