Software Is Getting Cheaper to Produce

For most of the software era, building something new inside a company began with a familiar calculation: how many engineers will this take, and for how long? A meaningful internal system might require backend and frontend engineers, infrastructure support, product management, data expertise, and increasingly machine learning. Even modest ideas could become six-month projects. Many never survived the spreadsheet.

AI is changing that calculation. Coding agents can write integrations, generate tests, trace bugs, refactor systems, and handle much of the mechanical work that once consumed engineering teams. At the same time, foundation models can interpret documents, reason over messy inputs, search unstructured information, and navigate workflows that traditional software struggled to automate. The result is not simply that software is becoming faster to build. More software is becoming economically worth building at all.

The Falling Cost of an Idea

Every company has a backlog of things it would like to automate. Salespeople assemble information from five systems before a call. Finance reconciles spreadsheets. Operations teams move information between tools. Support employees search through thousands of documents. Managers spend hours assembling reports from databases that already contain the answers.

Historically, many of these problems sat below the threshold for engineering attention. Saving a department $150,000 a year sounds attractive until solving the problem requires three engineers and six months. AI pushes that threshold downward from both directions. It reduces the cost of writing ordinary software while expanding the range of work software can perform.

Palantir spent two decades proving a version of this idea at the top of the market. Its Forward Deployed Engineers sit somewhere between software engineers, product managers, data engineers, and consultants. Instead of handing customers software and leaving them to implement it, Palantir puts technical people next to the operators who understand the problem and builds around the reality of the organization. AI has dramatically increased the leverage of that model.

At Panasonic Energy, Palantir's first use case was delivered in roughly one month, reducing a multistep operational process from around four hours to 15 minutes. At TrinityRail, an AIP inventory workflow built in approximately three months generated $30 million in savings. General Mills has reported approximately $40,000 per day, or $14 million annually, in savings from an AI-driven supply-chain system. These are not stories about programmers typing faster. They are stories about compressing the distance between an expensive business problem and the software capable of solving it.

Companies Need Skills in Fractions

The paradox is that software may require less labor while demanding a wider combination of expertise. A relatively modest AI automation might require backend engineering, machine-learning expertise, infrastructure work, product design, frontend development, and someone who understands how models behave in production.

A medium-sized company cannot hire 0.4 backend engineers, 0.15 ML engineers, and 0.1 infrastructure engineers. It has to hire people. Companies therefore either stretch generalists across domains or accumulate specialists whose expertise is critical for one project and barely required for the next. Hiring also comes with recruiting, onboarding, management, benefits, equity, retention, and the difficulty of finding excellent engineers in the first place.

A specialized external team has different economics. One project might require heavy ML expertise and almost no frontend work. Another needs integrations, data engineering, and infrastructure. Specialists can move between projects while customers consume only the combination of capabilities they need. AI makes this arrangement more powerful because the specialists themselves have become more leveraged.

This is why comparing a consultant's hourly rate with an employee's salary misses the point. The relevant metric is cost per useful outcome.

Palantir provides the extreme version of this model. Its customers are often enormous enterprises solving problems worth tens or hundreds of millions of dollars. But the same economics are moving down-market. Foundation models provide capabilities that once required dedicated research teams. Cloud infrastructure replaces enormous amounts of systems work. APIs expose payments, communications, voice, authentication, and data infrastructure. Coding agents compress implementation. The amount of organization required to build sophisticated software is shrinking.

This creates a new category of opportunity for companies like Lokos AI. The forward-deployed model no longer needs to begin with a Palantir-sized contract. A small specialist team can apply the same principle to a narrower problem: work directly with the people who understand it, bring in the right combination of technical capabilities, build the system, deploy it, and iterate from real usage.

From Headcount to Leverage

For decades, engineering capacity was roughly synonymous with engineering headcount. If a company wanted to build more, it hired more engineers. But headcount becomes a poor proxy for capability when output per person begins to diverge dramatically.

A medium-sized company may not need another permanent software team. It may need ten important problems solved. This is the premise behind the forward-deployed model at Lokos AI. The objective is not to maximize engineering capacity or sell as many engineering hours as possible. It is to minimize the distance between capital invested and capability created.

The comparison is therefore not five consultants versus five employees. It is the cost of accessing a multidisciplinary technical organization when needed versus owning one indefinitely. A company can deploy a relatively small amount of capital against a specific problem, draw on software, infrastructure, data, and AI expertise as required, and judge the investment against the economic value of what gets built.

Palantir demonstrated how powerful this model can become at the top of the market. AI is lowering the minimum efficient scale. What once made sense primarily for governments and Fortune 500 companies can increasingly make sense for a mid-sized company with a valuable problem that is too specialized to justify building an entire team around it.

A decade ago, ambitious software required accumulating people. Increasingly, it requires accumulating leverage: excellent technical judgment, reusable infrastructure, specialized expertise, and AI systems that multiply what each person can produce. The winners may not be the companies with the largest engineering organizations, but those that become best at converting relatively small amounts of capital into technological capability.

The question is no longer how many engineers you can afford. It is how much you can cause to exist with the capital you have.