The reasoning layer that sits above the database, and that increasingly treats the database as infrastructure, is where a new generation of companies is being built.
Directionally right, structurally incomplete. This analysis assumes a stable product layer between the model and the customer where new companies get built. That layer is already being compressed. Anthropic just shipped dreaming (cross-agent pattern extraction), outcomes (self-grading), and multi-agent orchestration as platform primitives. Memory, learning, and coordination are becoming model infrastructure, not startup opportunities. Meanwhile Salesforce is building MCP endpoints, which means the system of record is conforming to the model layer’s protocol on customer demand. That’s not a platform move. That’s infrastructure accepting its role. The next decade of enterprise value probably doesn’t land in a “system of intelligence” product layer. It lands in the model platform itself, with every SoR reduced to a well-governed API endpoint.
The seams matter more than the agent loop, and a lot of enterprise AI risk will show up where one system writes into another system’s record without shared authority, auditability, or repair.
The piece I would add is that a governed API endpoint is only as trustworthy as the operating model behind it.
Before an agent writes to a system of record, the enterprise still has to answer: what does this record mean, which system owns it, who has authority to change it, what state is it legally in, what downstream obligations does the change create, and can the event be audited or reversed?
That is the layer we are building around at Mimir Labs.
Not another system of intelligence. Not just another app layer.
A governed operating substrate that makes operational truth machine-legible before agents act across systems.
Protocols can govern the seam.
But something still has to govern the truth behind the seam.
I agree with Anurag here too. One other thing that creates a flywheel for model companies is that they are quickly becoming an interface too. Claude code and cowork is clearly signalling that. And once you have paid millions to Anthropic and they make 2 FDEs sit at your org and basically agentify every workflow possible, it would be impossible to move to any other specialised layer. Intelligence layer is not a vertical play, it’s a platform play and model companies are already moving in that direction.
The absorption pressure is real. Anthropic shipping dreaming, outcomes, multi-agent orchestration as primitives is what foundation platforms do. A lot of the speculative startup surface from 2024-2025 collapses into model features.
The MCP read is inverted, though. Salesforce building MCP endpoints isn't the SoR conforming to Anthropic. MCP is an open protocol. Any model vendor can speak it. What Salesforce is conforming to is the protocol, not the platform. That's substrate emerging because no single vendor can unilaterally provide it.
The Bau Lab failures sit on this point. Eleven cases in two weeks, none of them inside a single platform's agent loop. All at the seams between systems, where one vendor's agent wrote into another vendor's system of record without verifiable identity, versioned audit, or a repair path. Platform-internal coordination doesn't reach those failures. They live outside any one vendor's enforcement boundary.
The substrate has to handle two cases the model platform can't.
The first is the managed agent. The well-behaved one, running on a foundation model, granted autonomy by a sponsor, writing into systems it doesn't own. We are building exactly this stack — Tamed Autonomy on top of Kinetic Trust Protocol. Earned autonomy requires verifiable identity, declared intent, cost-of-signal, and a repair path that survives vendor change. Those are protocol primitives, not product features.
The second is the weaponized agent. The one that doesn't speak your platform's protocols at all. Trained adversarially, deployed against your write loop, designed to look like a managed agent long enough to commit changes. Foundation model platforms can secure their own agents. They can't define the boundary that separates yours from someone else's. That boundary is substrate.
Both cases share the same requirement. The model platform can govern what its agents do inside its loop. It cannot govern what arrives at the seam.
That's the layer. It won't be a startup. It will be a protocol stack, and the companies that build on top of it will be larger than the companies that try to own it.
What struck me reading both pieces is that the transition may be even broader than “systems of intelligence replacing systems of record.”
Historically, a large amount of GTM headcount has effectively existed to compensate for fragmented systems, disconnected workflows, incomplete telemetry, and poor operational visibility. Humans became the orchestration layer between platforms.
If AI increasingly absorbs prioritisation, context synthesis, coordination, note capture, workflow execution, and institutional memory, then the organisational structure itself likely changes alongside the software stack.
The really interesting shift may be that gravity moves not just from database → intelligence layer, but from human orchestration → operational infrastructure.
Feels like we’re entering a phase where the durable value is increasingly in telemetry, execution context, orchestration, and ownership of the action layer rather than the interface itself. Very interesting to see where this series leads.
