Insights from SALT on autonomous business, regulated industries, and the operating-model work most AI vendors don't price. Each one ends in a decision a CEO, COO, CTO, or Head of Strategy can make on Monday.
For CEOs, CFOs, and CTOs setting AI strategy. What enterprise AI actually pays off, where most investments fail, why the Microsoft stack changes the math, and the budget category most vendors don't tell you about.
For Founders, COOs, and plant leaders. What AI actually delivers on the plant floor today, where the "lights-out factory" hype falls apart, and what to buy if your shop is between $100M and $2B.
For health-system CEOs, CIOs, and CMIOs. Where AI pays off in healthcare, where regulation stops it cold, why your contract model matters more than your model choice, and which leadership role makes or breaks the deployment.
For Managing Partners, CEOs, and COOs of services firms. Why the billable hour stops working, how to manage agents like systems instead of staff, the platform decision most firms get wrong, and a three-year blueprint.
Three insights to start with. The foundation thesis, the contrarian management move, and the closing blueprint.
The destination matches the consensus. The implementation manual — and the discipline — is the constraint engineering nobody else is naming. The operator who asks "how much autonomy" goes shopping for platforms. The operator who asks "what constraints make bounded autonomy safe at scale" arrives at the destination.
The "agents as teammates" frame is leading operators into a specific failure mode HBR has already named: anthropomorphizing AI reduces individual accountability, increases unnecessary escalation, and lowers review quality. Treat agents like systems with named delegators, not like junior employees.
A defensible blueprint — workflow surface, agent inventory, auth layer, governance gates, talent overlay, contracting model — operators can mark themselves against. Every prior thread compounds here.
Filter by pillar, by tag, or search directly. The shelf updates as new positions ship and existing positions get graded.
The destination matches the consensus. The implementation manual nobody is publishing is the constraint engineering that gets an operator from pilot purgatory to scale.
Agent identity is separating from the broader non-human-identity category. Microsoft shipped first with Entra Agent ID. The 18-month category-formation window is open.
Vendors systematically underprice the layer where the actual transformation happens. Roughly 75% of an enterprise's AI transformation spend goes below the technology layer.
Structural argument grounded in shipped product. Microsoft is organized around governance primitives; the open stack around developer flexibility. Governance compounds. Flexibility decays.
Standard scenario frameworks treat fragmentation and autonomy as independent axes. They aren't. Fragmentation is causing the autonomy push. Collapse the four scenarios to two trajectories.
The obvious AI use case is now table-stakes. Operators chasing it as a differentiator are buying yesterday's playbook. Where the new differentiation actually lives.
Analyst research is written for $5B+ enterprises. The $100M–$2B operator's deployment pattern is structurally different and currently under-served. The playbook for the band buying Copilot.
Software-defined products and closed-loop digital twins force convergence between product life-cycle and manufacturing-execution stacks. Operators treating them as separate procurements lose.
Contract manufacturers, not in-house plants, will be the first scaled surface for autonomous procurement and quality agents — the regulatory and HR ceilings are lower at the supplier boundary.
Every public lights-out case study has 200–400% maintenance technician attrition or a hidden human layer just off-camera. The constrained-agency factory floor is the scalable pattern.
The enterprise IAM stack ends at the OT firewall. Agentic systems on the plant floor will force the agent-identity category to land here first, before the carpeted-office stack.
The operator question is not "how do we get to autonomy" but "how do we design the highest-value hybrid given a hard regulatory ceiling." The hybrid is the destination, not a transitional state.
AI deployment in provider organizations succeeds or fails on CxIO leadership and reporting structure; the technology decisions are downstream.
AI investments under FFS reimbursement cap out at administrative efficiency; only VBC contract structures convert AI into compounding margin. Contract structure determines ROI.
Certain workflows — death notifications, terminal-diagnosis communication, total-loss conversations — will not autonomize regardless of model capability. Designing around the empathy constraint is the operator skill.
The case-study evidence converges on a single precondition: multistep, documented, repeatable, frequent. Operators chasing MAS without that precondition are buying liability.
The EHR-centric architecture loses to API-driven cores with clinician-built UX. Existing EHR vendors will resist, then capitulate, then acquire.
HIPAA-grade audit and patient-consent delegation push agent-identity requirements past what generic enterprise auth can handle. Healthcare will produce the first specialized agent-auth vendor.
AI is a price-deflation event for time-priced services and a margin-expansion event for outcome-priced services. Firms that don't migrate the contracting model lose, regardless of AI adoption.
Most operators will not develop the in-house capability to run agentic systems at scale. The MSP category absorbs the operating-model work and becomes the durable form of mid-market AI delivery.
PSA selection is treated as a workflow-tooling decision; it is actually the load-bearing infrastructure decision for whether the firm can deliver outcome-based services profitably.
The "agents as teammates" frame leads operators into a specific failure mode HBR has named: anthropomorphizing AI reduces accountability and lowers review quality.
For firms under ~500 people, the build-vs-buy debate is over. The firms that will most regret it are the ones too big for Copilot defaults and too small to build the alternative.
A defensible blueprint — workflow surface, agent inventory, auth layer, governance gates, talent overlay, contracting model — operators can mark themselves against.