The SaaS Free Pass Is Over

Code is getting cheaper. Making it work is the business.

If better models were enough, the companies building them would not be investing this heavily in implementation.

On May 11, OpenAI announced the OpenAI Deployment Company and an agreement to acquire Tomoro, a deal expected to add approximately 150 forward-deployed engineers and deployment specialists.¹ On June 30, AWS announced a $1 billion investment in a dedicated Forward Deployed Engineering organization. On October 2, Anthropic announced a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027.^2,3

These announcements do not prove that every services business deserves a software multiple. They do show where major AI suppliers are committing resources: the work between a capable model and a functioning business system.

"The scarce thing is not another AI demo. It is a system a business can trust, operate, and improve."

That is the more useful challenge to the SaaS playbook. Software is not disappearing. But a subscription is a billing mechanism, not a moat. And human involvement is a cost to evaluate, not an automatic reason to dismiss a business.

The feature moat is under pressure. The production burden is not gone.

The build-versus-buy argument now has evidence behind it. In an August 2026 McKinsey survey, 32% of respondents reported at least one forgone software purchase after their organization used coding agents to build an internal alternative.⁴ That does not mean 32% of SaaS spending disappeared, nor that enterprises replaced their core systems over a weekend. It means internal development is becoming a credible alternative for at least some purchases in the surveyed organizations.

There is also experimental evidence for faster software production. A 2025 study pooling field experiments across 4,867 developers found a 26.08% increase in completed tasks with an AI coding assistant. The result applies to those tools and settings, not every development workload.⁵

The strategic implication is narrower than "code is free," but more actionable: a feature that was expensive to create may become cheaper for a competitor, incumbent, or customer to reproduce. Reproducing the interface is not the same task as migrating records, preserving permissions, validating edge cases, training operators, and keeping a service running. A serious build-versus-buy comparison has to price those obligations too.

The opportunity is the gap between using AI and changing the business.

In the same 2026 McKinsey survey, only 37% of respondents attributed any enterprise-level earnings-before-interest-and-taxes (EBIT) impact to AI. That separates adoption from demonstrated financial impact.⁴

This is the gap a deployment business should be paid to close. Not "we installed an agent," but "the work now happens faster, more reliably, or at a lower fully loaded cost." Lyft offers a concrete example. In December 2025, it reported an 87% reduction in average customer-support resolution time after deploying an AI assistant built with AWS and Anthropic technology. Its account describes an agent connected to backend context and capable of taking action, rather than just answering general questions.⁶

Forward deployed is an operating model, not a new rate card.

Palantir describes forward-deployed engineering as placing engineers close to customer problems while they work with core engineering to feed learning back into shared platforms. The feedback loop matters as much as the embedding.⁷

A useful FDE engagement combines customer context, engineering authority, and responsibility for production results. It should shorten the distance between discovering a problem, changing the system, and learning whether the change worked. The distinction for services firms is just as important. AWS's partner-led FDE model explicitly describes building reusable delivery capabilities inside consulting partners, with delivery IP remaining with the partner.⁸ But reusable means something specific: the next engagement benefits from tested connectors, deployment recipes, evaluation methods, or operational knowledge that the firm has the right to reuse.

Margins still matter. The category label does not settle them.

There is no single "AI margin." Bessemer's 2025 study of 20 selected high-growth AI companies described one group, its "Supernovas," averaging roughly 25% gross margin, and another, its "Shooting Stars," around 60%.⁹

Palantir provides an important counterexample. In Q2 2026, it reported a GAAP gross margin of about 84.7% and an operating margin of 47%. These are company-wide results; they do not isolate what FDEs caused, but they prove AI and high margins can coexist.¹⁰

Model the payoff. Do not assume it.

Better retention can justify deployment spending. It does not automatically justify any amount of deployment spending. Consider a hypothetical customer cohort generating $1 million in its first year. The investment case is not "trade 20 points of margin for engineers." It is "show that better retention, pricing, expansion, or delivery efficiency pays for the additional work." Move the sliders below to test the trade-off.

The uncomfortable truth: dependency can backfire.

The strongest counterargument arrived on September 29, 2026. Gartner predicted that by 2028, 70% of enterprises would abandon agentic AI built through vendor FDE arrangements because of rising costs and inability to evolve the systems themselves.¹²

The warning is relevant even without treating that percentage as destiny. A customer that cannot maintain its own system has not necessarily become a better customer. It may have become a future replacement project. The goal is a customer that chooses to keep you, not one that cannot survive without you. Document the system. Transfer operational knowledge. Make ownership clear.

