Only 11% of AI Agents Reach Production. Here’s What the Other 89% Are Missing.

After the AI coding boom, agentic apps multiplied. Connect a model, mount a knowledge base, add a few tools, and you can build an agent that retrieves, analyzes, and writes. Demos get smoother. Launches get louder. Startups multiply. It looks like spring.

But apps that enter core production have not grown at the same speed. Going from “complete a task once” to “handle real business every day” meets heavy resistance. Deloitte’s Tech Trends 2026 puts production-ready agents at 11%.

That number is not a verdict on models. It is a verdict on the gap between model capability and daily business operation.

At Huawei Connect 2026, Tao Jingwen, vice chairman of Huawei’s Supervisory Board, named the problem:

“AI intelligence advances faster than humans and industries adapt to it. We no longer lack powerful intelligence. We lack methods for people, industries, employees, and organizations to adapt to powerful intelligence.”

Three questions follow:

1. Which business scenarios deserve redesign around AI, not just an assistant?

2. How do we rebuild a knowledge system that was built around people so AI understands the business?

3. How do organizations, talent, and processes adapt to match the productivity of powerful intelligence?

Tao says enterprises must rethink how they organize work. Past processes were designed around people. With AI, they must redefine what goes to agents, what stays with professionals, and what requires joint decisions. He calls it “redesigning a car.” Technical teams are moving from “connect models to processes” to designing processes around human-AI collaboration, building a production system where both can operate together.

Competition in enterprise AI is shifting from model capability to systems engineering.

1. Quantity Is Not Quality

The agent field has no shortage of quantity.

New frameworks appear every month. Agents pop up fast. A team can build one that chats, retrieves, calls tools, and writes reports. It looks like the future, and it wins applause in demos.

Production is not a demo stage.

In a demo, an agent completes a task once. In production, it handles real business every day. Demos can ignore permissions, timeliness, exceptions, auditing, and cost. Production cannot. Those decide success or failure.

An agent that retrieves documents may not know they are outdated. One that analyzes problems may not understand business definitions. One that writes reports may not be accountable for results. Enterprises need a system embedded in processes, taking on tasks, accepting governance, and running continuously, not a clever assistant.

That gap explains why quantity is not quality.

Agent count can grow fast. Enterprise intelligence cannot be measured by agent count. Locally strong agents may have different information, permissions, and goals. They work in small scenarios. Local optima do not assemble into a global optimum. Without coordination, work duplicates or conflicts.

That is the resistance.

It lives in data definitions, knowledge representation, process boundaries, organizational collaboration, permission governance, and failure recovery.

The next stage of enterprise AI has a more basic question: how do we organize intelligence around shared business goals?

2. DIMAK: Five Layers of Dependency

To build that system, start with what enterprise intelligence is made of.

Huawei’s DIMAK framework names five: Data, Infra, Model, Agent, Knowledge. These are five words and a set of dependencies.

Data: Can the System Get Reliable Business Facts?

Data determines whether the system gets reliable business facts.

Putting files into a knowledge base is the starting point. The hard part: Are data definitions consistent across systems? Does update frequency meet business needs? Is quality trustworthy? Are permissions clear?

Take “inventory.” Warehousing says “in stock.” Sales says “available for sale.” Finance says “book inventory.” The model gets three numbers. Which one does it pick? Without context, it picks one and keeps going wrong.

Data means information can be accurately identified and used in specific business contexts. The hard part is judging whether sources are trustworthy, timeliness has not expired, and permissions are accessible. These quality indicators affect the result. When critical information is missing, the model can start down the wrong path.

Knowledge: Business Rules as Executable Code

Knowledge determines how business rules are understood and executed.

The system knows an order amount. That is data. Understanding which approval rule applies to that amount is knowledge. Deciding which process the task enters and which tool to call is agent execution.

What experts call “usually handled this way, but special cases need special procedures” must be broken into conditions, exceptions, and takeover mechanisms. Some rules belong in retrieval and explanation. Some belong in processes or tools.

