Six Interfaces Need Sufficient Closure
Artificial intelligence is advancing quickly, but “AI can do this” and “this has become productivity” are not the same thing.
Over the past two years, more people have begun using AI to write, research, program, and analyze, while companies have introduced it into customer service, contract work, software development, and internal management. Yet the same AI capability can create value quickly in one setting and remain stuck in pilots in another. The question is no longer only whether the model is strong enough. It is how AI capability enters real production.
Here, “capability” does not mean education, professional credentials, or a stock of knowledge. It means the actual ability of an actor to perform a particular cognitive or practical function.
In practice, for AI capability to become productivity, six interfaces need to reach sufficient closure: capability, verification, data, workflow, responsibility, and economics.
The first layer is capability and verification.
AI has a feature that distinguishes it from traditional tools: being generally capable of a class of tasks does not mean that every individual output will be reliable. Capability therefore has to be paired with verification.
Code can be tested. Customer service can be judged by whether the customer’s problem was resolved. An author can directly judge an article. A recommendation system can receive feedback through clicks, purchases, and retention. Different tasks require different forms of verification. What matters is not whether there is always a single correct answer, but whether there is an affordable feedback mechanism that corresponds to the real objective.
Capability and verification therefore form the first layer of closure: AI must not only be able to do the task; there must also be a way to know when its result is good enough to use.
The second layer is data and workflow.
Data closure does not mean giving AI access to an entire corporate database. It means giving the system enough real-world information and context to complete the task. If someone asks AI to revise an article, the article and the necessary background may be enough. A complex business process may require internal databases, customer records, permissions, and real-time status information.
Workflow addresses the other side of the problem: once AI has produced a result, can that result enter a clearly defined next step?
That next step does not have to be automated. Many tasks properly require human judgment at critical points. After AI completes a feasibility assessment, for example, the next step may be a human decision on whether the project should proceed. If the AI result enters an established decision and action process, and it is clear who decides, how the result is handled, and what happens when an exception occurs, the workflow can already be considered closed.
Workflow closure, then, is not the same as automation. It means that AI output is taken up by the real production process and moves into the next judgment or action.
The third layer is responsibility and economics.
As AI moves from providing answers toward taking action, responsibility becomes more important. Who is authorized to let AI do what? Where must it stop? When must a case be handed to a person? Who bears the consequences of the final result? Responsibility closure does not mean that every step requires human approval. Low-risk tasks can run automatically, but authority and consequences still need clear ownership.
Economic closure should not be reduced to the question of whether something makes money.
Productivity is not only about producing more physical goods. Articles, software, research, design, knowledge, judgment, and services are also outputs. When we spend time studying a question, the result may be more than the article as an intellectual product. It may also include a new judgment, a lower probability of future error, and a method that can be reused.
Economic closure therefore asks a broader question: for the object being observed, do the AI, human effort, time, capital, and risk invested ultimately produce net value sufficient to justify those inputs?
It is also useful to distinguish value creation, value realization, and value capture. If AI saves an employee an hour each day, value has been created. If the company’s output does not increase, that value has not appeared as higher company productivity. If the employee gains an extra hour of free time, the employee may have captured the value. The fact that the company did not sell more products does not mean AI created no value; it means the value was not realized in the particular object being measured.
For clarity, the six interfaces can be grouped into three forms of closure:
Task closure = capability + verification.
Production closure = data + workflow.
System closure = responsibility + economics.
These are not six gates that must be passed in sequence, nor does every interface have to reach 100 percent closure. Different tasks require different degrees of closure. Drafting an ordinary email and allowing AI to execute a financial transaction plainly demand different levels of verification, responsibility, and risk control.
Several paths by which AI capability enters production
The first and simplest path is direct entry into individual capability.
Why have so many ChatGPT applications spread so quickly? Because one person can often close several interfaces at once. The user defines the task, supplies the context, judges the result, decides whether to use it, bears the consequences, and decides whether the result was worth the effort.
Writing, summarizing, translation, information retrieval, and research assistance spread easily not only because AI can already perform them, but also because the costs of verification and use are low.
The second path is to embed AI in a mature production process.
Customer service, programming, document processing, and contract work often already have relatively clear tasks, data, workflows, and responsibilities. AI may create substantial value simply by entering one costly step in that process.
This kind of local innovation is not inferior innovation. If the real bottleneck is local, improving that step is the sensible solution. There is no reason to redesign an entire company merely because AI is available.
The third path is to expand the range of capabilities an individual or a role can call upon.
Some work that once had to be handed to another specialist can now be partly completed with AI. A salesperson can perform some data analysis; an ordinary user can write simple programs that were previously beyond reach.
But there is an important boundary. The fact that AI can do something does not mean the user has acquired the full professional capability. If the user cannot understand the result or detect errors and limits, the capability has not yet entered that person’s effective capability in a reliable way.
The fourth path is to reorganize workflow after the bottleneck moves.
Suppose AI cuts the time required to draft a report from four hours to thirty minutes, but obtaining the data still takes a day and approval still takes two. Cutting the drafting time from thirty minutes to ten adds little. Once the original bottleneck disappears, new bottlenecks appear in waiting, approval, verification, handoffs, and exception handling. At that point, the problem is no longer whether the model is strong enough; it is how the work should be recombined.
Local innovation improves the current bottleneck. Structural innovation becomes necessary only after the bottleneck has moved.
The fifth path is to redesign the system when necessary.
A further step is warranted only when task boundaries, departmental boundaries, software structures, approval mechanisms, and responsibility arrangements themselves have become obstacles to using AI effectively.
Even then, a shifted bottleneck does not automatically justify immediate restructuring. The bottleneck must be large enough and persistent enough to matter, and the gains must be weighed against restructuring costs, transition risks, and the speed of technological change. If the technology may improve substantially within six months, limited experimentation and preserving future options may be more rational than committing large sums to redesign the system today.
The best AI innovation is therefore not the most complex or the most systemic. It is the one that solves the real bottleneck at the lowest total cost.
Why individual AI applications spread faster than enterprise AI applications
The six-interface framework gives a simple explanation.
When an individual uses AI, the task, data, verification, workflow, responsibility, and value judgment are often concentrated in the same person.
A company is different. Data may belong to one department, the workflow to another, permissions to a third, risk to a fourth, system interfaces to IT, and budget authority to someone else.
The model capability may be identical, but an individual can close several interfaces personally, while an enterprise has to engineer and institutionalize those interfaces.
That is why slow enterprise adoption does not necessarily mean that AI capability is inadequate, or that companies do not want to use it. Often, the time-consuming part is connecting AI capability to the real production system.
AI is making cognitive capability increasingly easy to call from outside the individual. Capability itself is becoming cheaper, but productivity does not appear automatically.
An AI capability truly enters production only when capability, verification, data, workflow, responsibility, and economics reach sufficient closure.
The resulting productivity may take the form of more physical goods, but it can also appear as better software, higher-quality services, stronger writing, more reliable judgment, and reusable knowledge.
This is easy to miss when discussing the economic value of AI. Stronger models are only the beginning. What ultimately determines how much value these capabilities create is whether they can connect with human work and real production systems.
