AI adoption is three transformations pretending to be one

Based on The Agentic Awakening, published by Bessemer Venture Partners and written by Liran Eshel and Adam Fisher.

most companies begin AI adoption in the same way.

buy the licenses.

run a workshop.

share a few demos.

call it transformation.

that is visible progress. but it is not the full transformation.

i just finished The Agentic Awakening, a three-part report on what happens when companies try to move from using AI tools to working in an AI-native way. it is a long read, but a very good one: clear, practical, and full of examples from companies already doing the work.

its main idea stayed with me:

AI adoption is three transformations pretending to be one.

the system must change. the people must change. then the company itself must change.

the number that matters is not 10x

one of the most advanced companies in the study had 85-90% of its production code generated by agents. code generation became close to 10x faster.

but the full journey from idea to production improved by only 25-30%. it never crossed the report’s 50% ceiling.

why?

the engineers got faster. the company did not.

product reviews still took time. PM handoffs still created queues. planning still happened in long cycles. old ownership lines still slowed decisions down.

the report uses a good image for this: you can replace a slow car with a Ferrari, but if it stops at a red light every 200 meters, the journey will not become 10x faster.

AI does not automatically remove your bottlenecks. it reaches them faster.

1. the system must change

part one is about building the environment where agentic work can happen safely and repeatedly.

giving everyone access to a coding assistant is only the entrance. agents also need:

• infrastructure for isolated and parallel work
• an AI Ops function that owns the internal agent stack
• measurement that connects usage to quality and real output
• security, permissions, audit trails, and clear human approval points
• tests and verification strong enough for agents to check their own work

the last point is especially important. autonomy only grows as far as verification allows it to grow. if a team cannot prove that an agent’s output is correct, someone has to return to reading every line.

this is why tests, types, evals, observability, and safe sandboxes are not boring side work. they are the infrastructure that makes longer and more independent agent loops possible.

measurement must also go beyond tokens or lines of code. high usage can mean great output, or it can mean an expensive agent stuck in a loop. the better question is: how much useful, complexity-weighted work did the team ship, at what quality and total cost?

without this foundation, AI adoption stays a collection of impressive demos powered by a few motivated people.

2. the people must change

part two argues that conversion starts at the top.

a leader cannot only sponsor AI adoption from a strategy deck. they need to use the tools, understand what strong agentic work looks like, fund the missing infrastructure, and hold the standard when the organization moves back toward old habits.

the engineer’s job changes too.

the new job is less about writing every line and more about orchestrating and verifying work. an engineer may manage several agent threads at once, make architecture and product decisions, reject weak approaches early, and build the checks that let good loops continue without interruption.

that does not mean judgment becomes less valuable. it becomes the main job.

this transition is also personal. years of expertise may feel less visible when the code is written by an agent. control can feel weaker when reading every line is no longer possible. companies need to name that tension honestly, offer real training, pair experienced people with strong agent users, and give everyone a fair but clear path to the new standard.

early agentic work can even feel slower. every correction needs to become a test, rule, script, or piece of project memory. over time, the project learns how the team works. that accumulated context is what eventually creates speed.

3. the company must change

part three begins after the tools work and the people know how to use them.

this is where many companies hit the ceiling.

when a feature can be built in a day, a two-week sprint starts to feel strange. when engineers can prototype directly, long requirement documents and repeated handoffs become expensive. when one person can run a small fleet of agents, old team shapes stop matching the work.

the report points toward smaller autonomous squads, often two or three product-minded people. these teams own more of the path from customer problem to production. PMs and designers can bring working prototypes instead of only documents. planning cycles get shorter, and more decisions become reversible experiments.

this does not mean removing all control. it means moving control to the places where human judgment matters most.

the report suggests strong company-wide leadership across architecture, product, design, and security. these roles keep small teams aligned without turning every decision into another approval queue.

the same pattern will not stop at engineering. RevOps, finance, HR, support, and legal teams are already building agents and internal tools around their own workflows. AI Ops can become a company-wide function that helps every team move faster without creating shadow software and new security problems.

AI-native is a company design choice

so the useful question is not only, “which AI tool should we buy?”

it is:

• can an agent complete meaningful work safely?
• do leaders use the tools themselves?
• can teams make decisions at the speed they can now build?
• are we measuring outcomes, not only activity?
• which red lights still exist because the old speed required them?

if the system does not change, AI stays a demo.

if the people do not change, usage stays shallow.

if the company does not change, the new speed gets trapped inside old processes.

AI-native is not a tool choice. it is a company design choice.

if you are thinking about AI adoption beyond licenses, workshops, and demos, The Agentic Awakening is worth your time.