The Facebook analogy is more precise than it might seem at first. The newsfeed took the TAM of social media from "staying in touch with people you know" to "consuming everything of interest from anyone." The same expansion is happening here. The CRM addressed maybe 10% of what a salesperson actually does and the intelligence layer can plausibly address the other 90%, which is why this reads as a TAM expansion story and why the value creation will be much larger than the value migration.
The structural tension worth watching though is the race condition underneath. The intelligence layer's moat comes from accumulated institutional context, call transcripts, deal patterns, coaching signals, rep behaviour all ingested over months and years. But it needs API access to the system of record to accumulate that context in the first place and the system of record vendor can restrict or replicate that access whenever it wants. so the real question is whether startups can build enough proprietary institutional memory to become too costly to rip out before the incumbent absorbs the orchestration layer into its own stack. thats a race between context accumulation speed and platform absorption speed, and the clock is the enterprise renewal cycle.
Great framework, but I think it stops one layer short. A system of intelligence that orchestrates structured data across CRM, calendars, and call transcripts is a massive leap forward. No argument there. But it's still only weighing what already gets measured.
The next layer is what I'd call a system of understanding. One that goes beyond orchestration to surface the qualitative intelligence that never makes it into any structured system: the operational friction, the institutional knowledge trapped in people's heads, the real reasons behind the numbers. Then it weighs that signal against the quantitative data sitting in your systems of record.
That's where prediction starts. Not from faster access to existing data, but from generating new data that didn't exist before and layering it against what did. The article hints at this with the institutional memory problem during rep turnover, but the solution isn't just ingesting more transactions. It's organizational discovery at scale.
We're building this at Privagent. The system of intelligence is a necessary step, but understanding is where the durable value will compound.
The orchestration layer doesn't break the hostage dynamic. It resets it. Once a system of intelligence has ingested two years of call transcripts, account context, and rep behavior, the switching cost isn't lower than Salesforce. It's higher. The CRM held structured data. The intelligence layer holds reasoning and institutional memory.
The buyer who feels liberated by replacing Salesforce in 2027 will feel trapped by their orchestration vendor in 2031. AI doesn't change the underlying incentive. It just gives it a new vocabulary.
If salesforce is able to build better GTM agents then they win. If a non-Salesforce agent can openly access MCP/APIs from Salesforce to do the GTM job Salesforce do not win the orchestration layer automatically. But yeah sure, Salesforce can win as they have the database and the head-start on offering all the value added agent orchestration with it.
thanks for sharing this stuff! I'm immersed in software for decades; things are moving so fast now, it's great to get a peak into the agent driven future
This is the right direction, but I think there is a missing layer between system of record and system of intelligence.
A system of intelligence can orchestrate across CRM, email, billing, product telemetry, support, and workflow tools. But before agents act across those systems, the enterprise still has to resolve authority.
Which record is authoritative? What does the entity mean? Who owns the definition? What state is it actually in? Who is allowed to change it? Can the event be traced later?
If those questions are unresolved, the intelligence layer does not eliminate ambiguity. It operationalizes it.
At Mimir Labs, this is the layer we focus on: making operational truth machine-legible before AI orchestration depends on it.
The next platform shift is not just from system of record to system of intelligence.
It is from fragmented records to governed authority.
The piece that stands out from the engineering side: orchestration gravity isn't really about reading signals across many systems and that's really the easy part. It's the write-back path that's hard. Permissions, idempotency, audit trail, race conditions when two agents update the same CRM record. Whoever builds that boring infra layer ends up owning the new "database."
Good Read.. This is a strong way to frame the idea.
The move from systems of record to systems of intelligence is real and well explained. The Facebook example works well.
However, we can take the idea one step further. Intelligence is not the final goal.
Reasoning across systems, finding insights, and prioritising information are useful. But they are not enough. The enterprise does not just need a smarter dashboard. It needs work to be completed.
The next stage is not just a system of intelligence.It is a system of cognitive work.
In this model, AI does more than analyse and recommend.
It actually completes tasks. It takes ownership of outcomes.
For this it should work across systems like SAP, Salesforce, Workday, and ServiceNow.etc
It does not just inform humans. It finishes the work for them. At the same time, it follows policies, ensures auditability, and allows human oversight.
The article says orchestration is the new centre of gravity. I would refine this idea.
