Anirudh Warrier

The AI fluency goal for teams is L2, not L4

I’ve been thinking about AI fluency a lot lately.

Mostly because every company is now trying to figure out what “getting good at AI” actually means.

The phrase is already becoming too broad. Someone using ChatGPT to rewrite an email is using AI. Someone using Claude to summarize a sales call is using AI. Someone building an internal agent that reads calls, updates CRM, sends Slack nudges, and routes product feedback is also using AI.

All three are “using AI”.

But they are clearly not operating at the same level.

This is my current framing: the company-wide goal should not be to make everyone L4. The goal should be to get everyone to L2.

This is the compact version of the argument.

1. “Using AI” Is Too Broad

The default founder question is usually:

How do we get everyone to use AI?

This is a good starting point, but not a good operating question.

“Use AI” can mean too many things. It can mean writing better prompts, summarizing meetings, building small tools, deploying agents, designing agent handoffs, setting up governance, or creating closed-loop test environments.

Those are not the same thing.

So the better question is:

What level of AI fluency should each person in the company reach?

My current answer is:

most employees -> L2 natural builders -> L3 system owners -> L4

This matters because pushing everyone toward L4 sounds ambitious, but it is usually the wrong goal. It is expensive, distracting, and unnecessary for most roles.

The real unlock is getting domain experts to use AI inside the actual flow of their work.

2. The AI Fluency Levels

Here is the framework I am using.

L0: Casual AI usage

This is when someone uses an LLM for fun or for basic one-off tasks. They ask random questions, generate a few ideas, rewrite a sentence, or test something once in a while.

Nothing substantial has changed in how they work yet. AI is still a side activity.

L1: Prompting and personal productivity

This is when someone starts using AI to get their own work done faster.

They understand how to give context, ask better questions, summarize things, draft faster, clean up messy notes, and use internal tools to reduce manual work.

A sales rep uses AI to prep for a call. A marketer uses it to create campaign variants. A CS person uses it to summarize customer conversations.

This is useful, but the improvement is mostly individual. The work around them has not really changed.

L2: Multi-tool workflow automation

This is when someone uses multiple tools, agents, documents, data sources, and internal systems to automate parts of real work.

Sales rep uses workflows that pulls company context, recent call notes, CRM history, open opportunities, and creates a pre-call brief.

CSM uses AI to extract risks, expansion signals, renewal concerns, product asks, and next steps, then pushing that context into the right internal workflow.

Marketers are using AI to research ICPs, compare messaging, create variants, generate assets, and prepare something much closer to execution.

At L2, AI is not a separate tab. It becomes part of how people work.

L3: Internal tool and agent builder

This is when someone starts building systems for other people.

They can build and deploy agents, dashboards, automations, lightweight tools, or workflows that teams across the company can use.

The point is that this person can look at a repeated internal problem and turn it into something usable.

Examples:

This is valuable, but not everyone needs to be here.

L4: AI systems architecture

This is when someone can design reliable AI systems across a team or company.

This includes inter-agent handoffs, closed-loop testing, memory, context, constraints, governance, permissions, evaluations, reliability, cost, and failure handling.

At this level, the question is not just:

Can we build an agent?

The questions are:

Where should the agent stop? What context should it have? What should it never do? What should it remember? What should be validated? What happens when it is wrong? How do we test it before putting it in front of a team? How do we know whether the system is getting better?

That is a rare skill.

It is also not the level you should push everyone toward.

3. L2 is the inflection point

This is the main point.

If you are running a company, the goal is not to get everyone to become as good as your best AI builder or forward deployed engineer.

That sounds good in theory. In practice, it is expensive, distracting, and unnecessary for most roles.

You do not need every salesperson to understand agent handoffs. You do not need every CSM to understand memory architecture. You do not need every marketer to spend weeks figuring out APIs, permissions, tool limits, and governance.

You need each person to become meaningfully better at their actual job with AI.

That is the difference.

The salesperson should become better at sales. The CSM should become better at customer success. The marketer should become better at marketing. The founder should become better at decision-making, hiring, fundraising, customer feedback, strategy, and communication.

L2 gives you this.

It gives you the productivity lift without pulling everyone away from their domain.

4. The Tradeoff Curve

When thinking about AI fluency, I would look at four things:

ramp-up time productivity gain token and tool cost distraction creep

Ramp-up time is how long it takes someone to move from one level to the next. Even if AI tools are cheap, learning is not free. If someone spends 20 hours learning a new workflow, that time has to come from somewhere.

Productivity gain is the obvious one. Does the person actually become better at their work, or are they just spending more time playing with tools?

Token and tool cost is not just tokens. At L0 and L1, this is usually not a big issue. Nobody is maxing out context windows or running complex workflows. Even at L2, the cost is usually manageable because the use cases are tied to specific work.

But once you get into L3 and L4, the cost starts to include tools, infrastructure, retries, debugging, maintenance, testing, and all the time spent keeping the system usable.

Distraction creep is the most underrated factor.

From L0 to L2, people are mostly still thinking inside their domain. They are using AI to become better at the work they already own.

After L2, they start getting pulled into systems, agents, dashboards, automations, integrations, permissions, edge cases, and tool limitations.

That can be useful for the right person. It can also become a rabbit hole very quickly.

A marketer who should be thinking about messaging can spend three days debugging a workflow. A CSM who should be thinking about customer risk can suddenly get pulled into CRM sync logic. A sales leader who should be improving pipeline quality can start chasing agent experiments that never become reliable.

