Autonomous Business is the destination. Constrained Agency is how you get there.
The destination is real. The implementation manual nobody is publishing — the discipline of constraint engineering — is what gets enterprises from pilot purgatory to autonomous business at scale.
The dominant industry frame for the next decade of enterprise AI is Autonomous Business — the era after digital business, in which AI agents make decisions and act, organizations are organized around expanding autonomy, and the unit of value shifts from human-supported software to software with its own agency. McKinsey, BCG, Bain, Deloitte, KPMG, Accenture, MIT Sloan, a16z — every house has converged on the destination. CEO surveys are organized around it. Investment decks lead with it. The frame is right.
The frame is also incomplete. It describes the destination clearly and the path to the destination almost not at all. Operators following the conventional roadmap — pilot, expand, scale, eventually autonomous — are landing in pilot purgatory at a remarkable rate, and the failure mode is not what the analyst pieces predict. They are not failing because the technology is too immature, the data is too messy, or the change management is too slow. They are failing because they have skipped the discipline that the destination actually requires, and that discipline does not yet have a clean name in the market. This piece names it: Constrained Agency. Autonomous Business is where the decade is going. Constrained Agency is the operating discipline that gets you there. The operators who master the discipline lead the destination. The operators who chase the destination without the discipline keep paying for pilots that don’t scale.
This is the foundation of a insight series under SALT’s name. It establishes the relationship between the destination and the discipline, names the four classes of infrastructure that constrained-agency operators build, addresses the strongest counter, and sets out the operator move — what a Founder, COO, CTO, or Head of Strategy does on Monday to put the discipline to work.
Destination versus discipline.
Read against the public record of agentic deployments — Itaú Unibanco’s autonomous code agents at scale, NEC’s machine-buyer rollout in procurement, the case examples across asset-intensive manufacturing, the multi-agent systems being deployed against healthcare workforce shortages, the agentic-insurance hybrid models — every successful deployment has the same structural fingerprint.
Itaú made it work by inserting a named human owner per agent session and routing every agent action through the same code-review gates as for human-written code. NEC made it work by ruthlessly narrowing each agent’s scope to dynamic enough to justify an agent, low-risk enough to delegate, not so integral that employees lose self-worth. The healthcare multi-agent successes converge on a single precondition: multistep, documented, repeatable, frequent. Every win is achieved through bounded delegation, named accountability, and engineered constraint.
The macro narrative of Autonomous Business describes the destination correctly. The case-study record describes the path, and the path looks nothing like the slope the autonomy-expansion narrative implies. The path is not “incrementally remove humans from the loop.” The path is “incrementally engineer the constraints that allow machines to act safely, accountably, and recoverably — and progressively widen the bounded autonomy as the constraint infrastructure matures.”
Six moves.
Every successful agentic deployment in the public record reaches its bounded autonomy through aggressive constraint engineering, not autonomy granted on day one.
This is the empirical observation, and it is worth treating as load-bearing because the persistence of the pattern is what makes it informative. Itaú’s win came from the factory model — high-volume, repetitive, well-bounded tasks under standardized workflows, with named accountability per session. NEC’s win came from a deliberate three-axis filter on what work an agent was allowed to attempt. The healthcare multi-agent wins live in scheduling and intake rather than diagnosis or treatment planning. The manufacturing case examples are dominated by asset-intensive workflows where the agent operates within hard physical and regulatory constraints. There is no public case in this evidence base in which an enterprise granted an agent open-ended autonomy on day one and produced a sustained win. Every successful deployment reached its current state of autonomy by building up through constraint engineering, not by skipping past it.
Constraint engineering is a discipline, not a deficiency.
The most damaging assumption in the current market is that constraint is a temporary feature of the present technology — an artifact of model immaturity, governance immaturity, organizational immaturity, that will dissolve as the technology and the organization mature. This is wrong in the same way that “permissioning is temporary, eventually all systems will be open by default” was wrong about cloud computing. The reason organizations constrain agents is not that the technology cannot yet handle autonomy. The reason is that the value created by an agent is a function of how well the constraint matches the use case. The constraint is not friction to be removed. The constraint is the product. Designing the right constraints — what data the agent can read, what actions it can take, under whose authority, with what expiration, with what fallback when its scope ends — is the work that creates value.
The infrastructure required for Constrained Agency does not yet exist as a coherent stack.
When operators succeed today, they succeed by hand-rolling four classes of infrastructure that should be commodity but aren’t. None of the four is currently a discrete enterprise software category with a clean vendor map. Each is being assembled, today, at every successful deployment, from primitives that were built for something else. The first vendors to ship coherent products in each of the four categories will build category-defining companies. Microsoft has shipped first on Class 01 with Entra Agent ID and Agent 365’s unified registry; the rest of the stack is open.
The market is mispricing the constraint stack because the failure mode is silent until the deployment scales.