You cannot hand the full policy manual to a model and expect reliable rule execution. Does knowledge have a version? What if rules conflict? How does it adapt after a change? These problems need engineering. There is no shortcut.

Treat knowledge as executable business rules, not documents.

Model: Fit the Scenario, Not the Hype

Model handles understanding and reasoning. Selection and adaptation must answer scenario questions: Is domain understanding enough? Is tool calling reliable? Does deployment meet requirements? Is the performance gain worth the investment?

Current models cannot automatically evolve the best choice for every step. For clear, bounded tasks, a lighter model or a deterministic program can do the job. Answer these questions well, and model capability serves business goals. Otherwise, you do technology for technology’s sake and deploy a model just to have an “AI.”

Open-source models keep expanding. This year they reached 5T parameters, with version lifecycles of 3 to 6 months. Infra must openly support mainstream models so the industry can use them quickly and create value. Model toolchains and engineering practices must make training feasible, effective, and fast, and make leading models deliver inference performance.

Treat infrastructure as a production tool that creates value.

Agent: Turning Capability into Business Action

Agent determines how capability becomes business action.

Agents organize tasks, call tools, record state, and handle exceptions. They need to know where information comes from, what can be called, which actions need authorization, and where results get written back.

Enterprises also need accuracy, reliability, security, and efficiency. For low-risk scenarios like customer service and marketing, the system can suggest. For critical steps, set up permission management, manual review, exception takeover, and full traceability.

A technical solution must make AI run stably in a real environment and be locatable and recoverable when problems occur.

Infra: Can Intelligence Be Supplied Continuously?

Infra and Model determine whether intelligence can be supplied continuously in a way the enterprise accepts.

Infrastructure supports compute, storage, and networking. It must handle concurrency, long tasks, and load changes. Local reliability does not automatically mean tolerance for real-world noise. Evaluate stable operation across multi-turn reasoning, deep retrieval, and tool calling.

As more steps join, response time, resource use, and failure recovery affect experience, cost, and efficiency.

Treat infrastructure as a production tool that creates value.

Pearls and a Necklace

Tao compared discrete agents to pearls. They need to be organized around business processes.

That metaphor matches reality. Many locally strong agents have different information, permissions, and goals. Local optima do not assemble into a global optimum. Without coordination, work duplicates or conflicts.

Enterprise intelligence is not measured by agent count. It is measured by whether agents collaborate around shared business goals.

DIMAK maps enterprise AI dependencies: data provides facts, knowledge clarifies meaning and rules, models reason, agents act, and infrastructure runs underneath.

The model is the engine. DIMAK is the chassis. The Seven-Step Method is the roadmap.

3. The Seven-Step Method: Push Scenarios into Production

DIMAK describes needed capabilities. The Seven-Step Method describes the path into the business.

Huawei’s seven steps: Business Insight, Scenario Selection, Technical Solution, System Integration, Business Diffusion, Ecosystem Development, Continuous Evolution.

The names are familiar. The endpoint is production and continuous evolution, not a technical proof of concept.

Many AI projects fail at initiation. The team picks a model or builds several agents, then looks for requirements that show off the tech. That produces technically striking applications that may not solve the most important business problems.

The Seven-Step Method puts business logic and scenario selection first. The project must explain what the business is stuck on and which result should change after AI enters.

First: Business Insight and Scenario Selection, Where Should AI Go?

Do not start with “We have a model. What else can it do?” Start with real contradictions in production and operations.

Huawei works with customers to find high-frequency, complex, high-value problems. They judge business value, technical feasibility, and scaling potential. They prioritize scenarios worth investment and ready for implementation.

A good scenario has clear business goals, stable data sources, clear boundaries, and measurable metrics.

In manufacturing, equipment failure prediction, quality inspection, and production scheduling are all possible. But an enterprise cannot launch all three just because they are technically possible. It must ask: Which step most affects capacity and cost? Which has enough data? Which results will frontline employees accept?

Turn “what the model can do” into “what the business most needs to solve.” Then AI projects move past the demo.

Second: Technical Solution and System Integration, How Does AI Enter Existing Work?

