What are the chances the legacy systems cease to exist altogether if they take a defensive approach and innovators move elsewhere? For example, Xero just implemented API fees (defensive) and are not innovating. It seems like a great opportunity to build a great front door on top of the system of record, capture new innovative companies and also eat the incumbents existing user base.
Having just built an agentic reporting platform on Omni + Databricks, I found myself nodding along hard with this take.
Two things feel undeniably true:
• Agentic analytics isn’t hype — it’s an accelerant for analysts at every level. I never expected agents to write the majority of my SQL, but here we are.
• Agents ruthlessly exposes weak data governance. If your foundations are shaky, the agents won’t save you; they’ll just help you discover the cracks faster. Now is the moment to pay down that tech debt everyone’s been politely ignoring.
On the one hand, the author gets it right: once agents start taking action, ambiguity around “what is canonical” stops being a pedantic disagreement between finance and sales and starts breaking workflows.
However, in practice you can’t just slap agents onto a semantic layer and call it universal integration. Real systems still have to architect around data latency, access controls, cost management, and a lot more compute than you’re probably used to planning for. There's a reason OLAP and OLTP are separated (usually).
Where the article gets it right: All the old “boring” best practices — sanitized models, governed metrics, shared definitions — suddenly matter more, not less. There’s a reason big players like Microsoft, Google, and Databricks all want their stack to be the place where your semantic layer lives.
Whoever owns the language of the data — the metrics, definitions, and calculations — becomes incredibly sticky.
Sick read mate - really enjoyed this, and it maps to concepts I’ve been thinking about in banking.
Banks feel like one of the clearest examples of “systems of record that won’t disappear,” largely because AML, compliance, settlement and regulatory obligations are almost impossible for the next wave of fintechs to replicate.
What has changed, though, is that fintechs can increasingly sit entirely at the front door: leveraging a bank’s compliance stack, tokenising a virtual card on top of an existing card, and focusing purely on product, UX, and orchestration / initiation rather than balance sheet or regulatory heft.
It feels very similar to the GDS → OTA transition you describe - the system of record remains essential, but value accrues to whoever owns initiation, intent, and distribution.
Curious how you think this plays out long-term in banking?
Jamin, where do you think Bring Your Own Cloud (BYOC) fits in this paradigm? Nuon just recently announced that they were going open source which is a smart move given that many of their future potential clients are very large enterprises who are hyper sensitive to keeping their data safe. This will give those organization's internal developers time to fiddle with the code and advance the mission.
On the Open Source Pod last week (link below), the CEO Jon Morehouse gets into the reasons why they did this. It would seem the ultimate conflict is between customers and SaaS providers. When agents finally grow up and mature, customers will want to give them access to their data across tenants and have control. SaaS providers may be willing to offload the responsibility of storing and securing data to their customers, but I understand that they are hesitant to give up control given that they will lose the lock-in effects associated with storing your data with them.
This seems like a similarish dynamic to Apache Iceberg/Tabular. In that case, compute was separated from the storage layer. In another interview with the CEO of Eventual (link below) he posits that storage infrastructure is probably as good as it needs to be and the issue is that the compute/query/pipelines and all of that jazz needs to improve.
Thanks for sharing your thoughts. It seems those legacy SaaS companies are building their front end too, either building agents directly or providing tools to customers to build their own agents. The problem is the growth of their own agents isn’t impressive (yet), eg, Salesforce, ADBE, Workday, etc. However, their platform business, ie, providing tools to customers to build their own agents, has picked up growth, eg, Salesforce. So, to compete with LLMs’ agents, what are legacy SaaS’ competitive advantages? Existing data may help their platform business. But not sure there actual performance comparison between the self-built agents which is based on existing data and database vs. agents provided by LLMs.
Loved these two posts on systems of record and truth. Thanks, Jamin!
One additional consideration to ponder….these data platforms don’t seem to make money in their business models. E.g MDB, SNOW. Years of unprofitable growth…they have to give away their product to stay relevant in the market.
> asked ChatGPT to create a graphic for this, and it did a pretty good job! I coulnd’t get it to reverse the arrows from Airlines / hotels / car rentals going to the GDS (instead of the GDS pointing to them).
What are the chances the legacy systems cease to exist altogether if they take a defensive approach and innovators move elsewhere? For example, Xero just implemented API fees (defensive) and are not innovating. It seems like a great opportunity to build a great front door on top of the system of record, capture new innovative companies and also eat the incumbents existing user base.