A poorly-constrained agent in a bounded factory workload — migrate this service, summarize this ticket, stage this PR — works adequately. The ceiling is high enough that the missing constraint infrastructure does not yet bite. As deployments move into higher-stakes dynamic decision domains — procurement with payment authority, customer service with refund authority, clinical triage, financial reconciliation, contract negotiation — the failure mode flips. A wrong decision in a high-stakes domain is total: the audit chain points to no recoverable principal, the over-scoped service account did exactly what it was permissioned to do, the response time on revocation was hours rather than seconds, and the operating-model design did not anticipate the failure class that occurred. The 18-to-30-month window before the average enterprise’s agentic deployments cross from bounded factory workloads into dynamic decision domains is the window in which the constraint stack goes from “nice to have” to “you cannot ship without it.”
The discipline that wins this decade is not AI engineering. It is constraint engineering.
AI engineering — the work of training, fine-tuning, prompting, and chaining models — is becoming a commodity skill at the speed that any abundant skill becomes a commodity. Constraint engineering — the work of deciding what an agent is allowed to do, under whose authority, with what expiration, in what failure mode, against what governance gate, with what auditability — is barely emerging as a named discipline. By 2028, the operators who can hire constraint engineers (or train them) will be the operators who can ship agentic systems into production. By 2030, constraint engineering will be a recognized practice in the same way site reliability engineering became one in the 2010s — and for the same reason: the discipline emerged from operators who got tired of paying the cost of the absence of the discipline.
The operating-model layer is where the actual transformation budget gets spent — and where vendors are systematically underpricing the work.
The technology layer (foundation models, orchestration platforms, vendor agent SDKs) is well-priced and getting cheaper. The operating-model layer — the human roles, the delegation pathways, the governance gates, the cross-functional decision rights, the talent overlay — is unpriced because most vendors don’t sell it. Roughly three-quarters of an enterprise’s AI transformation budget will be spent below the technology layer, in the operating-model layer. The vendors who tell you Autonomous Business is a technology procurement decision are selling you the 25% and pretending the 75% is your problem. It is your problem. It is also where the actual path to the destination runs.
The strongest argument against this position.
The strongest counter is that Constrained Agency is a description of the present, not a discipline of the future — that the constraint pattern persists today only because the technology is immature, and that as foundation models become more reliable, agent frameworks more robust, and trust higher, autonomy will progressively expand and the constraint work will look quaint.
This counter is correct in the limit and wrong in practice. In the limit — twenty or thirty years out, with significantly more capable AI and significantly more developed governance frameworks — the constraint stack will be more automated, more standardized, and less hand-rolled. But the practical operator question is not the limit; it is the next decade. Inside the next decade, two structural facts dominate. First, regulatory and fiduciary domains have hard ceilings on autonomy that are not technology-relaxable: HIPAA, FDA, financial-services compliance, professional licensure, fiduciary duty. There is no historical evidence that ceilings rise with capability, and substantial historical evidence that they do not — every major regulatory regime has, over time, added requirements rather than removed them as technology has matured. Second, the human-in-the-loop is not a bug to be removed; it is the load-bearing accountability principal that allows the enterprise to deploy agents at all in regulated and high-stakes domains. The Autonomous Business destination still arrives, but it arrives via progressively widening bounded autonomy under engineered constraint — not via the dissolution of constraint.
Three things to do this quarter.
For a Founder, COO, CTO, or Head of Strategy of an enterprise that is currently in agentic AI pilot mode:
01 · Stop asking how much autonomy. Start asking what constraints. Replace “how much autonomy should we give the agent” with “what constraints make bounded autonomy safe at scale here, and how do we widen them as we learn.” The first question routes you into a vendor evaluation; the second routes you into operating-model redesign with a clear path to the Autonomous Business destination. The second is the higher-leverage one and the harder one, which is precisely why it is the work most operators are skipping.
02 · Audit your pilots against the four constraint-stack categories. Authority and identity, durable execution, governance and gating, operating-model design. For each, ask: what is doing this work today, and what will be doing it when we are running ten times the agent volume. Where the answer is “no one” or “we are hand-rolling it,” that is your investment priority for the next twelve months.
03 · Hire — or develop internally — for constraint engineering, not just for AI engineering. The AI engineering market is saturating. The constraint engineering market is empty. The single highest-leverage talent move an operator can make in this decade is to identify, recruit, or develop the people who can do this work, and to give them a seat at the operating-model design table — not bury them in a platform team.
Underneath all three moves is the same redirection: most of the path to Autonomous Business is operating-model redesign, not technology adoption, and the operators who organize around that fact arrive at the destination first. The operators who treat AI as a procurement decision keep landing in pilot purgatory and blaming the technology, when the actual problem is that they were trying to skip the discipline that the destination requires.
SALT’s standing position-review rhythm grades published positions against subsequent reality. Where positions falsify, SALT publishes the correction explicitly. The two predictions below carry this piece’s claim.