Syntic
All posts

Published October 4, 2026 in Future of Work

What an AI-Native Company Looks Like

By Syntic AI

When American factories first adopted electric power, many of them did the obvious thing. They took out the steam engine, put a large electric motor in its place, and kept everything else. The same overhead line shafts ran the length of the building. The same leather belts dropped to each machine. Machines were still arranged around the source of power rather than the flow of work.

The gains were small. The large ones came only after a new generation of plants was designed around "unit drive", a motor on each machine, which let engineers lay out single-storey factories in the order the product actually moved. The economic historian Paul David used this history in 1990 to explain why computers were not yet showing up in the productivity statistics: a general-purpose technology pays off when the organization around it is rebuilt, not when it is installed.

Most companies are at the motor-on-the-old-shaft stage with AI. They have added assistants and agents to processes designed for people who remember context and ask a colleague when something is unclear. An AI-native company is one whose operating model was designed for the new power source from the start. That depends less on which models it uses than on a handful of structural choices: how work is specified, how knowledge is kept, how access is granted, how quality is measured, how compute is budgeted and how all of it is governed.

Why the gains arrive late

David's account, and Warren Devine's engineering history From Shafts to Wires (1983), describe a long lag. Edison's first central power station opened in 1882. Electric motors supplied only a small share of factory drive power at the turn of the century, and manufacturing productivity did not accelerate sharply until the 1920s, after plants had been redesigned.

Erik Brynjolfsson, Daniel Rock and Chad Syverson formalized the pattern in "The Productivity J-Curve" (American Economic Journal: Macroeconomics, 2021). General-purpose technologies require large complementary investments in new processes, business models and human capital. Because those investments are intangible and poorly measured, productivity can look weak in a technology's early years and then overshoot once they pay off. Adjusting only for intangibles tied to computer hardware and software, they estimate US total factor productivity was 15.9% higher by the end of 2017 than official measures showed.

Earlier firm-level work by Timothy Bresnahan, Brynjolfsson and Lorin Hitt (Quarterly Journal of Economics, 2002) found that IT, workplace reorganization and new products worked as complements rather than substitutes.

Recent survey evidence, self-reported and so weaker, points the same way. McKinsey's March 2025 State of AI survey tested 25 organizational attributes and found that redesigning workflows had the biggest effect on whether companies reported an EBIT impact from generative AI. Only 21% of respondents whose organizations used generative AI said they had fundamentally redesigned even some workflows. The redesign is the investment.

Processes written to be executed

In most companies a process lives in three places: a slide deck describing the ideal, a wiki page nobody updated after the last reorganization, and the heads of the people who do it. Humans fill the gaps with judgment and hallway questions. Software cannot.

An AI-native company writes its processes as specifications that a machine can act on and a person can audit. That means explicit inputs and outputs, the systems involved, the decisions a step may make alone, the thresholds that trigger escalation and the definition of done. Exceptions are part of the spec rather than folklore.

Software engineering already does this. AGENTS.md, an open format its maintainers call "a README for agents", gives coding agents a predictable place to find build steps, test commands and conventions. Its site reports use in more than 60,000 open-source projects, and it is now stewarded by the Agentic AI Foundation under the Linux Foundation. The file format is not the point. Instructions once passed on informally have become a versioned artifact that changes through review, like code.

The same discipline applies to refunds, vendor onboarding or month-end close. Once an agent executes a process document, editing it is a change to production. It needs an owner, a change history and a test before it goes live.

Knowledge as a managed asset

Agents act on whatever context they are given. In a bolt-on company, that context is a search over shared drives: drafts beside final versions, superseded policies, decisions whose reasons left with the person who made them.

An AI-native company treats its knowledge base the way a finance team treats the ledger. Documents have owners, review dates and a status. Superseded material is retired. Above all, decisions are recorded with their reasons.

Software teams have a tested pattern here too. In 2011 Michael Nygard proposed the architecture decision record: a short note giving the context, the decision, its status and its consequences. His argument was that a newcomer who cannot see why something was decided has two bad options, to "blindly accept the decision" or "blindly change it". An agent is a permanent newcomer. Without a decision log it will follow outdated rules faithfully or undo choices it does not understand.

This has a cost. Someone has to own the knowledge base, and writing things down has to count as real work in performance reviews. Otherwise agents are only as current as the last person who updated the wiki.

Access designed before it is needed

When AI is added late, permissions are granted in one of two bad ways. The agent borrows a human's credentials and inherits far more access than its task requires, or every request goes through a ticket queue and the project stalls.

An AI-native company designs access up front. Data is classified before it is connected. Each agent role gets its own identity and the narrowest set of permissions its job needs. Actions are sorted by how reversible they are, and the irreversible ones go through an approval step.

