Code is no longer the wall. What should companies own in the AI era?

Let me put it bluntly: if a company still defines its core asset as "the code we wrote", that definition is getting old.

Code still matters. Products do not run on strategy decks. But with tools such as Claude Code, Codex, Cursor and CogGlance getting better, code is becoming easier to generate, replace and rewrite. It still needs engineering judgment, tests and release discipline. It is just less likely to be the moat by itself.

The numbers point in the same direction. Stanford HAI's 2025 AI Index says 78% of organizations reported using AI in 2024, up from 55% the year before, and that model performance on SWE-bench rose sharply in one year. Anthropic's Economic Index, based on real Claude usage, found that usage is concentrated in software development and technical writing tasks. In its dataset, AI use leaned more toward augmentation, about 57%, than direct automation, about 43%.

Put those facts together and the picture is fairly plain. AI has entered everyday work. But what it replaces first is not the company. It replaces chunks of work that can be described, checked and repeated.

So the question changes. If code is no longer scarce, what is?

AI assets are the systems that help AI do the right work

I would split AI-era company assets into seven buckets. None of them is just a codebase. Most of them eventually show up as systems, processes, data, permissions and habits.

1. Business context

AI's biggest weakness is not that it cannot write. It is that it often does not know why it is writing.

Valuable context includes who the customers are, why they buy, why they leave, what sales promised, how support handles edge cases, which requests should be rejected, which mistakes the company has already made, which compliance lines cannot be crossed, and which users care about speed versus stability.

In most companies, that knowledge is scattered across Feishu, Slack, CRM, support tickets, product docs, sales calls and people's heads. Yesterday it was just "documentation". In the AI era, it becomes the fuel that determines whether the model can do useful work.

Simon Willison's 2025 note on "context engineering" quotes Tobi Lutke and Andrej Karpathy making a similar point. Lutke said the phrase is better than prompt engineering because it describes giving an LLM enough context for the task to be solvable. Karpathy argued that industrial LLM apps are not about typing a clever prompt; they are about placing task descriptions, examples, RAG data, tools, state and history into the context at the right moment.

That should make managers uncomfortable. If your business context still lives only in veteran employees' heads, you do not have an AI asset. You have tired people.

2. Workflow assets

In the old software world, companies encoded workflows into systems: submit, approve, assign, review, archive. In the AI era, a workflow also has to say where AI enters, what it reads, what tools it can call, who checks the output and how failure is rolled back.

For a SaaS company, AI should not only help engineers write code faster. It can participate in the whole chain from customer feedback to requirement triage, prototype, test cases, staged release and updated support scripts. The asset is not the code for one feature. The asset is that the chain is structured, observable and accountable.

LangChain made a useful distinction in its discussion of multi-agent systems: tasks that mostly read are easier to parallelize than tasks that write, because actions carry implicit decisions and conflicting writes create bad results. For companies, that is practical advice. Do not start by letting a swarm of agents change production systems. First make reading, retrieval, analysis, summarization and validation reliable. Then open write permissions step by step.

3. Evaluation assets

Many companies can say, "We use AI." Ask them how they know the AI is right, and the room gets quiet.

Evaluation assets include gold answers, bad-case libraries, regression sets, review rules, model performance records, cost data and latency data. They are not glamorous. They decide whether AI can move from demo to production.

This matters even more for SaaS and internet companies. A wrong support reply may cause a complaint. A wrong billing change, permission change, compliance judgment or ad-spend recommendation can become an incident. Without evaluation, AI is just a confident contractor.

4. Private data and feedback loops

A data asset is not a pile of files dumped into a vector database. That is just moving the mess.

Useful data has clear sources, permissions and versions. It can be searched, corrected and fed back into the workflow. Support conversations, sales objections, product behavior, failed tickets, code review notes and A/B test results can become a company learning system if they are cleaned and connected.

