Rate the Hype: NL-to-SQL and Governance without the Vendor Tax

Summary

This video rates today's loudest data and AI trends—open table formats, semantic layers, MCP servers, vector databases, data mesh, and governance—using a four-tier framework based entirely on what Abridge deployed in production under HIPAA.
A live NL-to-SQL demo using a dbt and Unity Catalog semantic layer shows how adding business context dramatically improved query accuracy, delivering a 10x gain in data-to-insight speed.
An AI governance agent that scans for protected health information in real time demonstrates that enterprise-grade governance for regulated data is achievable without purchasing a dedicated vendor product.

Rate the Hype: NL-to-SQL and Governance without the Vendor Tax

Every year the modern data/AI landscape gets more crowded. Honest opinions about what's worth building are rare. We took some of the loudest trends (open table formats, semantic layers, MCP servers, vector databases, data mesh, and data governance) and rated each one: Adopt Now, Experiment, Watch, or Ignore. Not based on vendor pitches but on what we've deployed at Abridge, an AI-native healthcare company where we process clinical conversations under HIPAA in real time. Then we will prove it. We will present a natural language to SQL interface that 10xed data-to-insight speed for our consumers. A HIPPA scanner that shows governance doesn't require a vendor. Two engineers, two demos, and field-tested answers on what to ship in production and what you can skip.
Talk By: Christina Taylor, AI Platform Data Lead, Abridge AI ; Sumit Sirwani, Data Engineer, Abridge ;

Chapters

FAQs

Which data and AI trends are worth adopting now versus ignoring?

Based on production experience at Abridge, this video rates open table formats, semantic layers, MCP servers, and governance as Adopt Now. Vector databases are rated Experiment, and data mesh is rated Watch due to the high organizational maturity it requires.

How does a semantic layer improve NL-to-SQL accuracy?

This video demonstrates that NL-to-SQL accuracy improves significantly when AI models are given business context through a semantic layer built with dbt and Unity Catalog. Without the semantic layer queries miss business logic; with it, the system can translate natural language directly into accurate SQL against governed tables.

Why are MCP servers rated Adopt Now by Abridge?

The presenters describe MCP servers as the equivalent of TCP/IP for AI agents—a foundational connectivity standard that enables agents to interact with tools and data sources in a standardized way. Abridge's production experience led them to recommend immediate adoption rather than waiting for the ecosystem to mature.

How can healthcare companies implement data governance without a vendor product?

This video shows an AI governance agent built by Abridge that scans for protected health information in real-time streams, including live Slack data, without relying on a third-party governance vendor. The demo illustrates that teams can build HIPAA-compliant governance capabilities directly on their existing data platform.

