In the AI-Era, Communication is Becoming an Execution Skill

I have worked with developers for many years, and I have worked with all kinds of them.

Some are excellent communicators. Some are not.

Before AI became part of everyday software development, a technically strong developer could still be very good at their job without being particularly good at expressing themselves. It wasn't ideal, of course. Communication has always mattered in a team.

But if the developer understood the problem and had the technical ability to solve it, they could express a large part of that understanding through code.

I think AI is changing this equation.

As more execution shifts to AI, developers need to do something they could sometimes avoid earlier: they need to express what they understand clearly enough for another intelligence to act on it.

That makes communication more than a soft skill.

It is increasingly becoming part of execution itself.

Let us simplify software development for a moment.

In the pre-AI era, much of a developer's work could be described as:

Understand the problem → Execute the solution

Obviously, real software development has never been quite that simple. There are requirements, architecture decisions, testing, reviews and collaboration.

But the important point is that once a capable developer understood the requirement, much of the execution happened inside their own head and through their own hands.

They translated their understanding into code.

A developer who wasn't particularly articulate could therefore still be extremely effective. Their code could do a lot of the talking.

Communication became increasingly important as they moved toward senior, client-facing or leadership positions, but weaker communication did not necessarily prevent someone from being a strong individual contributor.

AI introduces another interface into this process.

Consider how the same developer may work today.

They receive a requirement.

They still need to understand it. AI has not removed that responsibility.

But instead of personally writing every part of the implementation, they may ask an AI coding system to produce, change, test or review some of it.

The workflow starts looking more like:

Understand → Communicate → AI executes → Evaluate → Refine

The first step remains human.

So do the last two.

But that new step in the middle "communicate" is becoming extremely important.

AI cannot see what is inside the developer's head.

It doesn't automatically know the history of the project, why a particular architecture was chosen, which business rule is unusual, what the client actually meant, which existing behavior cannot change or what compromises have already been rejected.

The developer has to transfer the relevant parts of that understanding.

And, increasingly, that transfer happens through language.

This is one of the things I find most interesting when watching people work with AI.

Imagine giving two developers access to exactly the same coding system.

The first says:

Build this feature.

The second explains what the feature is supposed to achieve. They describe the current behavior, relevant architecture, constraints, exceptions, dependencies and what a successful implementation should look like.

They both have access to the same AI.

But they haven't given the AI the same ability to help them.

This does not mean AI only benefits highly experienced developers. Large field experiments involving 4,867 developers at Microsoft, Accenture and another Fortune 100 company found a 26.08% increase in completed tasks among developers using an AI coding assistant, with less-experienced developers showing higher adoption and greater productivity gains.[1]

That is one of AI's great advantages: it can help people perform work that might previously have been beyond their immediate ability.

But as the problem becomes more complex, another issue becomes increasingly important.

How much of the problem does the human actually understand, and how much of that understanding can they successfully transfer to AI?

AI can help fill gaps in knowledge.

It cannot reliably compensate for context that was never communicated.

There is a lot of discussion about the cost of AI.

We count tokens. We compare models. We look at usage and computing costs.

Those things matter, particularly when AI is being used at scale.

But I think there is a much larger cost for software companies.

Building the wrong thing faster.

Suppose a developer hasn't understood a requirement properly or has understood it but communicated only half of the relevant context.

AI may still produce perfectly reasonable code.

The code gets reviewed. It gets tested. Someone integrates it. Eventually, perhaps the product manager or client sees it and says:

That isn't what we needed.

The cost of the AI tokens used to produce that code is probably the least important part of what just happened.

The real cost is the human time, testing, review, rework and delay created because the wrong problem entered execution.

AI makes production faster.

That is useful when we are moving in the right direction.

But speed works both ways.

Without clear direction, AI can become very efficient at solving problems no one actually has.

When I talk about communication, I am not talking about learning a collection of clever AI prompts.

And I don't simply mean having excellent English.

I mean the ability to take something you understand and make someone or something else understand it correctly.

