In The Great Compression, we documented how six model providers deployed over $200 billion to systematically absorb every middleware function standing between foundation models and enterprise workloads. In subsequent dispatches, we showed that compression logic extending into the services layer — through PE joint ventures with Anthropic and OpenAI — and then into the stateful runtime environment itself. That series asked what was being absorbed and by whom. This series asks what the enterprise inside the PE portfolio actually receives, and what precondition that delivery assumes is already in place.
On April 15, 2026, Thoma Bravo and Google Cloud announced a strategic partnership to accelerate AI transformation across Thoma Bravo’s enterprise software portfolio — companies worth more than $300 billion combined. The terms are specific: streamlined access to Gemini models and the Gemini Enterprise platform for agentic AI, forward-deployed Google engineers to solve deep technical challenges, and go-to-market channels through Google Cloud’s Marketplace and co-sell programs. Google Cloud’s chief product officer described the goal as having portfolio companies “deeply embed Google’s leading AI models and Agent Platform into the core of their product stacks.”
That is a precise description of what the announcement delivers. It is also an incomplete description of what agentic AI deployment requires. The announcement is not unusual in this respect — it reflects the same architecture appearing across every major PE/provider partnership now in progress. What Thoma Bravo’s portfolio companies are receiving is a governance and execution capability. What they are being given by name and by press release is a complete agentic AI solution. The gap between those two descriptions has a structure. Understanding it starts with naming what the capability actually is.
Naming What Was Just Delivered
The governance architecture that platforms like Google’s Gemini Enterprise, Salesforce’s Agentforce 360, and ServiceNow’s AI agent platform deliver has a specific architectural identity. It is the set of controls that define which agent behaviors follow explicit logic versus which reason freely — and govern what the agent does within those boundaries. On April 15, 2026 — the same day as the Thoma Bravo/Google Cloud announcement — Salesforce named it directly: the Agent Fabric release describes the capability as “guided determinism,” defining it as the ability to “define fixed handoff rules while LLMs handle the reasoning in between.” Two major signals from the same day. One PE firm embedding the capability at $300B portfolio scale. One provider formalizing it as a named product feature expanding across platforms. The term is Salesforce’s. The architectural analysis of what it delivers — and what it assumes — is the subject of this post.
Guided Determinism
Salesforce’s own definition: the capability to “define fixed handoff rules while LLMs handle the reasoning in between” — applied across Agentforce and now Agent Broker. The design decision that separates deterministic workflow paths from LLM-driven reasoning paths within a single agentic system.
What the term does not resolve: what the agent receives before the deterministic rules execute. Guided Determinism governs what an agent does with what it has received. It has no architectural reach over whether that context was complete, current, or fit for the decisions the agent must now make. That is the substrate condition. This post is about the gap between the two.
Guided Determinism is a real architectural contribution. The controls it encompasses — guardrails that restrict topic and action domains, trust layers that enforce permissioning, escalation paths that route to human review, topic restrictions that bound agent authority, audit trails that log reasoning steps — are all necessary components of any production-grade agentic deployment. This post does not argue otherwise. A portfolio company that deploys agentic AI without a Guided Determinism layer is taking on avoidable governance risk. The JV programs delivering these capabilities are providing something real and necessary.
The architectural question is not whether Guided Determinism is valuable. It is where Guided Determinism ends — and what it assumes is already in place on the other side of that boundary.
Where the Governance Layer Operates
Guided Determinism operates after context reception. By the time the deterministic rules run, the agent has already reasoned over whatever the substrate gave it. The governance layer inspects and constrains the agent’s actions, routes escalations, enforces permissions, and logs decisions. None of those functions reach the upstream condition that made the context available in the first place: whether the data the agent reasoned from was decision-grade.
What Guided Determinism Governs
Operates downstream of context reception. Governs agent behavior after reasoning over the substrate has begun.
- Which actions the agent is permitted to take
- Which topics and domains are in or out of scope
- When to escalate to human review
- What gets logged for audit and accountability
- How trust rules and permissions are enforced
What Substrate Fitness Governs
Operates upstream of the governance layer. Determines whether what the agent receives is decision-grade before reasoning begins.
- Whether retrieved data reflects current operational state
- Whether context is logically consistent across retrieval points
- Whether data structures support machine reasoning, not just human navigation
- Whether operational state is first-class and transactionally bound
- Whether retrieved context is decision-bound at consumption time
The distinction is not subtle once it is named. An agent operating under perfect Guided Determinism — correct guardrails, full audit trails, tight permission scoping, rigorous escalation paths — can still reason from a contextually incomplete picture of the world. If the data it received was stale, logically inconsistent, missing the operational state that would make a recommended action binding, or structured for how a human analyst navigates a report rather than how a machine reasons to a decision, the governance layer has no view of that condition. It governs the action the agent takes. It does not govern the picture the agent formed before taking it.
Guided Determinism governs what an agent does with what it receives. It does not resolve what it receives — and SaaS platforms positioned as complete agentic AI solutions are selling the former as if it were the latter.
The Substitution Happening at Scale
This architectural gap would be a contained engineering concern if it were being discussed openly in the deployment programs now running across PE portfolios. It is not. The Thoma Bravo/Google Cloud announcement describes the goal as having portfolio companies “transition into AI-first companies.” Vista Equity’s Agentic AI Factory anticipates deploying between four and eight billion autonomous agents across its portfolio. Salesforce calls Data Cloud “non-negotiable” for the agentic enterprise — then scopes Data Cloud to customer data within the Salesforce CRM boundary. Every one of these programs delivers the governance layer with precision and genuine capability. None of them contains a substrate fitness assessment as a pre-deployment condition.
The substitution that is happening at scale is this: the governance capability is being positioned as the complete answer to a problem whose harder half — whether the data substrate the agents will reason from is fit for agentic use — is assumed away. The assumption is not visible in the JV announcements, the deployment roadmaps, or the platform marketing. It becomes visible in the failure data that accumulates after deployment.
Gartner’s projection for the share of agentic AI projects that will be abandoned by end of 2026 — with the primary failure condition identified as lack of AI-ready data, not model capability or governance tooling. A separate measure shows 91% of organizations acknowledge that a reliable data foundation is essential for AI success, while only 55% believe they actually have one. That 36-point gap is the substrate assumption made visible at scale.
The 36-point gap between what organizations believe is necessary and what they believe they possess is not a technology awareness problem. Enterprise CIOs and architects know data quality matters. The gap persists because the enterprise is receiving a governance capability — deployed by credentialed engineers from credentialed providers — that does not require substrate fitness as a precondition. The deployment can complete successfully. The audit trail is clean. The guardrails are configured. The governance layer is operational. None of that verifies the substrate condition beneath it.
What Should Come Before
The argument this series is building is not that Guided Determinism is the wrong solution. It is that Guided Determinism is the answer to the second question. The first question — is the data substrate this agent will reason from actually fit for agentic use? — should precede provider selection, not follow deployment failure. That question has a structure. It maps to five specific architectural tests, developed across the Substrate or Scaffolding series, that determine whether an enterprise data architecture can serve as a decision-grade agentic substrate or whether the gaps it carries are extensible versus structural.
Some of those gaps can be layered in. Others require the data substrate to become something it was never designed to be. Guided Determinism cannot tell you which condition applies to your portfolio company. That is the assessment that should be running in parallel with — and in some cases before — the governance layer is configured.
The PE/provider partnership programs now being executed at portfolio scale are delivering the governance layer with precision. The substrate precondition those programs assume — that the data architecture beneath the agents is fit for agentic reasoning — is not visible in any program currently being deployed at scale. That gap is the subject of this series.
