Last year Twilio created a new role - Head of Marketing Innovation and AI, a job taken on by Piyush Sagar Mishra. This isn't Mishra's first time diving into AI at the firm. In 2024, the company began experimenting with generative AI (gen AI), and Mishra was involved in designing AI-focused workstreams, including an inbound Enterprise Sales Automation (ESA) that was highly successful. But the work he was doing with AI wasn't the only reason he was chosen for this new role, he says:
If you want to build something for the company, you want somebody who understands the workflows and has played the role of an operator and a consumer. So, before moving into this role, I was managing growth marketing. And growth marketing is a very interesting role at Twilio because, fundamentally, it's a PLG business. Individuals come in, they start building on Twilio, and at some point, they realize, okay, Twilio is good for my use case, so they go more enterprise, look for contracts and discounts.
That allowed me to look at the entire go-to-market funnel, so how do we work with sales and marketing? So having that understanding of how different teams collaborate on a single, unified go-to-market motion was a very strong plus. So I had done that bit in the past. In the growth marketing role, I'd also started building a lot of AI modules and products, and ESA was one of them. But having somebody who understands what marketing teams want, who has a background in data science and understands insights and analytics, and then can also deploy AI. I think that led to Chris believing that I could come in and play this role very well.
What's interesting about Mishra is that his background isn't in B2B marketing, it's in B2C. Before joining Twilio, Mishra worked for over ten years as a management consultant focused on B2C. He brought that knowledge to Twilio, wanting to figure out how to apply B2C principles such as personalization, upsell and cross-sell, pricing, and so on to a B2B go-to-market (GTM) motion.
Mishra has been in his new role for the last seven months, and there are three things he's very focused on:
Mishra said that if you look at marketing over the last couple of years, what you see are teams running 20th-century infrastructure, trying to deliver on 21st century expectations. He said creativity in the B2B martech landscape has historically been constrained by systems and systems of record, as well as traditional approaches to delivering marketing content.
People were frustrated, and AI provided an opportunity. Mishra's team developed a new internal AI framework called MAISON, short for Marketing AI Stack for Insights, Optimization, and Navigation (Mishra also referred to it as their marketing operation system). The goal of this new function was to build trust using AI and solve for the multiple use cases that different marketing teams have. For example, help growth marketing write email copy to upsell a product to a customer. Another example, help demand generation create ads that upsell a product to a customer segment.
There are many jobs to be done in marketing, and Mishra's team has started indexing them based on the impact they can unlock for the business and the number of teams that will benefit when they're deployed. And, of course, how scalable it is. After looking at multiple dimensions of these jobs, the firm created a set of agents and a roadmap to build them.
For Mishra it has been a learning process. They looked at what other companies were doing and identified the problems they were facing. The aim was to "balance the flexibility and creativity of stochastic AI models with the reliability of deterministic rules." In other words, without a solid, accurate knowledge base at the foundation, you can't fully trust what AI is doing;
So, who do you need to build the right knowledge base in the organization? Who do you need to do a little bit of white coding and, you know, just bring the GTM hacker concept? And at the same time, how do you work with a set of people who understand the output and can vet it so we can have a human in the loop? So the way we have set up the MAISON framework is: We identify jobs to be done, we build an application, we prototype it with a live marketing team, and then we get their feedback and we iterate.
The team includes 6 full-stack marketers who understand the GTM marketing motion broadly, plus another 30-35 people the team works closely with to support beta testing new applications and features.
As to whether Twilio is building new AI capabilities or leveraging AI within its existing tech stack, Mishra reckons that in a few years, there will be two types of companies: those that use AI as a bolt-on over their existing stack and those that use AI to unlock a whole new set of workflows around their existing stack. Twilio is leaning towards the second approach.
There are two reasons for leaning towards new workflows. First, once you start deploying AI into production, roles begin to evolve. So, for example, growth marketing managers are now also responsible for managing an agent. They need to understand the underlying prompts and how to update the knowledge base and context.
The second reason, which is extremely important, relates to governance and speed, which are tightly intertwined when it comes to AI:
Speed without trust is complete chaos, and trust without speed is completely irrelevant. So we have been thinking, how do we build trust fundamentally in the stack?
Right now, Twilio is focused on building a tool-agnostic agentic workflow. It doesn't want to build specific integrations to tools (at least not in the short term). Rather it aims to build a platform that can push data and content into existing systems (like Adobe, Marketo, and Salesforce) with only one human in the loop, rather than requiring a different person to govern and guardrail each tool.
The core thesis around MAISON is that it's a trust problem, says Mishra:
If you get the data in the right format and make it trustworthy, the models will do well. And if the models are trustworthy, the end users will do well. And if they do well, then the customers will be happy. So, how do we solve for trust has been one of the key focus areas for us.
