Precision Belongs in the Substrate
Four voices in the enterprise software discourse, arguing for incompatible cures, have converged this spring on the same diagnosis: the systems we run our businesses on do not carry the rules they pretend to. A platform CEO calls for more infrastructure. An analyst pushes back that the problem is trust. A consulting school revives the older claim that precision is a methodology, not a feature. The frontier AI labs ship product that proposes the agent itself as a substitute for substrate. Each is correct about something. None of them say the thing the field has been avoiding for thirty years, which is that precision is a property of the substrate, and a substrate that does not carry it cannot be rescued by the layers built on top of it, the practitioners hired to compensate for it, or the model deployed to synthesize around it.
The convergent moment
In April, Boomi’s CEO Steve Lucas told the assembled analysts at Boomi World that the value of enterprise software is migrating from applications down into infrastructure. Apps, he argued, are becoming less valuable; the substrate that carries data between them, governs it, and feeds it to AI agents is becoming more valuable. Companies need what he called an “active data foundation, not a data lake alone” — multi-layered infrastructure that turns enterprise data into something agents can actually act on. He cited the volume Boomi processes (hundreds of millions of business processes for thirty thousand companies) as evidence the foundation is real and that his platform is positioned to be it.
Jon Reed at diginomica took the claim seriously and complicated it [1]. He agreed that AI agents diminish the value of conventional SaaS. He pushed back on the implication that all software is becoming commodity, and on the assumption that more capable infrastructure produces more capable enterprises. Agents, he noted, are more brittle than humans. They depend more sharply on accurate context. Speed without trust is a faster way to be wrong. “Change only happens at the speed of trust” was the line he reached for, and it is the right line.
In the same window, on a different surface of the discourse, the precision-methodology school has been promoting a sharpened version of the older critique. Most ERP environments do not fail because the software is defective; they fail because the enterprise was not modeled with sufficient precision before the software was configured. Master data is a strategic asset, not an administrative detail. Weak classifications produce weak reports. Weak hierarchies produce weak roll-ups. Governance must be embedded in the architecture, not bolted on through policy. The cure is methodology — credentialed practitioners, structured frameworks, disciplined configuration — and the buyer’s job is to fund that discipline before the software is allowed to touch the operation.
A fourth voice has been adding itself to the discourse through product moves rather than through essays. The frontier AI labs — OpenAI, Anthropic, and the cohort behind them — are not trying to fix the substrate or train its custodians. Their bet is that a sufficiently capable layer of intelligence sitting above whatever messy substrate exists can produce useful, agent-actionable outputs without the substrate having to be cleaned up first. Anthropic’s Model Context Protocol is the standardized adapter, designed to let an agent reach into any data source without bespoke integration. OpenAI’s Forward Deployed Engineer model, paired with its acquisition of Rockset, is the deployed version: drop a small team and a real-time retrieval substrate into the customer’s environment and let the agent reason across whatever documents, spreadsheets, and legacy tables happen to be there. The implicit pitch — rarely stated this directly, but visible in the product shape — is that you do not need a deterministic database underneath the agent if the agent is smart enough. The substrate’s incoherence becomes a problem the model is asked to absorb, in real time, on every query, forever.
Four sermons. Four different cures. Read them next to each other and the underlying diagnosis is the same in every case: the system, as installed, does not know what it is supposed to be carrying. The schema does not enforce the rules. The vocabulary is not shared. The constraints live somewhere outside the system, in middleware queues or practitioner heads or stewardship policies or model context windows that nobody reads. Each of these voices is, in its own register, telling you the same thing about your enterprise software: you bought a frame that does not carry semantics, and someone is going to have to add the semantics back, somewhere.
The disagreement is not about the diagnosis. The disagreement is about who pays.
The three prevailing answers
The infrastructure school, of which Boomi is one articulate case and which includes most of the data-fabric, data-mesh, data-activation, and modern-iPaaS category, answers the question by selling another layer. If your applications do not agree about what a customer is, the integration platform will reconcile them in motion. If your data lake cannot be trusted by an agent, the active data foundation will enrich it on the way through. If your governance is incoherent, the platform’s policy engine will impose coherence at the boundary. The bet is that the substrate cannot be fixed, but it can be wrapped, and the wrapping is where the next decade of value will accrue.