Full transcript

[00:09] Welcome. Good morning everyone. So um every year the data and AI landscape gets more crowded, more vendors, more logos. And I used to give my students a handout called um the MAD landscape. If
[00:21] you look it up, it's in PDF format. But since 2025, they have added navigation menus just to help you comb through the map. So it looks like everyone has a pitch of
[00:34] something, but honest takes on what is worth building, that's little rare. This is why um we took some of the loudest trends in modern data and AI like MCP
[00:46] servers, semantic layers, governance, open table format, data mash, and we rated them. One out of four, adopt now, experiment, watch, and ignore. Uh, not
[00:59] based on elevator pitches or vibes, but based on what we actually shipped in production, what worked in a bridge. A bridge is a healthcare AP uh AI company
[01:11] where we ship um we we process clinical encounters and conversations in real time under HIPPA. So I'm going to bring up my colleague Sumit who keeps me honest.
[01:24] >> Hey everyone. Hi. Uh thank you Christina and thank you data bricks for giving us opportunity to present. Uh I'm Sumit Surwani. Uh I'm a data engineering leader and AI enthusiast. Uh working for upbridge. I have over 17 years of
[01:36] experience and I primarily focus on analytics and ML use cases providing data for that. >> And I am Christina Taylor. I am the data platform lead and over 12 years in data engineering space also teaches data
[01:49] engineering and writes about lakehouse architecture. Now the rating game how does that work? >> All right. Uh so let's go over the rating game once again. So what we did is we picked up some loudest trends in AI era and which are like one of one of
[02:02] the most popular ones and we tried to bucket them like is it a hype or do we need to adopt them now? Do we do we want to you know watch experiment them do a spike uh and then we are going to bucket them into these uh uh uh you know boxes
[02:15] you see and then followed by we are going to have some you know live demos thank you know hopefully everything will go well it's it's you know live demos kind of tend to have jitters but um and we are going to also explore some you
[02:28] know real but not real healthcare data I mean of course we have a sample Christina so let's go on the first one that we want to explore uh open table formats. What do you
[02:40] think? >> Uh adopt now. Well, adopt yesterday. >> Oh, that that was very fast. Uh could you you know care to explain uh why are we adopting it? >> Yeah, sure. Um Delta iceberg those are
[02:52] the settle golden standard. Why? And your your data is open. It's not locked in any vendor particular format. It's just based on parcade in your cloud environment. you take it anywhere with
[03:05] you regardless of your tool. >> So various various format like it can come in any format. Are there like more more reasons like you know why why is this like compelling enough to adopt? >> Sure thing. And I think it's faster,
[03:18] cheaper and better. So I don't have to tell you that um uh with open table formats you get schema evolution, time travel, asset compliance and other than interoperability uh uh for free. But um
[03:32] I what really made it or break it for us at a bridge is um iceberg is getting huge momentum and the v3 semistructured data support is this is actually the messy data stuff you see in production
[03:46] right so uh clinical conversation happens in real time it's not just structured columns or rows and you don't want to have to flatten every column everything in your JSON before you can
[03:58] do analytics so iceberg these three really made it possible for us and as for cheaper and better and what can better speak for it than actually a live demo de demo we're going to show you
[04:10] some benchmarks. >> Sure. Yeah. Let's dive into a demo also like is it like you know win over storage and reduce our compute costs there like you is this one of the getaway thing from this one? >> Yeah. Okay. So let's roll. Um so what
[04:24] I'm going to do is uh I'm going to load some bronze tables into my data lake. Um, so I've got um 12 uh 21,000 clinical encounters like scrambled data. Uh I'm
[04:38] going to put it into like three different formats. The original text format, it was generated by our audio recording process as well as iceberg v2 and iceberg v3. And just uh for context
[04:53] on what is a clinical encounter. So, it's a conversation happening in in real time between a doctor and a patient and it comes in a like a JSON blob in shape. That's all you need to know. All right.
[05:06] So, I've got uh my tables over here and these are the columns. So, uh I'm going to run my benchmark. >> What are we doing in this step? Like, you know, why are we creating
[05:18] benchmarks? >> Sure. So, first of all, we want a little more data. So, we're going to apply a scale factor of 12. So we have uh about uh 252,000
[05:32] rows, not just 21 21,000.
[05:45] two is we will apply a date range filter and three we do a point lookup by ID. And four it's an aggregation on two nested fields. And here are the results. So, uh, we're going to share this code
[05:58] at the end of the demo. So, you feel free to reproduce yourself. Um, but we can see that across the board, iceberg vastly outperforms CSV. Um, both in terms of compute and storage. But, um,
