← Blog

Lock, Stock and Two Thousand Approvals

Summary

A research-backed perspective on organizational inertia, AI-native productivity, and how to redesign enterprise workflows around governed agents.

RB
Roozbeh Bandpey

Co-founder and CEO of Bandpey GmbH

Lock, Stock and Two Thousand Approvals

In the 1890s, factories got electricity, and for about 30 years, productivity barely moved. Why? They pulled out the steam engine and bolted an electric motor into the exact same spot: same belts, same shafts, same floor plan. New power, old factory. That is how most companies are integrating AI right now: a chatbot bolted onto the old workflow. The factories that won did something different. They put a small motor on every machine, then rearranged the entire floor around the flow of work. That is when productivity exploded.

The real history is messier than the shorthand, but economists have studied it for decades. Paul David showed the gains arrived only once factories were redesigned around the new power source, and David and Wright date the surge in US manufacturing productivity to the 1920s. The electricity was there. The factory wasn’t ready for it.

The gospel of org charts

I have spent years consulting large enterprises on their operating models, structures, and data and AI strategy. The hardest part to change, sometimes close to impossible, is the organization itself. Until that changes, even the best technology barely moves the needle.

In one client workshop, a slightly technical question went unanswered by a room full of managers. It later took an eight-email chain and three weeks to reach the engineer who knew the answer.

Organizations carry two kinds of complexity. The first comes from the problem itself: regulation, system interactions, edge cases nobody can remove. That is the real work. The second exists because nobody was brave enough to simplify. Most of it is decisions that were never made. It is almost always the bigger of the two, and almost always defended by the people who created it.

A year ago, my company started integrating agents that actually execute work. On several productivity measures, we have improved at least 6x. Most operational decisions now take under ten minutes. They are hard decisions in complex territory, but the context, the authority and the accountability sit in one place.

Being small doesn’t give you this for free. Founders hold full authority and still grow into processes they never question, often because they built them.

The contrast between our speed and my clients’ inertia leaves me with two questions: how can mid-sized and large enterprises adapt, and how can I help them?

When headcount stops looking like growth

In AI-Native Firms, AI-native startups in the Y Combinator sample have roughly 25% fewer employees, a 13% higher share of engineers and a 15% lower share of managers than comparable startups. Yet their valuations are statistically similar. A broader US venture-backed sample shows the same direction, although a smaller difference in headcount.

For years, investors read headcount as a proxy for scale: a company that was hiring was a company that was growing. AI breaks that shortcut. The better measure of an organization is how much work it gets done, and how fast.

AI-native startups hire more engineers and fewer managers.

Workforce shares, Y Combinator sample · AI-native firms also have about 25% fewer employees

Source: Kim & Koning, AI-Native Firms (2026), YC sample, 2,233 firms. Industry/cohort-adjusted differences: engineers +4.5 pp, managers −3.7 pp, entry-level −4.0 pp, applied to the reported non-AI reference shares. Shares are separate measures and need not sum to 100%. Observational, not a causal headcount forecast.

Caveat: these are young companies. Nobody should read this as permission to cut a quarter of an enterprise workforce by installing agents. What it does suggest is more interesting: a business designed around AI may never build some of those management and coordination layers at all.

The productivity gain is in the workflow

In Mapping AI into Production, researchers ran an experiment with 515 startups. All received AI training and technical resources. Some also learned how to identify work that could be redesigned around AI.

Over ten weeks, that group discovered 44% more AI use cases, completed 12% more tasks, and was 11 percentage points more likely to have paying customers. Revenue results were promising but less certain; founders also reported needing less external capital.

Capital needs fell. The revenue gain is still uncertain.

515 startups, randomized, 10 weeks · effect multipliers with approximate 95% intervals (log scale)

Source: Kim, Kim & Koning, Mapping AI into Production (2026), Tables 3–5. Tick: exp(β); whiskers: approximate 95% intervals exp(β ± 1.96 SE). Both models use ln(outcome+1), N=479. Revenue p<0.10; capital p<0.05. Transformed estimates, not arithmetic-mean ratios. Capital sought is stated demand, not spending. Dashed line = no effect (1×).

Every startup in the study had the same access to a model. The difference was mapping where AI changes the flow of work.

Locked and stuck in approvals

A company can feed its documents to the most capable LLM available and still achieve very little if every department keeps its own silo. The model may summarize an answer beautifully. The customer request still needs three systems, four approvals and someone to move information between them.

Real agents need access to the right systems, explicit permission boundaries and hard rules: what they may do, what they must never do, and when a person must decide. They should present the decision-maker with the evidence, the applicable rules and the proposed action.

In that operating model, the human coordinator stops chasing emails and starts making decisions. Agents handle retrieval and routine execution. People retain judgment, accountability and control over exceptions.

Agents also expose who benefits from the old structure.

If a manager's authority depends on department size, removing layers can threaten a budget and a career path. If a vendor makes money selling Copilot licenses per employee, a workflow that needs fewer seats may not be its favorite commercial outcome. Meanwhile, executives and boards are already dealing with decision fatigue because information moves through clogged organizational veins.

Some approvals earn their place: security, financial controls, regulated decisions. Those can run as rules and be audited automatically, instead of sitting in an inbox for another week.

What large enterprises can build on

Established companies start with what startups lack: customers, distribution, capital and a brand people trust. The open question is whether they can turn decades of know-how into something agents can use.

Canaries in the Gold Mine links recent productivity gains from AI to exactly that kind of know-how. In US public firms, 57.5% of AI-related jobs build what the researchers call organizational capital, compared with 31% of other jobs.

Organizational capital is the know-how built into how a company works: reusable processes, decision rules, coordination and systems that stop people from reinventing the wheel.

The biggest return from AI shows up in how the whole organization learns, coordinates and acts, more than in any one employee’s output.

The Agentic Facthory

I built Facthory for the problem I kept seeing in large enterprises: the knowledge exists, but it is stuck between departments, systems and approval chains. Facthory puts that knowledge, the people and the agents in one shared context, with explicit permissions and hard rules.

Take that workshop question again. An agent finds the answer in the engineering systems, checks which rules apply and sends the manager a one-page decision brief: the question, the evidence, the applicable rule and a recommended action. The manager decides in minutes. The three weeks of email disappear, and people spend their time on the calls that genuinely need judgment.

The factories that won rewired the floor around the work. The companies that win with AI will rewire authority, decisions and rules the same way.


Research note: The startup comparisons are observational; the experiment concerns early-stage firms, not large-enterprise transformations. The 6x productivity figure describes our internal experience, not an independently measured industry benchmark.

About the author

Roozbeh Bandpey

Co-founder and CEO of Bandpey GmbH

Roozbeh Bandpey is an engineer and entrepreneur in Munich with more than 20 years of hands-on software engineering. He is the co-founder and CEO of Bandpey GmbH, which builds Facthory, Emolysis, Copolly and Inteliwert, and a Databricks Solutions Architect Champion.

Related insights and work