Completion is the new centre of gravity.
The real value will come from systems that finish work, not just think about it.
Good article. But are systems of records just a database + UI? Not quite. There’s alos the business logic. Most of the work goes into the business logic. When I click a button to “Run Payroll” in Gusto, it doesn’t just update a database somewhere. It does a lot more. The button and database aren’t even the important part.
The Jason Lemkin data point is the cleanest signal here: 10+ Salesforce seats down to 2 human plus 1 API seat, but spend up 83 percent from $12K to $22K. Pricing migrates from per-seat to per-action while the database stays sticky for the same compliance and audit-log reasons that always kept it sticky. The risk to the system-of-intelligence thesis is that CRM usage is rising in your own GTM survey, which means the agents are reinforcing the SoR gravitational pull at the same time they are abstracting it. Watch whether Salesforce's headless API pricing leaves room for the orchestration layer to capture the upside, or prices it out before the new layer scales.
If you are managing commercial properties the gateway to gathering the condition of your asset is the camera. In Multi-Family and Single-Family rentals they trade and rent on their condition. Check out www.mytracker.ai for the first mobile only camera app to open up your new system of intelligence. #computervision
The market call is right. Orchestration above the database is where the next decade of GTM value accrues. CRMs become inputs.
The thesis doesn't address one layer below: what makes the "trusted vendor" trusted, when the vendor's agent is pulling signals from a dozen sources and writing structured changes back into the system of record. That's the surface the Bau Lab's Agents of Chaos paper documented as catastrophic earlier this year — eleven failure cases in two weeks, all traceable to missing substrate primitives (verifiable identity, versioned audit, cost-of-signal, capacity bounds, repair protocols).
When humans owned the write loop, trust was a marketing word. When agents own it, trust becomes a concrete protocol-level engineering problem.
That's the next investable layer. The orchestration layer is where value goes. The substrate beneath it is where the failure modes live — and the open-infrastructure version of that substrate will likely be a larger long-term play than any captive container.
Directionally right, structurally incomplete. This analysis assumes a stable product layer between the model and the customer where new companies get built. That layer is already being compressed. Anthropic just shipped dreaming (cross-agent pattern extraction), outcomes (self-grading), and multi-agent orchestration as platform primitives. Memory, learning, and coordination are becoming model infrastructure, not startup opportunities. Meanwhile Salesforce is building MCP endpoints, which means the system of record is conforming to the model layer’s protocol on customer demand. That’s not a platform move. That’s infrastructure accepting its role. The next decade of enterprise value probably doesn’t land in a “system of intelligence” product layer. It lands in the model platform itself, with every SoR reduced to a well-governed API endpoint.
This is the right layer of the argument.
The seams matter more than the agent loop, and a lot of enterprise AI risk will show up where one system writes into another system’s record without shared authority, auditability, or repair.
The piece I would add is that a governed API endpoint is only as trustworthy as the operating model behind it.
Before an agent writes to a system of record, the enterprise still has to answer: what does this record mean, which system owns it, who has authority to change it, what state is it legally in, what downstream obligations does the change create, and can the event be audited or reversed?
That is the layer we are building around at Mimir Labs.
Not another system of intelligence. Not just another app layer.
A governed operating substrate that makes operational truth machine-legible before agents act across systems.
Protocols can govern the seam.
But something still has to govern the truth behind the seam.
I agree with Anurag here too. One other thing that creates a flywheel for model companies is that they are quickly becoming an interface too. Claude code and cowork is clearly signalling that. And once you have paid millions to Anthropic and they make 2 FDEs sit at your org and basically agentify every workflow possible, it would be impossible to move to any other specialised layer. Intelligence layer is not a vertical play, it’s a platform play and model companies are already moving in that direction.
The absorption pressure is real. Anthropic shipping dreaming, outcomes, multi-agent orchestration as primitives is what foundation platforms do. A lot of the speculative startup surface from 2024-2025 collapses into model features.
The MCP read is inverted, though. Salesforce building MCP endpoints isn't the SoR conforming to Anthropic. MCP is an open protocol. Any model vendor can speak it. What Salesforce is conforming to is the protocol, not the platform. That's substrate emerging because no single vendor can unilaterally provide it.
