Build-to-fit is the new procurement.

THE BET. Enterprise software was built for a world where integration was expensive and code generation was slow. Both assumptions flipped. A tuned blueprint now ships faster than a SaaS trial converts. And it fits the workload a generic product cannot.

THE TELL. The workloads where generic SaaS has always failed — underwriting, revenue cycle, clinical operations, regulatory document review, reinsurance placement — are the first to move. Not because AI is cheaper. Because the fit, finally, is possible.

THE MATH. A blueprint baseline deployed in eight to fourteen weeks, tuned in milestones, priced against a measured pre-engagement baseline. Every phase has defined exit criteria. The risk profile of buy. The fit of build.

01 The procurement decision was a cost decision. Cost inverted.

The old buy-vs-build decision was never a strategy decision. It was a cost decision dressed up as a strategy decision.

Buying won by default for two reasons. Integration was expensive — system integrators charged three to five times the license fee in services, and the bill arrived before the value did. Custom code was slow — twelve to twenty-four month builds with high failure rates, where the market the software was built for had already moved by the time it shipped. Set against those two costs, even a SaaS product that fit the workload poorly was the rational choice. The cheap, narrow, mediocre fit was the cheap, narrow, mediocre fit, and the alternative was worse.

Both assumptions flipped at the same time. Code generation collapsed the build cost. Agent-owned integration surfaces collapsed the integration tax. The blueprint that took twelve months to specify and twelve months to deliver now takes eight to fourteen weeks end-to-end, gated against measured outcomes phase by phase.

8–14 weeks

Blueprint-to-production deployment window

Versus 12–24 months for the prior generation of custom build

The economics that made procurement the safe choice are the economics that no longer hold. Procurement still wins for commodity workflows where the SaaS market has compounded for two decades. Procurement loses everywhere else.

02 The workloads where SaaS never fit are moving first.

The pattern is consistent: where the workload is tuned to a single firm’s data shape, where SMEs are the bottleneck, where every quarter of generic-tooling friction has a measurable cost — that is where build-to-fit lands first.

The shape of this thesis is easiest to see across the Blueprints — the working library of workloads where the same operational geometry shows up at firm after firm in the same industry, and was never a SaaS shape to begin with. Different firms; same workload. Each one packaged as a tuned baseline rather than a custom build from zero.

In financial services and insurance:

• Submission Flow — submission intake, extraction, enrichment, and first-pass risk selection for commercial and specialty carriers. The work that used to be hand-keyed rekeys and PDF wrangling is an agent loop that reads, classifies, enriches, and routes. Broker intake formats vary; underwriting guidelines vary by segment. No generic underwriting product ever modeled either.
• Asset Flow — asset servicers running corporate actions, multi-custodian reconciliation, and client reporting. Custodian formats vary, client reporting templates and voice profiles vary, LPA and side-letter terms vary by fund. The configuration surface of any generic platform was always going to be smaller than the variability of the actual work.

In healthcare and life sciences:

• Revenue Flow — claim-to-cash for providers and diagnostic labs. Eligibility verification, coding assist, denial classification, appeal drafting. Payer mix changes the geometry of every step.

The common geometry across these workloads: variability that cannot be parametrized, SMEs gated behind mechanical work, vendor tools that captured the surface and never the substance. These are not workloads a SaaS roadmap was ever going to absorb. They are the workloads the next decade of enterprise software is being commissioned against.

03 The honest counter-case.

The thesis is real. The counter-case is also real. Four positions deserve the strongest version of themselves before being answered.

Total cost of ownership over three to five years. Custom systems carry maintenance, drift, re-platforming, and on-call cost that SaaS vendors amortize across thousands of customers. Where it holds. Commodity workflows — CRM, payroll, ITSM, identity — where category-leading SaaS has ten times more feature investment than any single customer can justify funding. The TCO math wins for the buyer there, and probably will for another decade. Where it breaks. Workloads tuned to a single firm’s data shape. The maintenance cost of not fitting — SME workarounds, stitching tools, manual reconciliation, slowed throughput — is higher than the maintenance cost of a focused custom build. The TCO comparison is honest only when you count the cost of the misfit, not just the cost of the fit.

The talent gap inside the enterprise. Most enterprises cannot operate ML and agent platforms in-house. Where it holds. This is real, and it is the reason “build” failed for twenty years. The platform team that could keep a custom system alive existed at one in fifty Fortune 500s. Where it breaks. A managed-services partner removes the constraint. The build is yours; the operations are operated. We will return to this in the final chapter — it is the precondition for everything else.