Small companies may not have massive datasets. That is fine. They can still have dense datasets. A complete history of 50 real customers often beats 100,000 dirty log lines.

5. Tool and permission orchestration

For AI to do real work, it needs tools: repositories, databases, ticket systems, payment consoles, email, CRM, cloud platforms and deployment systems. The asset is not how many tools are connected. The asset is the permission design.

Which tasks are read-only? Which can draft changes? Which require human confirmation? Which can execute automatically? Which actions need audit logs? Which data must never be sent to an external model?

This becomes the company's AI operating system. The clearer it is, the more work AI can safely do. The messier it is, the less management will trust it.

6. Brand, distribution and user trust

When features become easier to copy, why should users stay with you?

The answer is usually not code. It is trust, domain fit, service, switching cost and how deeply the product sits inside the customer's workflow. AI helps competitors build similar features faster. It also helps you improve faster. What remains is whether you understand the vertical better, respond better and learn from usage faster.

7. Organizational learning

The most underrated AI asset is habit.

If every employee uses AI as a one-off chat box, the company may get faster, but it will not compound. The better pattern is to preserve good prompts, context packs, evaluations, failure cases and workflow templates so the next person does not start from zero.

This is not the old knowledge-base story. Old knowledge bases were mostly written for humans. Now part of the knowledge base has to be written for agents as well. A document for people and context for an agent are not the same artifact.

What small companies should build

Small companies usually lack time and order, not model capability. They should not begin with a complex AI platform. They should start with assets that become useful immediately.

• A customer-question library.Collect real questions from sales, support and founder conversations. Tag customer type, scenario, outcome and the correct answer. This is one of the cheapest and most useful AI assets a small company can build.
• Standard operating flows.Break delivery, pricing, content production, release and after-sales work into steps. AI can help at each step, but each step needs inputs, outputs and acceptance criteria.
• Reusable context packs.Company description, product boundaries, pricing rules, forbidden claims, industry terms and typical customer stories. Employees should not have to explain the company to AI every morning.
• A lightweight evaluation set.Do not start with an elaborate benchmark. Prepare 30 to 100 real questions, standard answers and unacceptable mistakes. Run them whenever you change model, prompt or workflow.
• Founder judgment turned into rules and cases.In many small businesses, the real edge sits inside the founder's judgment. In the AI era, that judgment needs to be turned into rules, examples and counterexamples.

For small companies, the goal is not to look advanced. It is to onboard people faster, reduce repeated work and make customer responses more consistent.

What large companies should build

Large companies have the opposite problem. They have data, systems and budgets, but context is fragmented, permissions are complex and responsibility is spread across layers. AI projects easily become disconnected pilots.

• A unified context layer.Connect product, customer, contract, permission, process, knowledge and historical decision data. Not all data should go to models. The point is to provide the smallest sufficient context inside the compliance boundary.
• Enterprise evaluation and audit.Every important AI use case needs quality metrics, red-line cases, replay mechanisms, human review rules and clear ownership.
• Model routing.Different tasks deserve different models. Not every problem needs the most expensive model, and not every sensitive task can go to an external API. Routing, cost, latency and safety policy become assets.
• AI permission governance.Large companies cannot only ask whether AI can do something. They have to ask who authorized it, who owns the result and how failure will be investigated.
• Cross-department process redesign.If AI only makes each department a bit faster, the gain is limited. Larger gains come from shortening end-to-end flows, such as customer issue to product fix, compliance review to release, or market signal to sales action.

For large companies, AI assets are the structures that turn organizational complexity into something machines can understand, execute and monitor.

SaaS and internet companies have a different asset map

SaaS and internet companies can misread this shift because they already know how to build software. Their first instinct is to treat AI as faster coding. That is only the first layer.

For SaaS companies, at least five AI assets matter more.