I don’t think they will. Everyone knows the high profit margin in the front end business.
Having just built an agentic reporting platform on Omni + Databricks, I found myself nodding along hard with this take.
Two things feel undeniably true:
• Agentic analytics isn’t hype — it’s an accelerant for analysts at every level. I never expected agents to write the majority of my SQL, but here we are.
• Agents ruthlessly exposes weak data governance. If your foundations are shaky, the agents won’t save you; they’ll just help you discover the cracks faster. Now is the moment to pay down that tech debt everyone’s been politely ignoring.
On the one hand, the author gets it right: once agents start taking action, ambiguity around “what is canonical” stops being a pedantic disagreement between finance and sales and starts breaking workflows.
However, in practice you can’t just slap agents onto a semantic layer and call it universal integration. Real systems still have to architect around data latency, access controls, cost management, and a lot more compute than you’re probably used to planning for. There's a reason OLAP and OLTP are separated (usually).
Where the article gets it right: All the old “boring” best practices — sanitized models, governed metrics, shared definitions — suddenly matter more, not less. There’s a reason big players like Microsoft, Google, and Databricks all want their stack to be the place where your semantic layer lives.
Whoever owns the language of the data — the metrics, definitions, and calculations — becomes incredibly sticky.
(edited by ChatGPT)
Sick read mate - really enjoyed this, and it maps to concepts I’ve been thinking about in banking.
Banks feel like one of the clearest examples of “systems of record that won’t disappear,” largely because AML, compliance, settlement and regulatory obligations are almost impossible for the next wave of fintechs to replicate.
What has changed, though, is that fintechs can increasingly sit entirely at the front door: leveraging a bank’s compliance stack, tokenising a virtual card on top of an existing card, and focusing purely on product, UX, and orchestration / initiation rather than balance sheet or regulatory heft.
It feels very similar to the GDS → OTA transition you describe - the system of record remains essential, but value accrues to whoever owns initiation, intent, and distribution.
Curious how you think this plays out long-term in banking?
Jamin, where do you think Bring Your Own Cloud (BYOC) fits in this paradigm? Nuon just recently announced that they were going open source which is a smart move given that many of their future potential clients are very large enterprises who are hyper sensitive to keeping their data safe. This will give those organization's internal developers time to fiddle with the code and advance the mission.
On the Open Source Pod last week (link below), the CEO Jon Morehouse gets into the reasons why they did this. It would seem the ultimate conflict is between customers and SaaS providers. When agents finally grow up and mature, customers will want to give them access to their data across tenants and have control. SaaS providers may be willing to offload the responsibility of storing and securing data to their customers, but I understand that they are hesitant to give up control given that they will lose the lock-in effects associated with storing your data with them.
This seems like a similarish dynamic to Apache Iceberg/Tabular. In that case, compute was separated from the storage layer. In another interview with the CEO of Eventual (link below) he posits that storage infrastructure is probably as good as it needs to be and the issue is that the compute/query/pipelines and all of that jazz needs to improve.
Nuon interview
https://open.spotify.com/episode/0h4GATlJKgIvND1RHyoErL?si=a98e0833066044a8
Eventual interview
https://open.spotify.com/episode/28g7CkjxPajcmYcxyP2j5F?si=0d88bef738b54a4b
Thanks for sharing your thoughts. It seems those legacy SaaS companies are building their front end too, either building agents directly or providing tools to customers to build their own agents. The problem is the growth of their own agents isn’t impressive (yet), eg, Salesforce, ADBE, Workday, etc. However, their platform business, ie, providing tools to customers to build their own agents, has picked up growth, eg, Salesforce. So, to compete with LLMs’ agents, what are legacy SaaS’ competitive advantages? Existing data may help their platform business. But not sure there actual performance comparison between the self-built agents which is based on existing data and database vs. agents provided by LLMs.
Loved these two posts on systems of record and truth. Thanks, Jamin!
One additional consideration to ponder….these data platforms don’t seem to make money in their business models. E.g MDB, SNOW. Years of unprofitable growth…they have to give away their product to stay relevant in the market.
Great piece!
great post. happy holidays
> asked ChatGPT to create a graphic for this, and it did a pretty good job! I coulnd’t get it to reverse the arrows from Airlines / hotels / car rentals going to the GDS (instead of the GDS pointing to them).
Gemini/Nano Banana would have done this easily