TL;DR AI has reshaped coding and customer support faster than it has reshaped marketing, and Kashish Gupta thinks the reason is that most AI-for-marketing work has been lazy. Kashish is the founder of Hightouch, which started as reverse ETL for customer data and now builds the AI marketing platform for enterprises like Domino's, Spotify, and PUMA. The hard part was never generating text. It was generating on-brand images and video without hallucinating the wrong logo or font, and the fix was a coding loop on SVG plus a per-brand context layer, not a single-shot image generator. The next unlock, he argues, is giving agents a browser so they can read the real-time signals marketing runs on.
Watch the full conversation with Kashish Gupta on Navigators.
Hightouch started as a reverse ETL tool, the plumbing that moves a company's customer data into its marketing and sales systems. From there it built a customer data platform, a self-serve layer where marketers could segment audiences and orchestrate journeys on top of that data. Over the last few years the center of gravity moved again, from human marketers pointing and clicking to agents that run marketing for you. Today Hightouch does AI for marketers: deciding which campaign to run, building it, and running it in the wild for the largest consumer companies.
We sat down with Kashish for the fifth episode of Navigators, our series on how builders are shipping with AI, to talk through why marketing has trailed coding, how Hightouch drove hallucination out of generated creative, and why the next inflection point runs through the browser.
Why is AI in marketing behind coding?
Coding and customer support were among the first enterprise functions to adopt AI, because both have a verifiable loop. You can check whether code runs and whether a support ticket got resolved. Marketing has no clean success signal like that, because so much of it comes down to taste. Kashish thinks the harder problem is self-inflicted, that teams reached for the shortcut instead of doing the context work.
The work that closes the gap is unglamorous. Every brand keeps its assets somewhere different, one in a digital asset manager like Adobe, another in Figma, another in a pile of past ad creative and ad-platform metrics. Hightouch spent about a year building the pipelines to pull all of that into one place per customer, then teaching the model to query against it. How you do it for Domino's, whose ads show people eating, is different from how you do it for Spotify, whose ads show artists and the app.
Why is content the real bottleneck?
Once marketers can run personalized campaigns, they immediately need personalized content to fill them, a different piece for every audience. That is more creative than any human team can produce by hand, so the demand lands on AI. Generating text is largely solved: retrieval over a brand's past copy is enough to write new copy. Image and video are where it breaks, because you cannot fit a brand's entire ad history into a context window, and even if you could, the model would still hallucinate.
That eighteen months went into two things: teaching models to stay on brand, and building a vector database of each brand's real guidelines as embeddings. Getting hallucination down from roughly 20% to 1% is what earns a marketing team's trust, and Kashish says the work that used to take eight weeks per brand now takes about two days.
How do you generate on-brand creative without hallucinating?
Hightouch does two things. First, it turns a brand's past advertising into embeddings and tags and teaches the model to query that database, so new creative resembles what the brand has actually shipped. This is the per-brand context layer. Second, and this is the part that fixes determinism, it treats image generation as a coding problem rather than a generation problem.
The result is variance where you want it and fidelity where you need it. Selling a shoe, you can let the background be generated, urban in one execution and sporty in the next, while the shoe itself stays the brand's real product image. Hightouch wraps this in a canvas that looks like Figma, so a marketer sees their past ads and performance, generates new ones, and drags, drops, or edits text directly instead of prompting back and forth.
Can AI have taste?
Taste is the objection every CMO raises, in two parts. The first is capturing a creative director's judgment, which is hard because they often cannot describe their own guidelines, so it has to be learned from their approved work. The second is the fear of convergence, that Hightouch's model will generate the same thing for every brand that uses it.
Hightouch runs exactly that: a blue-sky sub-agent that checks what a brand and its competitors have never tried. It scrapes competitors' ads from Facebook into a separate database and looks for white space. For PUMA, the agent found the brand overperforms on suede and certain lifestyle lines, so it leaned into that advantage while proposing categories PUMA had not explored.