This is not always bad.

But it needs to be intentional.

5. The Jump From L0 to L2 Is the Best Company-Wide Bet

The jump from L0 to L1 is more effort than people think.

Not because the tools are hard, but because habits are hard. People have to stop thinking, “I’ll just do this myself,” and start thinking, “Can AI help me get to a better first version faster?”

That shift takes time. Some people adopt it naturally. Some people need very specific examples from their own work before it clicks.

Once it clicks, the productivity gain is good. The cost is low. The distraction is low. This is worth doing.

But L1 is not enough.

At L1, people save time individually. The company does not necessarily get a workflow-level improvement.

The jump from L1 to L2 is the one that matters most.

At L1, AI helps the person. At L2, AI starts improving the workflow.

Someone at L2 knows how to bring in the right context, use the right tools, connect the right inputs, and get to a useful output. They are not trying to become an engineer. They are not trying to design company-wide AI architecture. They are still operating inside their domain, but they are using AI to make real work move faster.

This is the sweet spot.

The ramp-up time is higher than L0 to L1, but the productivity gain is significantly higher too. The token cost is still manageable. The distraction creep is still low because the person is focused on doing their actual job better.

For most teams, this is where the biggest company-wide lift will come from.

6. L3 and L4 Are Specialist Paths

L2 to L3 is where the curve starts changing.

At L3, the person is no longer just improving their own workflow. They are building systems for other people.

This can create leverage, but it comes with a much higher learning curve. Now they need to understand internal tools, automations, agents, dashboards, integrations, and enough system design to make things usable.

This is where distraction creep becomes real.

The productivity gain can be high, but it is not always immediate. A lot of time goes into learning, experimenting, debugging, and figuring out why something that worked once does not work reliably for a team.

Some people should absolutely go to L3. These are the people who naturally see repeated problems and want to build tools around them.

Support those people.

But do not make L3 the default expectation for everyone.

L3 to L4 is even harder.

This is not about using AI better. It is about designing AI systems that can be trusted, governed, evaluated, and improved.

That requires systems thinking. It requires understanding handoffs, constraints, memory, context, permissions, reliability, cost, testing, failure modes, and team adoption.

This is a lot.

The ramp-up time is very high. The distraction creep is very high. The cost is high. And for most employees, the productivity gain is not proportional to the effort required to get them there.

That does not mean L4 is not valuable. It is extremely valuable.

It just means L4 is not a company-wide training goal.

You need a few people who can operate at L4.

You do not need everyone to become L4.

7. Domain Expertise Still Matters

This is the part that I think gets missed.

Saying L4 is rare and valuable does not mean every L4 person is more useful than every L2 person.

An L2 person with deep domain expertise can beat an L4 person with shallow context.

A salesperson with L2 AI fluency and deep customer understanding can be more valuable than an AI-native generalist who does not understand the market.

A CSM with L2 fluency and years of customer pattern recognition can produce better outcomes than someone who can build agents but does not understand customer risk.

A marketer with L2 fluency and good taste can beat someone who understands AI systems but does not understand positioning.

The rarest person is someone with both:

deep domain expertise L4 AI fluency

Those people are incredibly valuable.

But you cannot build your whole hiring plan around assuming everyone will be that.

That is not a strategy. That is hope.

8. What Leaders Should Do

The practical strategy is simple.

First, get everyone to L2.

Not through generic AI training. Not through a list of 50 prompts. Use real workflows from their own function.

Sales should learn AI inside account research, call prep, follow-ups, CRM hygiene, and pipeline inspection.

CS should learn AI inside call analysis, risk detection, expansion signals, product feedback, and renewal prep.

Marketing should learn AI inside research, campaign creation, content repurposing, messaging, and experimentation.

Product should learn AI inside feedback analysis, spec writing, user research, and prioritization.

Founders should learn AI inside decision-making, hiring, fundraising, customer feedback, strategy, and internal communication.

Second, identify the natural L3 builders.

Some people will naturally start creating workflows and internal tools. They will see repeated problems and want to automate them.

Support those people, but do not force everyone into that path.

Third, hire or develop a few L4 people.

These are the people who can think about the system as a whole. They should design the architecture, governance, memory, handoffs, evaluations, and reliability layer.

Their job is not to replace everyone.

Their job is to make everyone else more capable.

Finally, protect domain expertise.

AI fluency without domain judgment is dangerous. The goal is not to replace domain experts with AI-native generalists.

The goal is to make domain experts more powerful with AI.

That is the real unlock.

9. A Small Interactive Experiment

I am also trying to turn this idea into an interactive module that helps every employee in your team to start thinking differently about AI.

Because I do not think this framework is best understood only by reading it.

The interesting part is not memorizing what L0 to L4 means. The interesting part is feeling the tradeoffs.

When does AI save time? When does it improve a workflow? When does automation become useful? When does it become a distraction? When does someone need to stay focused on their domain? When does someone need to become a builder? When does a team need an L4 architect?

That feels more useful as an interactive game than as another AI guide.

Not a course.

Not a prompt library.

Something closer to a small simulation of how AI fluency actually shows up at work.

The Big Picture

For founders, the important questions are:

My current answer is:

Get everyone to L2
find your L3 builders
hire or develop a few L4 architects
protect domain expertise

Do not confuse AI fluency with AI distraction.

The goal is not to make everyone L4.

The goal is to make L2 the new baseline.