Meghana Atluri

Meghana AtluriSep 10, 2026

NHS England launched a national registry for AI scribe suppliers this year. To get on it, suppliers have to self-certify against requirements covering clinical safety, technology and data protection. NHS organisations can then use the registry and the evidence behind it as part of their own procurement and assurance process. That struck me because it's a very different definition of "ready" from the one you usually see in an AI demo. A model working well in a consultation isn't the finish line. Someone still has to be comfortable with the clinical risk, the data handling, the integration, the monitoring, and ultimately who is responsible when something goes wrong. That's probably the more useful way to think about healthcare AI adoption. The bottleneck isn't always whether the technology works. Sometimes it's whether an organisation can take responsibility for using it.

Meghana AtluriMay 19, 2026

There's a version of market expansion that looks great in a board deck and falls apart on the ground. I've been thinking about this a lot recently, watching a wave of companies treat Europe as a single market to enter rather than a collection of distinct ones to navigate. The product is the same. The pitch is the same. The GTM motion is lifted directly from wherever it worked before. And then it doesn't work. And nobody is quite sure why. In my experience, the diagnosis is almost always the same thing. The strategy was built for the market that was assumed, not the market that actually exists. Buying behaviour is different. Trust signals are different. The people who need to say yes, and the order in which they need to say it, are different. What counts as a credible proof point in one country doesn't land the same way two borders over. None of this means expansion is harder than it looks. It means it's different than it looks. The teams that figure it out fastest aren't necessarily the ones with the biggest budgets or the strongest product. They're usually the ones who invested early in understanding the market as it actually is, not as it was projected to be. And more often than not, that understanding came from one or two people on the ground asking the right questions before the rollout began. The best market entry insight I ever got didn't come from a consultant or a report. It came from a stakeholder in the third meeting who felt comfortable enough to tell me what everyone else in the room already knew but hadn't said. That's what local really means. And it's very hard to shortcut. What's the most surprising thing a market taught you that no amount of desk research could have?

Meghana AtluriAug 18, 2026

Most people trying to automate GTM with Clay skip the hardest step, and it isn't in Clay at all. Before I built any scoring workflow, I spent longer than I expected just interrogating my own judgment. Two accounts, same firmographics, same size, same industry: why would I call one first? I didn't have a clean answer the first time someone asked me directly. Mildly embarrassing, given I'd apparently been doing this correctly for years without being able to explain it. Most of the time it came down to something like a recent leadership change, or a specific hire that told me this company was in motion right now, not just one that theoretically fit the ICP. That instinct had to become an explicit rule before anything could score it. The shortcut most teams take is grabbing a scoring template that already exists somewhere and wiring it into Clay, treating the tooling as the hard part. What that actually does is import someone else's judgment instead of encoding your own. It'll run. It just won't be right. Nobody finds out until close rates quietly drop three months later. The judgment underneath the automation is what actually made the difference. If you can't say out loud why Account A beats Account B, a model won't figure it out either. It'll just guess with more confidence than you did. If you've tried to automate any part of your GTM motion, where did you actually get stuck: the tooling, or figuring out what you were trying to encode in the first place?