[06:12] there's a little bit of detail here. Uh, between v2 and v3, you can see that uh, v3 is a little bigger on disk. Um, this is where we trade off um, performance versus storage. So remember that um the
[06:25] variant type so in v3 it's just store it natively in parquet format uh where you can extract uh or or other nested fields uh versus in v v2 you often have to do
[06:39] roundtrip data conversion between string and the structure. So in V3 we trade off on storage um to gain for better query performance but keep in mind storage
[06:51] cheap compute is expensive. So no um I hope I've convinced you that um open table formats you should adopt yesterday risk falling behind if you
[07:03] don't. >> Awesome. Uh let's go on to the next one. >> Yeah. Okay. Go. I got one for you. Semantic layers. All right. Uh, semantic layer uh there is a lot of talk about
[07:16] and probably it's just coming uh into the picture once again that there used to be talk about but then AI happened and now it's one of the you know again one of the popular areas. Semantic layer I would also say adopt now.
[07:28] >> Okay. So that's a very strong word for a semantic layer. >> Yeah, it's actually a battleground. Uh so let me try to you know explain this u uh a little dumber version. Uh have you built a Lego? Uh I'm sure like you know
[07:41] everybody would have built a Lego uh and let's say you know we are kind of throwing in or I'm giving you like a thousand bricks and then asking you to you know build a car from it and then they are like not sorted and you don't
[07:53] know what what exactly uh type of bricks you have gotten or how they go well together. I do this all the time with my son and then we first start sorting and then start learning about the blocks and then we start building. Uh similarly
[08:06] like you know uh when you throw data bricks on AI uh AI also needs to know like you know how they interact with each other. Uh how will you do uh how how the referential integrity works or you know what type of data uh a field
[08:19] represents. For example, Christina just explained like an encounter is in the context it is uh nothing but a visit happening between clinician and patient tracked as an encounter and our ambient
[08:32] model kind of you know logs that uh you know whole encounter ti ties up with an ID. This detail is important enough for AI to you know learn so that it can serve the use cases. Of course, you know, uh you have I'm sure you have
[08:45] heard this a lot of times, but AI can be as accurate as the data you feed into it. And that accuracy is defined at the semantic. >> That's a clever analogy. Data breaks for AI. Um
[08:59] need to kind of dumb it down for the agents? uh thumb down is not the phrase I was you know going for but again you know uh just supplying enough context so your engineers are not frustrated with the wrong decisions or the wrong thing being
[09:12] built uh you're not burning tokens uh you will burn less tokens if your data is you know clean enough also it's tagged really well. So again not dumbing it up but again you know supplying all additional context to it. So maybe it um
[09:25] dumb it no smart it up like or so that it doesn't outsmart itself or just try to give me the most verbose answer in the spirit of token maxing. Okay, I got you. Um
[09:37] >> nice adopt now offer semantic layer. >> Let's dive into demo for semantic layer. >> Uh yes, >> let me just clear out the screen. Let me
[09:49] ask me some questions. But first of all um I am going to build some silver tables and gold tables based on our bronze data. So I'm using dbt here uh and the dbt spark package in a local
[10:03] environment. Uh okay >> Christina just run make dbt command. I know the screen is a little bit you know uh shadowed by the prompts there as well. >> Yeah. So I promise I am going to make it a little more visual for you. Uh, so we
[10:16] can actually just look at a data lineage instead of this um verbose script. Okay, sweet. I got my um DB docs generated
[10:29] my favorite feature the lineage graph. So you can see from bronze encounters we have uh clean the data uh we have did standardization. So we got our silver layer and some gold layers with separate
[10:43] metrics uh like there's encounter metrics um patient 360 view and deidentified data. So that's what DEI stand for uh while I'm here. I'm also going to register the tables in Unity
[10:57] catalog which we'll use for the second half of the demo. So um this is our governance layer. Um I actually have two catalog here locally.
[11:10] I'm using um the iceberg catalog and open source unic catalog sits on top of it. Uh read only as um the governance layer. And the reason for that was explained in a little greater detail in
[11:23] uh the repo. But uh I have a UNC catalog docker image pointed to the the second latest version on dockerhub.0 um 040. Okay. Sweet. Um, let me also
[11:38] bring up the Unity catalog UI. Okay, server is running. Now we got uh bronze tables and then all these uh silver
[11:50] tables and gold tables. Okay. Uh, great. And looks like everything is working as expected. Uh, now ask me some questions.
[12:02] >> All right. So this is basically a natural language to SQL kind of you know converter that we are going to demo. So let's uh and then again you know bringing back to the healthcare real but not real data and then you know encounter context uh let's actually you