• In-product behavior data.Where users get stuck, which features are tried once, which paths lead to renewal, which permission settings cause errors. This data can guide AI-assisted product improvement, support, sales and retention.
• Domain workflows.Generic CRM, project management and support tools will become easier to clone. The harder thing to copy is the workflow of a specific industry: cross-border ecommerce selection, clinic scheduling and compliance, construction change orders, education renewals.
• Customer-success knowledge.SaaS value is not only features. It is whether customers actually succeed. Best practices, implementation plans, failure cases, training materials and customer segmentation should become callable assets.
• Embedded-agent feedback loops.If users operate AI inside the product, the product can learn from questions, intent, edits, accepts, rejects and reuse. That is closer to real work than page clicks.
• A trusted integration network.SaaS products connect customer data and workflows. The company that connects more systems safely sits closer to the customer's work.

The SaaS moat shifts from a feature list to domain context, feedback data, workflow embedding and trust. Code is the surface layer.

What counts as AI-era talent?

The AI era does not remove the need for people. It stops rewarding people who only complete isolated tasks.

In the past, someone who could write code, make visuals, write copy or build spreadsheets had a clear seat in the company. Those skills still help, but AI compresses them. The new gap is elsewhere: who can define the problem, organize context, judge the output, place AI inside a real workflow and take responsibility for the result.

• Problem definition.Turning "build a system" into goals, constraints, users, data, risks and acceptance criteria. AI executes clear work well. It does not remove strategic vagueness from management.
• Context organization.Preparing the right material for AI, knowing what must be included, what will distract it and which historical decisions cannot be lost. Many jobs will become a kind of context engineering.
• Review and acceptance.Seeing where AI is wrong instead of being impressed by fluent output. Engineers review code. Operators review tactics. Finance reviews definitions. Legal reviews risk.
• Workflow redesign.Not adding an AI button to every old step, but redesigning the split between people and machines. Some steps should disappear. Some should merge. Some can be automated. Some must stay human.
• Tool composition.Connecting models, databases, automation tools, internal systems, monitoring and permissions. Many people can use AI once. Far fewer can turn it into a production workflow.
• Ownership.When AI is wrong, "the model did it" is not an answer. In a company, a person still owns the result. Stronger AI needs clearer owners.

What managers have to change

Managers should stop treating AI as a small productivity tool for employees. That view is too narrow.

A better view is that AI changes the unit of work inside the company. A role used to own a long chain of actions. Now many actions can be split off to AI, while people define, authorize, inspect and handle exceptions. Management shifts from watching what people did all day to designing how tasks are decomposed, assigned, checked and accumulated.

• Hire for leverage, not only skills.Someone who can use AI to organize ten times more output may be more valuable than someone who can manually finish one task. Interviews should test problem breakdown, output validation and workflow design.
• Evaluate results and system contribution.Using AI to work faster is not enough. Did the person turn the work into a template, evaluation, workflow or data asset that helps the next run?
• Open permissions gradually.If AI never gets tool access, it stays a chatbot. If it gets too much access too soon, it becomes a risk. Management needs staged permissions, not a swing between total ban and total freedom.
• Spend on capability, not just subscriptions.Buying AI tools is not transformation. The real spending goes into data governance, evaluation, workflow redesign, employee training and audit.
• Allow jobs to change.Some old roles will shrink. Some people will become AI workflow designers, evaluation owners, data stewards and agent operators. Management cannot understand new work only through the old org chart.

The short version

Code still matters in the AI era, but code is no longer enough to define a company's core asset.

The assets worth building are business context, private data, evaluation systems, workflows, permission governance, user feedback, organizational learning and the people who can connect them.

If a company only has AI-generated code, it will soon discover that competitors can generate code too.If it has clear context, reliable evaluation, real customer feedback, stable workflows and people willing to own the result, then AI can become its asset instead of someone else's API.

References

• Stanford HAI, The 2025 AI Index Report
• Anthropic, Introducing the Anthropic Economic Index
• Simon Willison, Context engineering
• LangChain Blog, How and when to build multi-agent systems