The methodology school answers the question by selling discipline. The substrate is what it is; the cure is to model the enterprise precisely before the software is configured, hire practitioners credentialed in that modeling discipline, treat the configuration as engineering rather than administration, and produce, through the practitioners’ care, the precision the substrate cannot supply on its own. The bet is that humans, properly trained, can carry the semantics the software does not, and the visible artifact of their care — the configured system — will hold long enough to be useful.
The agent-as-substitute school answers the question by selling intelligence. The substrate is what it is; the cure is to drop a sufficiently capable model on top of it, equipped with a standardized adapter (Anthropic’s Model Context Protocol is the canonical example) and supported by deployed engineering (OpenAI’s Forward Deployed Engineer model paired with the real-time retrieval substrate it acquired with Rockset), and let the model synthesize whatever the operator needs to ask of the data. The bet is that the model can carry the semantics the substrate does not, that the customer’s experience of querying clean data is itself the deliverable, and that the substrate’s underlying incoherence becomes invisible behind the agent’s synthesis fast enough that nobody has reason to ask whether the substrate has been fixed.
All three answers preserve the assumption that the substrate is not the place where precision lives. The infrastructure school relocates precision into a wrap layer; the methodology school relocates it into the configurer; the agent-as-substitute school relocates it into the model’s continuous synthesis. None of the three asks whether the schema itself, the constraint engine, the enforced vocabulary, the operational state machines, could be the place where precision lives — durably, declaratively, in a form that survives the departure of any one consultant, the replacement of any one middleware vendor, and the deprecation of any one model release.
Why all three answers fail
The infrastructure school fails because translation is not agreement. A platform that reconciles three systems’ definitions of customer in flight is not producing a definition of customer; it is producing a series of pairwise mappings that will diverge again the moment any one of the source systems changes. The wrap layer can move the inconsistency, smooth it, present it more legibly to whatever sits above it, but the inconsistency does not stop existing. It is held in suspension by the wrap layer’s continuous attention. Withdraw the attention — defund the platform, lose the integration team, accept the pricing pressure of the next vendor consolidation — and the inconsistency resumes immediately, having never actually been resolved.
When agents become consumers of the wrap layer’s output, the failure mode worsens, for the reason Reed identified. A human reading a reconciled report has some chance of noticing that the customer count differs from the customer count in the source system; agents do not notice. They take the wrap layer’s reconciliation as ground truth, reason confidently from it, produce fluent and incorrect actions, and surface no warning the way a thoughtful human reader would. The faster the loop, the more the brittleness compounds. The infrastructure school’s product is most useful precisely in the regime where its failure mode is least observable.
The methodology school fails because practitioner precision is evaporative. The configured system that the disciplined practitioner produced on day one is a snapshot of their understanding on day one. The operation evolves; the schema does not. The configuration drifts because every operational change is a small renegotiation of the model, and the model lives outside the system, in the practitioner’s head and the project documentation that the practitioner wrote and that nobody on the customer’s side ever reads after go-live. Three years in, the original consultant has moved to another engagement. Five years in, nobody at the customer can tell you why the chart of accounts is structured the way it is, what the W3 work-order substate was meant to capture, why the part-number convention has three exceptions, why the cost-center hierarchy contains a level that nobody uses. The precision was in the practitioner. The practitioner is gone. What remains is the residue of their care, which has now become a slowly rotting artifact that the customer is paying to maintain without understanding.
This is not a failure of the methodology school’s principles. The principles are correct. Master data is a strategic asset. Weak classifications do produce weak reports. Governance does need to be embedded somewhere structural. The school’s failure is in where it embeds them. By placing the embedding in the human practitioner’s discipline rather than in the schema’s enforcement, it produces an architecture that requires continuous, expensive, unnamed labor to remain coherent — and a vendor relationship in which the absence of that labor is itself a billable line item.