[12:14] know uh try to see if we can derive a SQL uh to look at the customer leaderboard. Everybody like you know likes to uh see the how the competition is going within the customers who is most engaged and least engaged. >> Uh okay show me the customer leaderboard
[12:28] >> and then we have not provided any context. uh everything you know is in is in our data model and then uh at the semantic layer and as you can see uh it was able to uh you know generate the query give us the results and submit
[12:40] care network kind of leads with the highest uh over 5,000 encounters but again you know we never provide this context in our prompt uh that encounters define the leaderboard criteria >> so it knows to look at the encounter
[12:53] metrics in uh in our gold data set and I guess it also knows how we can measure customer activity. Let's try another one. >> Let's let's bring in some ambiguity now. So, let's look at trend uh of
[13:05] recordings. We'll use recording as a as a term. Uh you know, how how how is the monthly trend of recordings? Let me just do like a little more some recent stuff. Let's see if we can figure
[13:18] out what we mean by monthly recording user.
[13:31] this is the uh you know this is how the monthly trend look like. >> Yeah. So semantic layer has the company knowledge to know that recording in our sense means clinical conversations or aka the encounter document uh not maybe
[13:44] the recording we're doing as of right now. So let's try one last one. >> All right. Uh we'll again you know again we are just demoing the text so we'll ask who is our least engaged customer. >> Okay. Uh
[14:03] let's see if there's some outreach work we need to do. >> Again the engagement definition is something that we are not supplying to the uh to the agent that is running. But again the least engaged customer is granite state with like you know the least number of encounters that that has
[14:15] been locked for them. >> Yeah. Our customerf facing team often uses such tools and as a followup. Uh but the point we're making is the semantic layer is the translation layer between um your data and your business
[14:29] logic. So adopt now. >> Awesome. Let's go to the next one. Uh this is for you. Uh vector databases. >> Ah
[14:44] >> Experiment. this is AI summit and then I'm sure like you know there are a lot of you know companies selling vector databases and branding could you could you explain why experiment >> yeah for sure and uh vector search and vector the the concept itself is really
[14:58] worth learning so if you recall originally um a vector is just a list of numbers right so you know um there's vectorized scanner for parket files and you know there's um vectorzed engine for
[15:11] data bricks photon they they I just in that they are able to process a large number of things at the same time instead of one at a time. So a vector is a list of numbers and in terms of a vector search. So they are just um think
[15:25] about a vector in space it's a representation of their distance and the nearest neighbor. To give you a concrete example, uh when you say a king, a queen
[15:37] and a monarch, they are going to show up uh close in space. And I do hope the distance between the the king and the monarch and the queen and the monarch are the same one.
[15:51] learning and it's real for semantic search, right? And if you have uh hundreds of millions of vectors and that kind of scale uh what I want to remind you is the major CL platform they
[16:04] already ship it. So vector search is a first class citizen in data bricks and vector itself it's a type in snowflake and in bigquery you can also do it right in SQL and postgrass has pg vector and
[16:18] uh like Oracle all the clouds providers they're all in. So there is a native solution and you can start with. >> All right. So you're not saying like never never use it's just you know uh
[16:30] explore what we have in house uh first and then never use the >> Yeah. So not at all definitely experiment and this is real for AI applications and what I'm saying is uh
[16:42] start with what you have and you can reach for a dedicated vector space when you have hundreds of millions of vectors and sub uh subsecond latency and that kind of use cases. So um experiment
[16:55] >> awesome. >> All right, let's go to the next one. >> Okay, so I have one for you. Um, data mash. I really like the sound of it. >> Data mesh. Uh, it's it's uh, you know,
[17:08] it's very popular. Again, you know, it's it's it's I would say it's also kind of a dream as well. Uh, it requires uh, you know, a lot of effort. So, right now, I'm going to say watch >> what? But people love data mash.
[17:20] >> Uh, they do. I know. And then again, you know, as I as I was saying, like it's it's kind of a dream where, you know, every team kind of owns the uh, their own data decentralizing data. But you know what? data eventually has to come
[17:32] together to serve use cases like business intelligence. So anyway like you know all your products data would eventually combine uh to help sales to help like you know uh uh people uh to
[17:45] promote uh your products when it combines it presents challenges uh that's why I said watch because you know we there are like you know few few things uh that all orcs should have uh
[17:57] like uh they have to follow their compliance they have to you know uh they have to have like data engineering skills they have to have platform skills to manage the data within their pods irres uh you know instead of having a centralized uh centralized data