None of this is new doctrine. NIST's security control catalogue, SP 800-53, requires least privilege for "users (or processes acting on behalf of users)", limited to what is "necessary to accomplish assigned organizational tasks". What changes is how much depends on it: an agent with broad access and a misread instruction can do damage at machine speed. We cover agent identity and policy enforcement in AI agents as the execution layer for business.

Quality as a system, not a review step

The bolt-on pattern for quality is a human who checks the output. That works for a pilot. At volume, reviewers tire, skim and start approving by default.

An AI-native company builds an evaluation system the way a manufacturer builds quality control. Each important process has a set of test cases drawn from real work, including the awkward ones. Changes to prompts, models, tools or process specs are run against that set before release. Production is sampled continuously, and every failure becomes a new test case.

Defining "good" is harder than it sounds. Shankar and colleagues (UIST 2024), studying people building automated graders for model outputs, described "criteria drift": people need criteria to grade outputs, but grading outputs is how they discover their criteria. Quality standards are not written once. They are owned by the people who know the work and revised as failures reveal what matters.

Compute budgeted like headcount

In most companies AI spend sits in an IT software line or on a team's credit card. That hides the number that matters: what a unit of work costs when software does it.

An AI-native company budgets compute the way it budgets people. Each process has an owner accountable for cost per completed task, not just total spend. Budgets are set per workflow, with alerts when a change in model, prompt length or retry behaviour drives costs up. Human review and error correction are counted alongside model usage, because a cheap step that produces expensive rework is not cheap.

Planning changes too. "Should we hire for this?" and "should we run this as a workflow?" become one question, compared on cost, quality, risk and how easily the work scales back down. The headcount side of that choice is the subject of why the next company may have more AI employees than humans.

A different management cadence

The weekly review in a bolt-on company tracks the old indicators: tickets closed, calls handled, hours logged. Those can look healthy while automated work quietly degrades.

An AI-native company adds indicators for the automated part of the business. How often does each workflow escalate to a person? How often do people override or reverse what an agent did? What share of the evaluation set passes after the latest change? Which failures recurred this week, and which process spec or knowledge document did they trace back to?

The rhythm resembles site reliability engineering: incidents are reviewed without blame, root causes lead to changes in specs or permissions, and owners are named. The manager's job moves from supervising tasks to maintaining the system that performs them.

Governance as part of the operating model

Governance is where companies most often confuse documents with systems. Two frameworks help, if described accurately.

The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) is voluntary guidance from the US National Institute of Standards and Technology. It is organized around four functions: Govern, Map, Measure and Manage. A Generative AI Profile (NIST AI 600-1) followed in July 2024, and NIST says the framework is being revised. It is not a certification and does not, on its own, show compliance with any law.

ISO/IEC 42001:2023 is an international standard for an AI management system. Like ISO 27001 for information security, it sets requirements for establishing, running and continually improving a management system, and organizations can seek certification against it from accredited bodies. Certification shows that the management system meets the standard. It does not show that any particular AI system is safe or accurate.

Their value is structural. They push leadership to assign accountability, keep an inventory of AI systems, assess impact before deployment and monitor afterwards. Built in from day one, those practices are cheap to maintain. Retrofitted, they usually mean documenting systems nobody designed.

How AI-native companies fail

Designing around AI creates its own failure modes.

Over-automation. In April 2018, after heavy automation slowed Model 3 production, Elon Musk wrote that "excessive automation at Tesla was a mistake" and that "humans are underrated". Automating a step before it is stable locks in its flaws and makes them harder to see. Toyota's production system offers the opposite principle, jidoka, or "automation with a human touch", in which machines stop themselves when they detect an abnormality and any operator can halt the line.

Brittle processes. A precise specification is efficient until reality changes. A new product, regulation or customer habit can make a spec obsolete overnight, and an agent will keep following it. Specs need review dates and owners who watch for drift.

Skill atrophy. When agents handle routine cases, people lose practice at the work they oversee, and the escalations that reach them are the hardest ones. Companies have to keep humans doing some of the work, rotate review duties and train for exceptions. The team-level side of this is explored in from chatbots to digital coworkers.

Optimizing what is measured. Cost per task and escalation rate are easy to push down by handling hard cases badly. Metrics need counterweights, such as audits of resolved cases and complaint rates.

The floor plan, not the motor

Electrification is sometimes told as a story about patience: wait long enough and productivity arrives. That misses what happened. The gains went to companies that redrew the floor plan, put power where the work was and rebuilt jobs around the new layout. Companies that kept the line shafts got a quieter engine.

The AI version of the floor plan is less visible but just as concrete: processes that can be executed and tested, knowledge with owners, access designed per role, an evaluation set that grows with every failure, compute managed as a cost of production, governance inside the operating rhythm. None of it requires a frontier lab. It requires treating these artifacts as the company's real infrastructure, an argument we extend in intelligence is becoming infrastructure.

Sources