1. Can you explain the actual problem?

2. Can you separate important context from unnecessary information?

3. Can you explain an exception that changes the solution?

4. Can you describe what already exists and what cannot be disturbed?

5. Can you recognize when your instructions have been misunderstood?

6. Can you explain why the proposed solution is wrong?

7. Can you define what a correct outcome should look like?

These sound like communication skills.

But they are also evidence of clear thinking.

It is difficult to explain a problem clearly if you haven't understood it clearly yourself.

This is not only something I am noticing inside a software company.

A 2026 study of 163 developer–Claude Code interaction episodes identified what the researchers call a human-AI context gap: a divergence between the context held by the developer and the context available to the AI. The study links unresolved context gaps to communication and coordination breakdowns, functional errors and escalating corrective effort.[2]

Other research on AI-assisted software development similarly shows that the developer's role does not disappear when AI produces code. A 2026 mixed-methods study involving 76 developers found that AI-assisted programming still involved developers actively creating code as well as evaluating and verifying generated results.[3]

That is an important distinction.

AI is not simply replacing:

Human developer → AI developer

The division of work is changing.

Some execution moves toward AI, while understanding, context, judgement and responsibility remain heavily dependent on the human.

Communication is what connects the two.

There is an easy conclusion to draw from AI:

1. If AI can write increasingly good code, developers won't need as much technical knowledge. I don't think that is the right conclusion.

2. If AI proposes an architecture, somebody still needs to understand whether it fits the system.

3. If AI changes a database query, somebody needs to recognize the implications.

4. If it misunderstands a business rule, somebody needs to catch it.

5. If it produces a solution that works today but creates problems elsewhere, somebody needs enough understanding to see that.

The better AI becomes at producing plausible solutions, the more important judgement may become.

So I don't see communication replacing technical capability.

I see communication becoming the bridge between human capability and AI capability.

A developer needs enough knowledge to understand the problem, enough communication ability to transfer that understanding, and enough judgement to evaluate what comes back.

Remove any one of those and the value of the others falls quickly.

If I were beginning a career in software development today, I would absolutely learn how to work with AI.

But I would not make “learning AI tools” the entire plan.

1. I would learn my technical field properly.

2. I would learn to understand requirements before starting implementation.

3. I would learn to write clearly.

4. I would learn to explain complicated things simply.

5. I would learn to ask questions when something doesn't make sense.

6. I would learn to document decisions and assumptions.

7. I would learn to challenge an answer that sounds convincing but doesn't actually solve the problem.

These abilities mattered before AI.

What has changed is their leverage.

In the pre-AI world, a developer could understand a problem and express much of that understanding directly through code.

In the post-AI world, more of the execution can be handed to AI.

But before AI can execute what you want, you first have to make it understand what you want.

That is why I increasingly believe that, for people working in technology, communication is no longer only a soft skill.

It is becoming an execution skill.

[1] Zheyuan (Kevin) Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng and Tobias Salz - “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.”
Microsoft Research. The randomized field experiments covered 4,867 developers at Microsoft, Accenture and an anonymous Fortune 100 company and reported a 26.08% increase in completed tasks among developers with access to an AI coding assistant.
https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/

[2] Dov Te'eni and Adi Katz - “Reliable Vibe Coding: The Human-AI Context Gap in Software Development.”
HHAI 2026 / SAGE Journals. The study analyzed 163 developer–Claude Code interaction episodes and identified recurring communication and coordination breakdowns centered on human-AI context gaps.
https://journals.sagepub.com/doi/10.3233/FAIA260505

[3] Yinan Wu, Ze Shi Li, Kathryn Thomasset Stolee and Bowen Xu - “How Do Developers Interact with AI? An Exploratory Study on Modeling Developer Programming Behavior.”
arXiv, 2026. Mixed-methods study involving 76 developers; AI-assisted developers remained involved in creating code and evaluating/verifying generated results.
https://arxiv.org/abs/2604.16393