Why did bring your own database win enterprise?
Data was Hightouch's first platform and remains the foundation. Every CMO has far more customer data than they use to personalize experiences. While competitors offered to store that data, Hightouch pioneered the opposite: a composable, serverless approach that plugs directly into the customer's own warehouse.
Bring-your-own-database only makes sense in enterprise, because SMBs have no warehouse to bring, and that forced Hightouch to be enterprise-led. It also compounds: as the customer's data improves, so does everything Hightouch builds on top, from the semantic layer it shipped in 2020 to the insights agents that now read against it.
What does an agent-run marketing org look like?
Solving access surfaced the next problem. Once marketers could reach all their data, they did not know what to do with it. Pre-LLM, surfacing opportunities automatically was impossible. Now research agents run around the clock, flagging customers who have gone quiet, segments getting too many emails, or audiences with no content built for them.
The same shift changed how Hightouch builds itself. Kashish and his co-founders follow PRs directly and have Claude summarize the week's commits against what people said they would ship. Solutions engineers ask Claude how a product works instead of interrupting engineering, reading straight from the codebase. Teams of one to five engineers with no PM now ship entire products, like a Figma integration that indexes a customer's design assets, built in under a week.
What is a real moat in marketing AI?
Hightouch frames advantages as thickness versus thinness. A thin advantage disappears tomorrow when someone builds it for free. A thick one is genuinely hard to build. Their thickness is in helping each brand develop its own insights and context layer, which takes real code, from scripts that tag a decade of email history to research agents that cross-reference it against first-party purchases.
Hightouch's own data strategy, oddly, is to have none. It never wants to own the data. It wins by helping brands build their strategy and writing it back to them, creating the insights layer rather than storing it. That choice keeps the moat in the work, not in lock-in.
What happens when the signal an agent needs has no API?
Marketing runs on real-time signal, and most of it lives nowhere an API can reach. A CMO can plan for the Fourth of July six months out. They cannot plan for the US losing the World Cup last night, or a storm freezing pipes across upstate New York, the just-in-time moments that make a consumer experience land. Getting that signal means reading the open web.
Hightouch is building listeners that check Twitter, Instagram, the weather, and a brand's own SKU performance, then return a report with ideas. Kashish's sharper insight is temporal: if you scrape a competitor's active ads week over week, the ads that stay live are the ones that work, so storing that history as time series turns scraped pages into a read on what is winning. That distance between what marketing needs and what exposes an API is the gap a browser fills, and it is the layer Browserbase runs. We give agents a real browser they can drive at scale, with proxies that hold their reputation, verified access to sites that check who is asking, and a recording of every session so you can see what happened when a run fails.
How does Hightouch decide what to build versus buy?
The rule is simple. Anything verticalized for marketers, Hightouch builds. Anything horizontal, it buys. For generation it uses several foundation models plus open-source ones, since different models are good for different things, and Hightouch acts as the arbitration layer on top.
When something sits in your critical path, a model or a piece of infrastructure, Kashish looks for scalability and resilience under spiky load, the kind you get when a customer asks for content against every possible World Cup outcome at once. Paul's addition: buying from a fast-moving startup can behave like an early investment, where the product you buy today is meaningfully better in a year and you are the customer getting the attention.
Where this leaves us
Kashish sees the next inflection points clearly. Image and video generation are gated by the foundation models, so Hightouch is patiently waiting on them, and each jump, like the last generation of image models, has moved the product a lot. SQL and insight generation are gated not by coding agents but by semantic understanding, which Hightouch has been building since 2020. Data, he feels, they can just keep grinding on themselves.
The throughline is that marketing lagged coding not because the models were incapable, but because the context work was skipped. The teams that win will be the ones that do it, learning a brand's real visual language, wiring up its data, and reaching the signals that live outside any API. For that last part, the open web that no integration covers, that is the layer we run. If you are building an agent that needs to reach those parts of the web, you can create your first session in a couple of minutes.