From day 1, Mishra's team has been focused on change management and has tried to mirror the workflows marketers use today because, like most other jobs, people get used to doing things a certain way and don't adapt well to change. So the team examined a typical day (and weeks and months) and productized it, making it more efficient. This has allowed marketers to come to the screen and recognize what it is.
So what does this AI platform look like? The knowledge base sits at the foundation. On top of that is the sandbox, where you find the approved models. On top of the sandbox are the agents, and then the use cases. The platform offers a lot of flexibility because you can select the best model for a given use case.
According to Mishra, building a knowledge base is what most companies initially get wrong. They get excited about AI and start rolling out capabilities without properly developing the underlying knowledge base. In their defence, they probably didn't know they had to solve that problem.
But stitching data together is very important and has to be step one, Mishra said, which is what his team did from day one. With the knowledge base foundation in place, over the last seven months, the team has shipped four agents from its roadmap.
One of these relates to Account-Based Marketing (ABM). Organizations talk about doing one-to-one personalization, but the truth is it's often segment-based ABM, and requires a lot of manual work, including creating segments and content, developing rollout plans, executing, and measuring performance.
Twilio does it differently now. It takes a list of accounts and put them in a backend process that includes a multi-agent suite. These agents perform the research and enrichment on each company, its problems, its relationship with Twilio, who the buying committee is, what the value prop looks like, and cost, and then create the content.
These are real one-to-one customer experiences, including landing pages, web, email, slide decks for Sales, even different languages - all done in 30 minutes. The content is then added to Adobe to action.
This application is still in a pilot phase, with rollout expected this quarter. Mishra claims it's a game-changer, shrinking time-to-market to a day or two from weeks or months.He adds that users can also put a human in the loop to vet, iterate, and revise the content (adding another week to the timeline).
Another use case Mishra shares relates to the Creative Studio and the question of what the image model would look like in 2026. We all know how bad most models are at handling text in images, which most marketers do. The answer, suggests Mishra, is to de-couple the text over the images. A brand team creates content for a specific market and changes the content in the image to adapt it for another market, then builds a custom text editor on top of the image, allowing them to change the text. This small thing has shrunk shipping time to days.
Along with these larger applications, Mishra's team has also built some small widgets around their 'jobs to be done' list (e.g., taking a web page, localizing it, and shipping it directly to Adobe Experience Manager (AEM) in minutes).
These three examples have been the most adopted use cases so far, driving efficiency and productivity in the market team. But more importantly, Mishra customer experiences have improved. By sending highly personalized emails, this leads to more engagement than has been seen over the last two years with three to four times the engagement rate from email copy alone.
And there are marketers who are technical and like building things with AI in Twilio. Mishra said when all this started, Twilio saw small experiments happening in different teams, but people didn't know what others were doing.
So his team took the best practices from everyone, kept the technical nuances, and gave marketers the flexibility to change the model used if they wanted. Marketers could go back and forth, selecting different models, testing the results, then choosing one and moving to production:
We have a skills document that says, for this use case, this is what works very well today, and we passed that in the channel. Most of the endpoints are configurable, so depending upon the use case, you automatically pick the model up. And in the past, the model updates would happen very drastically. Like, the entire format of the endpoints would change. I think companies are realizing that that's very hard work for consumers, so they're trying to keep the changes minimal in the core base. I hope it's just going become like a toggle in 6 months, where you can just go and say, switch from 5.2 to 5.4, maybe, and things should work as is.
Last year may have been the year companies realized they needed to start doing something with AI. But the truth is, many are still struggling with adoption and trust is still a problem; you have to make the data trustworthy, and a knowledge base is needed, MIshra reminds us:
If you do not start with that, you are gonna end up with slop. No matter how much models will improve, they don't know your institutional knowledge. When you are in a product or R&D organization, you use GitHub or standard tools, so everything is documented natively. Sales and marketing do not work like that. Marketing is basically saying, "Oh, don't look at this document; go ask Sarah." And everybody is doing that. How do you ask Sarah and institutionalize that for the agents? That's a hard problem.
I think 2026 will be the year we start getting a lot more systems and best practices for building knowledge bases. And there is so much on Twitter right now, it's like tons and tons of use cases around context graph and knowledge graph. I think it's just a data stitching problem. Five years ago, everyone said, single source of truth, customer 360. I think this is the year when we'll actually make it happen.
Two things stand out for me from this interview. First, creating a team within marketing that is focused on understanding how AI can help improve marketing functions, a team of people that truly understand both marketing and the business, is smart. This team can work closely with marketers to understand how they work and where AI can best help. It's only a matter of time before the team expands to support other, related workflows.
Second, it was clear from talking to Mishra that he's passionate about his work. Not only building with AI, but building something that marketers and GTM teams trust and can easily adopt. He often spoke of trust and made it clear that it comes down to the knowledge base that supports all AI workflows.
Mishra's team is really just getting started, and it will be interesting to check in on them in six months and see how things are going, especially given that AI changes fast and the struggle to keep up is real for all organizations.