If agents are to participate, connecting a model and a knowledge base is not enough. Connect existing data, knowledge, applications, and production systems.

The agent must know where information comes from, what can be called, which actions need authorization, and where results go back.

Enterprises need accuracy, reliability, security, and efficiency. For low-risk scenarios, the system can suggest. For critical steps, set up permission management, manual review, exception takeover, and full traceability.

A technical solution must make AI run stably in a real environment and be locatable and recoverable when problems occur.

Enterprise AI projects often require data governance, knowledge engineering, interface changes, compute configuration, and process redesign at the same time. Replacing a model does not solve a production system.

Third: Business Diffusion, Ecosystem Development, Continuous Evolution, From One Scenario to Many

After a scenario is validated, the valuable output is not just the project. It includes reusable data interfaces, knowledge specifications, agent templates, evaluation methods, and governance mechanisms.

The enterprise can extend validated capabilities along the value stream. It can also work with partners to turn single-point experience into industry solutions.

Diffusion is not simple copying. Different departments and industries have different data definitions, business rules, and responsibility boundaries. AI systems must collect feedback, optimize models, knowledge, and processes, and adjust to business changes.

Through iteration, an enterprise moves from an agent that solves one problem to system intelligence that covers multiple steps and works collaboratively.

That is the point of the Seven-Step Method: gain business insight, select the right scenario, complete technical work and system deployment, spread along the value chain, and turn a pilot into long-term production capability through ecosystem cooperation and continuous evolution.

4. Developers’ New Homework

Enterprises are asking more from AI technical teams.

Prompt design, model API calls, and agentic frameworks matter. They now sit inside a larger engineering chain.

For new agentic AI, developers need to understand business goals. They must organize data and knowledge, design task flows, define what agents can do, and answer for results.

Beyond model intelligence, they must handle task state, role handoffs, failure recovery, and human intervention. These decide whether business runs continuously.

Without mature solutions to these problems, AI cannot connect to real enterprise scenarios.

Enterprises care about what the connected AI system completed, what its basis was, and whether results withstand inspection and produce value.

A complete technical architecture can still stall when organizational collaboration lags.

Tao’s emphasis on industry-vertical organizations and short-chain operations responds to this need.

At the conference, “New Methods, New Organization, New Tools” answered three questions:

• The Seven-Step Method answers how to advance.

• DIMAK answers what engineering capabilities are needed.

• The new organization answers who collaborates, how, and who is accountable for results.

For agent developers, the next direction is clear: understand the business deeply, build maintainable agent systems, and prove value through real operation.

5. Organizations’ New Homework

Enterprise AI is not a solo performance by the technical department.

It needs business teams, data teams, knowledge experts, process owners, security and compliance, frontline employees, and external partners. Processes built around people must be redesigned around human-AI collaboration.

What goes to agents? What stays with professionals? What requires joint decisions? There are no standard answers. Every enterprise must answer for itself.

If the organization stays in traditional project silos, data team handles data, model team handles models, business team handles requirements, security team handles compliance, agents will struggle to enter production. Production systems need end-to-end accountability, not each group managing its own segment.

Competition in enterprise intelligence is extending from model capability to organizational capability.

Models can be bought. Tools can be integrated. Frameworks can be learned. The ability to organize data, knowledge, models, agents, and infrastructure around business processes must be grown inside the enterprise.

Conclusion

As AI takes on real tasks, competition extends from model capability to the enterprise’s ability to organize intelligence.

Investment effectiveness comes down to whether intelligence becomes trustworthy action and continuously improves the business system. It does not come down to model size, agent count, or demo spectacle.

Enterprises need an accountable system more than a chatty model.

Demos pursue amazement. Production pursues stability.

Locally optimal agents cannot assemble a globally optimal business.

AI implementation means letting processes regrow around humans and AI, with models connected along the way.

From the agent surge to production deployment, enterprise AI is crossing the last mile. That mile lives in systems engineering, knowledge engineering, and organizational engineering.

Whoever completes this restructuring first turns intelligence into productivity.