The scorecard should get harder, not trendier.

Keep the financial disciplines that work. Add the operating evidence needed to understand what the revenue actually costs to create and sustain. A company that can demonstrate these things has an economic argument. A company that merely says "we have FDEs" has a staffing description.

Keep measuring	Add evidence about	The question it answers
Revenue & recurring revenue	Contracted recurrence, repeat projects, expansion, cancellation exposure	How durable is the revenue, and what type is it?
Gross margin	Account-level contribution after inference, support, and continuing delivery	Does successful usage create attractive economics?
Acquisition & payback	Time to first verified value and unrecovered deployment investment	How much capital is required before an account earns its keep?
Retention & expansion	Cohort results, customer outcomes, and reasons for renewal	Are customers staying because the system works?
Engineering output	Delivery hours per comparable deployment and reuse of tested assets	Does experience make the next engagement more efficient?
Customer satisfaction	Operational ownership, maintainability, and successful knowledge transfer	Can the customer operate the system and still choose to buy more?

Code is getting cheaper. Accountability is not.

Software founders should not assume that recurring billing protects a product whose differentiated value is becoming easier to reproduce. Services founders should not assume that an AI label turns repeated manual effort into software-like leverage.

The premium belongs to whoever can turn capability into results repeatedly, profitably, and without making the customer a hostage.

How Enfuse forward-deployed engineering solves for this.

At Enfuse, forward-deployed engineering is built around the scorecard above rather than the org chart. Small senior teams embed with the customer's operators, security staff and platform owners. They ship a working system inside the customer's boundary, then hand it over. We do not stretch out billable hours, and we do not leave behind a black box only we can run.

Four practices keep the economics honest:

• Time to verified value, not time to kickoff. Every engagement starts from one operational outcome with an agreed measure. Work is scoped to prove that outcome early, so unrecovered deployment cost stays small and visible.
• Reusable, tested assets. Delivery runs on the Sovereign Runtime + Factory: governed deployment patterns, evaluation harnesses and production components that carry from one engagement to the next. Each deployment should take fewer engineering hours than the last comparable one.
• Model-agnostic and boundary-aware. Frontier models plan where policy allows. Open models served with vLLM run on customer-controlled infrastructure, in Azure, Google Cloud, Azure Local or fully offline, where the data requires it. Customers are never locked to one model vendor or one hosting choice.
• Ownership transfer by design. Runbooks, architecture records, evaluation suites and operator training are deliverables, not extras. The goal is a customer that can run the system without us and still chooses to keep working with us.

That is the difference between a staffing description and an operating model. If you are weighing build versus buy for an AI system that has to run inside your boundary, talk to an Enfuse architect or review our forward-deployed engineering services.

---

Sources and evidence notes

Research checked October 4, 2026. Announcements, self-reported results, forecasts, and illustrative calculations are distinguished in the text.

1. OpenAI. OpenAI launches the OpenAI Deployment Company. May 11, 2026. (Announced company launch and agreement to acquire Tomoro.)
2. Amazon / AWS. AWS invests $1 billion to embed AI forward deployed engineers with customers. June 30, 2026.
3. Anthropic. Anthropic invests $100 million to train 10,000 engineers. October 2, 2026.
4. McKinsey. The state of AI in 2026: On the road to ROI. August 25, 2026.
5. Microsoft Research / Cui et al. Evidence from Three Field Experiments with Software Developers. June 2025.
6. Lyft. AWS and Lyft: Bringing Agentic AI to life for Riders and Drivers. December 1, 2025.
7. Palantir. Architecture center overview. Accessed October 4, 2026.
8. AWS Partner Network. Introducing Forward Deployed Engineering for Partners. 2026.
9. Bessemer Venture Partners. The State of AI 2025. August 13, 2025.
10. Palantir Investor Relations. Palantir reports Q2 2026 financial results. August 3, 2026.
11. Stanford HAI. The 2025 AI Index Report. April 7, 2025.
12. Gartner. Predicts 70% of enterprises will abandon agentic AI built by vendor FDEs by 2028. September 29, 2026.
13. Intercom. Fin AI Agent outcomes. Accessed October 4, 2026.