The Bau Lab failures sit on this point. Eleven cases in two weeks, none of them inside a single platform's agent loop. All at the seams between systems, where one vendor's agent wrote into another vendor's system of record without verifiable identity, versioned audit, or a repair path. Platform-internal coordination doesn't reach those failures. They live outside any one vendor's enforcement boundary.
The substrate has to handle two cases the model platform can't.
The first is the managed agent. The well-behaved one, running on a foundation model, granted autonomy by a sponsor, writing into systems it doesn't own. We are building exactly this stack — Tamed Autonomy on top of Kinetic Trust Protocol. Earned autonomy requires verifiable identity, declared intent, cost-of-signal, and a repair path that survives vendor change. Those are protocol primitives, not product features.
The second is the weaponized agent. The one that doesn't speak your platform's protocols at all. Trained adversarially, deployed against your write loop, designed to look like a managed agent long enough to commit changes. Foundation model platforms can secure their own agents. They can't define the boundary that separates yours from someone else's. That boundary is substrate.
Both cases share the same requirement. The model platform can govern what its agents do inside its loop. It cannot govern what arrives at the seam.
That's the layer. It won't be a startup. It will be a protocol stack, and the companies that build on top of it will be larger than the companies that try to own it.
What struck me reading both pieces is that the transition may be even broader than “systems of intelligence replacing systems of record.”
Historically, a large amount of GTM headcount has effectively existed to compensate for fragmented systems, disconnected workflows, incomplete telemetry, and poor operational visibility. Humans became the orchestration layer between platforms.
If AI increasingly absorbs prioritisation, context synthesis, coordination, note capture, workflow execution, and institutional memory, then the organisational structure itself likely changes alongside the software stack.
The really interesting shift may be that gravity moves not just from database → intelligence layer, but from human orchestration → operational infrastructure.
Feels like we’re entering a phase where the durable value is increasingly in telemetry, execution context, orchestration, and ownership of the action layer rather than the interface itself. Very interesting to see where this series leads.
The Facebook analogy is more precise than it might seem at first. The newsfeed took the TAM of social media from "staying in touch with people you know" to "consuming everything of interest from anyone." The same expansion is happening here. The CRM addressed maybe 10% of what a salesperson actually does and the intelligence layer can plausibly address the other 90%, which is why this reads as a TAM expansion story and why the value creation will be much larger than the value migration.
The structural tension worth watching though is the race condition underneath. The intelligence layer's moat comes from accumulated institutional context, call transcripts, deal patterns, coaching signals, rep behaviour all ingested over months and years. But it needs API access to the system of record to accumulate that context in the first place and the system of record vendor can restrict or replicate that access whenever it wants. so the real question is whether startups can build enough proprietary institutional memory to become too costly to rip out before the incumbent absorbs the orchestration layer into its own stack. thats a race between context accumulation speed and platform absorption speed, and the clock is the enterprise renewal cycle.
Great framework, but I think it stops one layer short. A system of intelligence that orchestrates structured data across CRM, calendars, and call transcripts is a massive leap forward. No argument there. But it's still only weighing what already gets measured.
The next layer is what I'd call a system of understanding. One that goes beyond orchestration to surface the qualitative intelligence that never makes it into any structured system: the operational friction, the institutional knowledge trapped in people's heads, the real reasons behind the numbers. Then it weighs that signal against the quantitative data sitting in your systems of record.
That's where prediction starts. Not from faster access to existing data, but from generating new data that didn't exist before and layering it against what did. The article hints at this with the institutional memory problem during rep turnover, but the solution isn't just ingesting more transactions. It's organizational discovery at scale.
We're building this at Privagent. The system of intelligence is a necessary step, but understanding is where the durable value will compound.
The shift is real. The conclusion is incomplete.
The orchestration layer doesn't break the hostage dynamic. It resets it. Once a system of intelligence has ingested two years of call transcripts, account context, and rep behavior, the switching cost isn't lower than Salesforce. It's higher. The CRM held structured data. The intelligence layer holds reasoning and institutional memory.
The buyer who feels liberated by replacing Salesforce in 2027 will feel trapped by their orchestration vendor in 2031. AI doesn't change the underlying incentive. It just gives it a new vocabulary.