[18:10] engineering team. >> H so who actually pulls it off? >> Uh I would say mature uh very mature orgs who have like you know already figured out how to use their data for
[18:22] which use cases. They are able to serve from uh you know uh they they are able to own a product within the pod end to end. Those are the ones like you know could could could do better but you know
[18:34] for the startups or for for the companies they are still uh figuring out how to use the power of the data uh and then how to be uh you know still be compliant uh using this using their data. It's kind of again you know uh I
[18:47] would say watch if it if you have figured it out you can definitely you know try it out. >> H I see. So I guess if I'm data engineer one just trying to build out my first reliability pipeline maybe I can skip this one for now. Gotcha. Um watch
[19:06] >> all right uh this is I'm sure like you know you have heard this multiple time and then lot of sales a lot of sales pitches as well and uh AI agents kind of you know all of this so MCP model uh context protocol what do you think Christina
[19:18] >> h adopt no definitely matter of fact this is the entire plumbing the whole agent AI platform runs on >> okay uh could you explain a little bit
[19:30] more uh when you say like the whole agent platform learns into it like what exactly they do. >> Yeah. So um again I guess if I'm explaining this to your seven-year-old son uh this is like MCP servers are USBC
[19:44] for AI agents. They are like uh the TCP protocol for the internet. So just think about as an open standard for connectivity between these um AI models and your tools. So things like cloud and
[19:58] codeex uh genie they are very very smart and they're really powerful but they can't do anything without access to your data and MCP provides such an interface
[20:11] and this is also the whole theme of today right and open data lake open format open source and open standards and I'm a CP is this part of this standard
[20:23] >> all right uh by the way I will code TCP in front of my son, he'll just say it's just some sort of TV. But again, uh uh does it like you know when you said plumbing as well, I caught that about like you know is it opening the doors uh
[20:37] for your agent to access like different systems? >> Yeah, actually hold on to that thought. Um exactly. So I I I just thought MCP server is the door to your entire data estate. But if you think about would you
[20:49] leave your house u wide open to the public all day long 247? Well, you you probably wouldn't do that. So, we need some kind of lock and we we want to make sure that the right people have the
[21:02] right access at the right time. Um, so the principle is the agent should never have more access than uh the person whoever you are giving it access to,
[21:14] right? So, which I guess leads us naturally to our last >> that's so that's the catch for like you know adopting. So adopt but again uh yes make sure that you know we
[21:26] >> open the door but keep the lock and the key. Uh I got one for you governance. >> All right so this is kind of a good segue to to our last one. Uh again as we talked about MCP uh adopt but open the
[21:40] door but again you know put a lock on it. So this one I would also say adopt now and this is this is non-negotiable especially when you're using AI agents and then you know you're using MCP and uh giving giving it a bunch of accesses.
[21:54] >> Oh okay. So I think you might get some push backs when you mention non-negotiable especially say oh lock it down and slow down that that's the worst thing that you want to hear. This is Silicon Valley. Can you explain that? Well, yeah, governance kind of, you
[22:07] know, can be interpreted as, you know, slowing you down, but it's also very important specifically for like, you know, uh, industries like healthcare that we are in. Uh, HIPPA is a real thing and, uh, we have to be compliant.
[22:20] We must be compliant and protect sensitive data to our patients. How we use it, where where do we store it, how long we store it, are we deidentifying it, uh, and, uh, who have access to it.
[22:32] So all these you know are like you know some of the uh limitations that we should exercise uh implementing like you know uh the better governance framework for our AI agents not only for the
[22:44] databases but now for you know agents as well. >> H okay so uh how about we actually prove it in some demos? >> Sure. Uh it's my turn for the demo and I'm going
[22:56] to be a little brave here and then share my slack which is a live Slack I uh and um I'm going to engage uh one uh AI bot which I already created in the background in interest of time. Uh but
[23:08] again you know this is the interface and there's an LLM running behind it. I named my bot as hippo lens. What it's going to do uh is I'm going to ask it to run a scan and I have the internet is you know fast enough here because LM is
[23:21] running uh I already get the response. What it did it basically scanned what Christina created uh uh in our data model the you know medallion structure it scanned if tables could contain PHI
[23:35] and then how it knows is based on the semantic layer uh and then this is what uh you know it came up with with some of the risks for the tables there. Um can you explain what it means here like critical high medium?