Vendor SLAs and pre-certified compliance. SOC 2, HIPAA, FedRAMP, regional residency — packaged software ships these. Where it holds. Net-new compliance lift is a real cost and a real risk. Saying “we built it ourselves” to a regulator is not a posture that ages well without the paperwork. Where it breaks. Blueprints inherit certified substrates — AWS HealthOmics, Bedrock, certified VPC patterns, certified deployment surfaces from the model providers. The compliance cost is amortized inside the substrate, not paid again per workload.

Best-of-breed feature lap. Salesforce, ServiceNow, Workday, SAP have decades of compounding feature investment. Where it holds. At the system-of-record layer, where the integrations are the product. Where it breaks. In the intelligence layer above the system of record — the layer that reads what the system of record holds, decides what it means, and routes the next action. That layer is what is being commissioned now. The system-of-record vendors will compete in that layer. They will not start with the lead they had in the layer below it.

The deeper treatment of where SaaS still wins is the next essay in the series. The summary is: the categories that fit a generic product still buy a generic product. The categories that never did are commissioning the build instead.

04 Why “build” no longer means what it used to mean.

The “build” of 2008 was a custom application written from scratch, integrated through bespoke ETL, run on infrastructure the team also built. Every part of it was unknown surface. The schedule slipped because every part of it was unknown surface.

The “build” of 2026 has a different shape. The substrate is pre-written. The integration surface is owned by an agent loop, not by a hand-written connector. The schedule is gated phase by phase against a pre-engagement baseline that was measured before any code was written. The pricing is anchored to the improvement, not to the wishlist.

Concretely:

• Pre-written blueprints — industry-tuned baselines for the workloads that recur (underwriting submission triage, audit table extraction, clinical-data abstraction, regulatory observation classification). The next carrier starts where the previous carrier ended.
• Agent-owned integration — the connector that used to be a quarter of the engineering cost is now an artifact the agent maintains. Schemas drift; the agent rebinds.
• Measured exit criteria — every phase has a number it has to clear before the next phase starts. The Sprint runs against the customer’s own data. The Integrate phase runs against the customer’s own KPIs.
• Judgment-only humans — the human work moves from implementation to judgment. The SME reviews the agent. The agent does the typing.

The risk profile of a build, with the unknown surface of a build removed.

05 What comes next: managed fleets of agents built by power business users.

The first wave of agents is being commissioned by the platform team. The second wave will not be.

Power business users — analysts, claims adjusters, pricing actuaries, clinical operations leads, audit seniors, compliance assessors — are already building agents. The frameworks are progressively-disclosed enough that the marginal new agent does not require a software-engineering team. Bedrock Agents, the AI SDK, internal agent platforms, low-code agent surfaces inside the system-of-record vendors. Every one of those reduces the bar to commissioning a new agent. Every one of them moves the commission decision out of the platform team.

The endpoint is straightforward to predict and uncomfortable to plan for. Every Fortune 1000 will end up with hundreds — and within a few years thousands — of agents in production. None of them will be governed by the stack that governed the prior generation of microservices. Most enterprises do not yet have an answer for the operating model that follows.

This is where a new managed-services category emerges. Not a model-training partner. Not a build partner. A fleet operator. The managed-services partner that:

• Orchestrates the agents the business builds — across reasoning models, tool surfaces, and data domains
• Observes the fleet — every agent’s behavior, latency, cost, error surface, drift signal
• Evaluates continuously — against domain-specific quality gates the business defines, not generic LLM benchmarks
• Governs lifecycle — promotion, retirement, prompt and model versioning, rollback
• Holds the security posture — credential scope, data exfiltration, prompt-injection surface, audit trail
• Owns cost — token economics across providers, routing decisions, budget enforcement per agent
• Manages drift — when the data shape changes underneath the agent, when the model provider changes underneath the workload

This is the layer Provectus is positioning into. The work is already underway in the engagements that show the pattern. The cancer-diagnostics agentic document-intelligence pipeline operates a fleet of agents under a HITL cockpit that trains the fleet from human corrections. The UK compliance provider’s agentic platform orchestrates classify, validate, summarize, and route across regulatory document workflows that still expand in domain coverage. The pieces of fleet operations are visible in both — orchestration, observability, evaluation, lifecycle, governance — even when the fleet is still small. The pattern scales.

The build-to-fit decision is upstream of the agent-fleet decision. You make the first one when you commission a workload. You make the second one when you realize you have forty of them and no one is operating the estate.

The thesis stands on its own: build-to-fit is the new procurement, because the cost economics that made procurement the default are the cost economics that no longer hold. The honest counter-case is that the categories that fit a generic product still buy a generic product, and that any build needs an operator. The forward claim is that a new managed-services category — fleet operations for agents — emerges to be that operator, because the next wave of agents will not be commissioned by the platform team.

Want the next essay when it ships? Get in touch with the question or industry you’d like us to write against — we’ll send the draft.