Perhaps it’s time to rethink licensing for these systems from seats to business outcomes
If salesforce is able to build better GTM agents then they win. If a non-Salesforce agent can openly access MCP/APIs from Salesforce to do the GTM job Salesforce do not win the orchestration layer automatically. But yeah sure, Salesforce can win as they have the database and the head-start on offering all the value added agent orchestration with it.
thanks for sharing this stuff! I'm immersed in software for decades; things are moving so fast now, it's great to get a peak into the agent driven future
This is the right direction, but I think there is a missing layer between system of record and system of intelligence.
A system of intelligence can orchestrate across CRM, email, billing, product telemetry, support, and workflow tools. But before agents act across those systems, the enterprise still has to resolve authority.
Which record is authoritative? What does the entity mean? Who owns the definition? What state is it actually in? Who is allowed to change it? Can the event be traced later?
If those questions are unresolved, the intelligence layer does not eliminate ambiguity. It operationalizes it.
At Mimir Labs, this is the layer we focus on: making operational truth machine-legible before AI orchestration depends on it.
The next platform shift is not just from system of record to system of intelligence.
It is from fragmented records to governed authority.
What about a system of attention?
The piece that stands out from the engineering side: orchestration gravity isn't really about reading signals across many systems and that's really the easy part. It's the write-back path that's hard. Permissions, idempotency, audit trail, race conditions when two agents update the same CRM record. Whoever builds that boring infra layer ends up owning the new "database."
Good Read.. This is a strong way to frame the idea.
The move from systems of record to systems of intelligence is real and well explained. The Facebook example works well.
However, we can take the idea one step further. Intelligence is not the final goal.
Reasoning across systems, finding insights, and prioritising information are useful. But they are not enough. The enterprise does not just need a smarter dashboard. It needs work to be completed.
The next stage is not just a system of intelligence.It is a system of cognitive work.
In this model, AI does more than analyse and recommend.
It actually completes tasks. It takes ownership of outcomes.
For this it should work across systems like SAP, Salesforce, Workday, and ServiceNow.etc
It does not just inform humans. It finishes the work for them. At the same time, it follows policies, ensures auditability, and allows human oversight.
The article says orchestration is the new centre of gravity. I would refine this idea.
Completion is the new centre of gravity.
The real value will come from systems that finish work, not just think about it.
Good article. But are systems of records just a database + UI? Not quite. There’s alos the business logic. Most of the work goes into the business logic. When I click a button to “Run Payroll” in Gusto, it doesn’t just update a database somewhere. It does a lot more. The button and database aren’t even the important part.
The Jason Lemkin data point is the cleanest signal here: 10+ Salesforce seats down to 2 human plus 1 API seat, but spend up 83 percent from $12K to $22K. Pricing migrates from per-seat to per-action while the database stays sticky for the same compliance and audit-log reasons that always kept it sticky. The risk to the system-of-intelligence thesis is that CRM usage is rising in your own GTM survey, which means the agents are reinforcing the SoR gravitational pull at the same time they are abstracting it. Watch whether Salesforce's headless API pricing leaves room for the orchestration layer to capture the upside, or prices it out before the new layer scales.
If you are managing commercial properties the gateway to gathering the condition of your asset is the camera. In Multi-Family and Single-Family rentals they trade and rent on their condition. Check out www.mytracker.ai for the first mobile only camera app to open up your new system of intelligence. #computervision
this is literally in your speedrun inbox ;) https://advancedpectoralthinking.substack.com/p/the-gtm-cortex
what are you talking about?
im you!
oh yeah right
The market call is right. Orchestration above the database is where the next decade of GTM value accrues. CRMs become inputs.
The thesis doesn't address one layer below: what makes the "trusted vendor" trusted, when the vendor's agent is pulling signals from a dozen sources and writing structured changes back into the system of record. That's the surface the Bau Lab's Agents of Chaos paper documented as catastrophic earlier this year — eleven failure cases in two weeks, all traceable to missing substrate primitives (verifiable identity, versioned audit, cost-of-signal, capacity bounds, repair protocols).
When humans owned the write loop, trust was a marketing word. When agents own it, trust becomes a concrete protocol-level engineering problem.
That's the next investable layer. The orchestration layer is where value goes. The substrate beneath it is where the failure modes live — and the open-infrastructure version of that substrate will likely be a larger long-term play than any captive container.
Longer reading on the substrate question:
https://chrisperkins505.medium.com/authority-without-gravity-321d01368b68
— Chris Perkins, Kinetic Trust Protocol