[23:48] >> I'm just going to ask about uh every single thing you know we we tend to ask AI agent to clarify and then you know I'm just I just asked it and then you know it kind of explains like why it's uh it's calling something critical
[24:00] versus high for critical it says oh you have a table but you have you don't have governance enabled on it. Sorry I see that it's probably not visible but yeah it says you have a table but you don't have a governance enabled on it. It's
[24:12] definitely uh you know uh critical enough because it thinks that it has PHI in it and uh this this is catching my eye de it's it's why a agent is flagging
[24:24] something that is like you know deidentified. So I'm just going to ask it. So this is just to demo you that you know uh how we have simplified being compliant with an agent. we are not you know play paying a third party
[24:36] to monitor the data or you know monitor the access control but the AI agent is able to do for us and uh when it makes mistakes uh uh we can provide a feedback so I'm right now asking okay why are you
[24:48] flagging DE ID uh it says oh because it sees like you know birth here or some sort of you know patient identifier information but you know what it's hashed so I don't want to you know uh see this model next time so I'm just
[25:00] going to ask it to put it in an exclusion criteria because this is deident deidentified and then we have verified it. I know we are like moving a bit faster on this one but
[25:12] uh once I you know ask it to exclude it's able to you know uh prepare an exclusion list and then uh uh uh add this model there so that it's not part of the scan uh or not flat in the scan
[25:25] next time. So I'm just going to run a scan once again. All right. uh as you can see like it's not flagging the DE ID anymore. This is also one of like you know one of the
[25:37] most important gift that you can give your compliance team as well because it's something that you know we built on the slack. Everybody uses Slack and then you know uh there's a dashboard uh coming in where your compliance team can
[25:50] see what are the findings and how many models we have. I know it's a very small structure but within an arc it's going to be kind of you know hundreds or thousands of tables uh sensor data you
[26:03] know scattered scattered there and then you are I'm sure you are managing somehow like you know the governance but again you know this is something that is watching over the door that you have opened something like ADP kind of you know watches after installing security
[26:16] cameras uh cameras but again you know in this case you don't have to pay the vendor tax >> so how does it know that what your rules for so how does the gain the knowledge of what is compliant and what is not. >> So I'm going to show uh this
[26:29] architecture. So this is basically uh what we demoed uh what Christina created uh the medallion structure the unity catalog where the columns are tagged uh uh having PHI or nonPHI. It also uh you
[26:43] know talks to notion we use an as an example via MCP where you could store your policy docs or compliance can store policy docs. uh we have like you know governance policies. It can be in GitHub
[26:55] or it can be like you know a separate config. Uh but again with the with the MCP it can talk to pretty much anything. It talks to DBT and it can be interacted. You can choose the interface
[27:07] Slack uh teams anything like you know you can you can enable this boton and uh yeah uh that's how it makes it decisions. >> Yeah that's pretty neat. It looks like uh uh the agent does have company
[27:19] knowledge and it's able to learn from well the door the the different connections it had to internal systems. >> Yeah, it's it's it's one of the powerful you know uses of kind of you know AI to
[27:33] uh make governance not not uh something a blocker or slows you down or bottleneck. That's that's what I was looking for.
[27:46] uh six things um between adopt now watch experiment and uh skip. So what do you think like AI does everything AI assists everything?
[27:59] >> Oh yeah uh we have to evolve with with AI that's what that's what we we think at a bridge as well and then you know of course you know AI does most of the things but we are kind of you know the ones who are accountable. So uh how do
[28:13] you think like you know the data engineering role we we being the data engineer should be evolving with AI? >> That's a really good question. I get that from my students all the time. So I give you my honest take here is um AI
[28:25] doesn't replace data engineering. AI needs data engineering. Don't ever let AI tell you what to build. You tell it what to build. So um give it a semantic layer. Give it the MCP connectors. give
[28:39] it the right access to the right tools but uh you do everything else that's the highest level thinking and the judgment so the battle tested answers that they cannot just tell you um again and if
[28:52] there's one word of advice I have is don't let the bot tell you what to build and you tell it what to build and what's your hot take >> uh no I'm in 100% agreement here so we
[29:05] kind of you know demoed here all the latest trends and how basically AI can can work around uh and simplify your tasks. >> So rate the hype yourself and and build
[29:17] what you can know. Thanks everyone. >> Thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.