The agent-as-substitute school fails because synthesis is not enforcement, and an agent has no operational state. The pitch — that a sufficiently capable model can read across whatever messy data exists and produce useful answers without the data having to be cleaned up — confuses the question of what was the answer with the question of what the system is permitted to record. A constraint engine refuses operations that violate the rules the business has committed to, before they enter the data. An agent synthesizing across the result has no upstream presence. It cannot prevent a malformed transaction from being written; it can only describe, after the fact, what was written. When two source systems disagree about what a customer is, the agent does not produce a definition of customer either; it produces a real-time guess at which definition was meant in this query, and the guess will differ at 9:01 from 9:02 if the context window shifts, the model is updated, or any input changes by a row. There is no canonical answer. There is no version-controlled answer. There is no answer an auditor can reproduce on a different day.
The brittleness Reed identified for agents reading off poor substrate worsens, not disappears, when the agent is asked to be the substrate. A faster agent without a substrate is a more efficient producer of confident, irreproducible answers, and the customer relationship that depends on it is a continuous-attention relationship in the same shape as the integration platform’s, with the same withdrawal risk: stop paying for the agent and the synthesized coherence stops the moment the agent stops. The substrate beneath has not changed during the entire engagement. What was synthesized was never persisted. The amplification problem the essay has already named is now terminal — there is no clean thing for the agent to amplify, only the agent’s own ongoing performance of clean-thing-ness, billed by the token.
All three schools have a structural reason to want the substrate to remain inert. The infrastructure school sells more layers when the substrate doesn’t carry the rules. The methodology school sells more engagements when the substrate doesn’t carry the rules. The agent-as-substitute school sells more compute and more deployment hours when the substrate doesn’t carry the rules. The buyer who reads all three pitches together notices the agreement: nobody at the table is recommending that the substrate carry the rules. That recommendation, if it were to come, would shrink all three businesses.
The substrate path
The substrate path is the one none of the three schools is incentivized to argue: that precision can be a property of the substrate itself. That the schema can be the contract. That state machines can belong to operational entities, with the legal transitions enforced by the system rather than encoded in a workflow appendix that the next consultant will rewrite. That the semantic vocabulary — what a customer is, what an order is, what a part is, what a work order is permitted to do — can be shared across modules and across tools, declared once, enforced everywhere, rather than redefined per integration. That a constraint engine can sit at the boundary, refusing operations that violate the rules the business has actually committed to, before they enter the data and have to be reconciled out of it later.
This is harder to build than any of the other three cures. You cannot bolt a constraint engine onto a schema that was designed without one; the existing data violates constraints that did not exist when it was written, and the act of imposing them surfaces a backlog of corrections that no customer wants to fund. You cannot impose a shared semantic vocabulary on a product that has shipped twenty years of integrations against its old vocabulary; partner ecosystems break, customers’ bespoke configurations break, the upgrade path becomes treacherous. You cannot make the schema carry operational state machines without redesigning the schema, which is the part of the system every vendor has the strongest reason to leave alone.
So the substrate path is rare. The handful of products that have actually attempted it are not the ones with the largest marketing budgets. The three schools that argue for the alternatives have, between them, most of the discourse — the analyst relationships, the conference circuit, the LinkedIn thought-leadership layer, the certification programs, and now the model deployment teams. The substrate path’s argument has to be made from outside the institutions that are funded by the alternatives, which is why it is so seldom made well.
Why agents make the choice urgent
Reed is right that agents amplify whatever substrate they sit on. He is also right that change happens at the speed of trust. The two observations point at the same conclusion, and the conclusion is structural: the trust has to live in the substrate, because no other layer scales.
A human operator can carry trust on behalf of a brittle system. The operator looks at the screen, recognizes that the displayed value is wrong, knows that the work order is held for engineering review even though the system says in process, and quietly reconciles the gap in their own behavior. They route around the system’s incoherence with their own. The system survives because the operator is patient with it. The reporting layer is wrong by a known and roughly bounded margin; the operations continue.
An agent cannot do this. The agent has only what the substrate exposes. If the substrate says in process, the agent acts on in process. There is no second channel — the supervisor’s whiteboard, the line lead’s Slack thread, the planner’s spreadsheet — through which the agent reconstructs the operational truth. The agent’s behavior is bounded above by the substrate’s accuracy. A faster agent on a less accurate substrate is a more efficient producer of confident wrong actions, and the marketing is going to celebrate the speed and not the accuracy because the speed is the part that demos.
This is what Reed means when he says agents are more brittle than humans, and it is what none of the three schools can answer — not the infrastructure school with more middleware, not the methodology school with more practitioners, not the agent-as-substitute school by replacing the substrate with the model. The agent regime has, as a strict architectural requirement, a substrate that does not require human patience to be useful. That substrate cannot be wrapped into existence. It cannot be configured into existence. It cannot be synthesized into existence. It has to be designed.
The durability question
The choice the field has avoided is whether enterprise precision should be durable or evaporative.
Durable precision lives in the system. It survives the departure of the consultant, the replacement of the middleware vendor, the rotation of the operations team, the absorption of the company into its acquirer’s stack. It is expressed in the schema, enforced by the constraint engine, encoded in the operational state machines, defended by the shared semantic vocabulary. It is harder to build because the system has to mean what it says. It is cheaper to own because nobody has to keep paying for the meaning to be reapplied.
Evaporative precision lives in people and in services. It is the configured system after the consultants left. It is the integration platform that holds three systems in suspension as long as the contract is paid. It is the data steward who knows what the seventeen flavors of customer in the warehouse actually mean and is going to retire in six years. It is easier to sell because every renewal is another opportunity to sell it again. It is more expensive to own because the bill never stops, and the substrate is no closer to carrying the meaning at year ten than it was at year one.
The vendor, consultant, and frontier-lab tiers all have a structural preference for evaporative precision, because evaporative precision is a recurring revenue line and durable precision is a one-time investment in something that subsequently asks for less of their help. The buyer’s interest runs the other way, but the buyer is rarely positioned to recognize the choice as a choice. The pitches arrive separately. The infrastructure pitch, the methodology pitch, and the agent-as-substitute pitch are presented as alternatives to each other, not as alternatives to the option that none of the three pitchers has any reason to mention. The buyer picks one of the three and pays for as long as the system runs.
Why the law cares about the answer
The durable-vs-evaporative distinction sounds like an aesthetic preference until you remember that the operations of a real enterprise produce numbers the law has opinions about.
Public companies sign 10-K filings whose numbers are auditable to the underlying transactions. Manufacturers carrying ISO 9001 or AS9100 demonstrate lot traceability through quality records that survive after the consultant leaves. Defense suppliers under CMMC produce immutable evidence of every change to a controlled record. FDA-regulated manufacturers reproduce, on demand, the genealogy of any product from raw material to shipment. Software vendors providing financial systems answer to ASC 606’s specific construction of revenue, with auditable timing for every recognition event. The list runs longer than this paragraph.
The honest description of how enterprises actually satisfy these regimes today is not that the substrate is perfect — it usually isn’t — but that a layer of human-traceable work bridges the gap between imperfect substrate and audit-acceptable practice. Auditors are realists. They know the chart of accounts in the QuickBooks file is messy. They accept compensating controls — documented manual reviews, period-end reconciliations, sub-ledger reconciliations, journal-entry approval workflows — that demonstrate the gap was caught and managed even when the system itself didn’t catch it. They accept human attestations — the CFO’s SOX 302/404 certification, the quality manager’s internal audit signoff, the reviewer’s sign-off on the manual reconciliation — because those attestations carry personal false-statement liability that makes them load-bearing. They accept lot-traceability via paper travelers and spreadsheets, because the spreadsheet, once saved, is a frozen artifact another person can reproduce. The bridges share three properties that make them auditable: they are deterministic when frozen, reproducible by another person, and backed by signed personal liability. The substrate does not have to be pure. The bridges have to hold.
The agent-as-substitute pitch is dangerous to compliance not because the substrate underneath is imperfect — it was already imperfect, and the bridges were carrying the load — but because the pitch dissolves the bridges without replacing them with anything the regimes recognize. The agent does not replace the substrate; it replaces the human-traceable work that was making the imperfect substrate compliant in the first place. When the CFO certifies AR today, they are signing based on their own review or their staff’s documented review; the reasoning chain the auditor will follow at audit time is human-traceable, the manual reconciliation can be re-run, and the personal certification carries weight because the executive can defend the work. When the agent does the review and the CFO signs, the reasoning chain is not reproducible by another person at audit time, the artifact the auditor would re-run is not frozen — the same query later produces a different answer — and the personal certification is hollow because the executive cannot defend reasoning they did not perform. Spreadsheet reconciliations are crude and ugly, but they are crude and ugly in a way auditors have decades of precedent for accepting. Agent synthesis is fluent and fast, but it is fluent and fast in a way auditors have zero precedent for accepting, and the absence of precedent is not a temporary condition the industry will fix; it follows from the artifact never being frozen.
The natural objection — and the one any executive will raise within thirty seconds of hearing the argument — is that the agent’s actions can be attributed to the user account that prompted them, which appears to satisfy the WHO half of the audit requirement. It does not, and this is the place where the regimes are sharper than the casual reading suggests. Audit trails require WHAT, not just WHO. Tracing to a user account tells the auditor who issued the prompt; it does not tell them what the system did, whether the system was validated to do it, or whether the same input would produce the same output a year later. SOX-equivalent regimes require evidence of the operation of internal controls, not just the existence of a transaction. ISO 9001 and AS9100 require validation that the procedure followed was the documented procedure. FDA’s 21 CFR Part 11 requires the system itself to be validated for its intended use. None of these accepts user-attribution as a substitute for the system’s own record of what it did. And the obvious-looking solution makes the executive’s exposure worse, not better: when every synthesized action is attributed to the logged-in user, the personal liability runs to that account. The audit log that was supposed to protect the user has named them.
This is the dimension along which the agent-as-substitute pitch will fail first, not last, because the failure mode is structural and the regulators are patient. The first agent-synthesized number that ends up in a 10-K and the auditor refuses to sign — the first agent-inferred lot trace that fails an FDA inspection, the first agent-reconciled deferred revenue that ASC 606’s evidence requirements reject — are not edge-case events careful customers will avoid. They are the consequences the pitch writes into its first deployment. The customer who bought the agent because they did not want to fix the substrate has not even tried to fix it; they have removed the bridges that were keeping the imperfect substrate compliant in the first place, with interest accruing in the form of regulatory exposure that is invisible until it is suddenly the only thing that matters.
The buyer’s test that follows is, for these regimes, the test the law is going to apply on the customer’s behalf whether the customer asked for it or not.
What the buyer can ask
The plain test is one the buyer can apply to any product they are evaluating, and to any system they have already installed, and the test is unkind enough to be useful.
Suppose you fired the consultants tomorrow. Suppose you cancelled the integration platform. Suppose the practitioner who configured the system retired next quarter and was not replaced. What does the system, on its own, know? Does it know what a customer is? Does it know what a work order is allowed to do? Does it know which transitions are legal and which are not? Does it carry a vocabulary that the modules agree on without external translation? Does it refuse, at the schema level, to record an operation it is not permitted to record? Or does it accept whatever it is told, defer enforcement to a layer that will no longer be present, and resume — quietly, immediately — being the inert frame it was when you bought it?
If the answer is the latter, the precision in your system is evaporative. You did not buy precision. You rented it, and the rent will keep coming due as long as the system runs.
The argument the discourse is not yet making, and the one this essay exists to make, is that this rent is avoidable. A substrate that carries its own rules is buildable. The fact that the field’s loudest voices are not building it does not mean it cannot be built. It means the buyers who want it have to ask for it specifically, in the language of the substrate path, because the three schools that own the conversation are not going to surface the option on their own.
The vocabulary the infrastructure school has been validating — active data foundation, agent-ready substrate, trustworthy context — is the right vocabulary for the question, even if the answers being offered against it are still the old answers. The vocabulary the methodology school has been validating — precision is strategic, weak classifications produce weak reports, managers only manage from systems they trust — is the right vocabulary for the diagnosis, even if the cure being prescribed still locates precision in the practitioner. Take the vocabulary. Refuse the cures. Ask whether the system, on its own, when nobody is paid to mind it, knows what it is supposed to be carrying.
If it does, you have durable precision. If it does not, the precision is evaporating right now, and the next layer you are being sold is a meter on the rate of evaporation, not a fix.
Jon Reed, “AI infrastructure layer now more valuable than software? Boomi CEO Steve Lucas raises the stakes at Boomi World,” diginomica, April 2026. Reed’s framing of trust and brittleness — “change only happens at the speed of trust” — is the editorial counterweight to Lucas’s infrastructure-as-asset claim, and the line this essay extends. ↩︎