How to integrate AI in an organization: process first

An order of work for bringing AI into a company: make the process explicit, name the risks, and widen the autonomy of agents as the structure supports it.

To integrate AI in an organization, start with the process: make explicit how the work is done, how evidence is left and how the real state of a project is known, and bring the agent in after that. Autonomy comes last, as a consequence of the structure around the agent, and not as the first decision.

What it means to integrate AI in an organization

Integration is more than handing people an assistant. In a framework I wrote in 2023, I described four ways in which a language model enters a company:

  • Productivity tools for employees. People use assistants to improve their own productivity.
  • A call to a model API. An application gains a function by calling a model.
  • Internal decision-making. A model sits behind internal applications that inform decisions.
  • Customer-facing applications. The input of a customer guides the prompt, and the response reaches the customer directly.

Choosing among them matters, and it is not the first step. The first step is the process the model will work inside.

Why the process comes before the autonomy of an agent

On one project, where I was one of a small group of people, the work did not start with the idea of giving an agent more autonomy. It started by making the processes explicit: how to leave evidence, how to connect a story to a change, how to review, how to follow up and how to understand the real state of the project.

Many of those practices already existed, distributed across people and tools. Making them consistent and observable reduced the ambiguity of the system.

That changed the relationship with the agents. At first they required a lot of supervision, specific instructions and a constant transfer of context. As the environment became more structured, an agent no longer had to rebuild everything from scratch. It could understand where the project stood, which rules to follow, which actions were valid and what evidence it had to leave behind.

Autonomy was a consequence of having strengthened the system around the agent.

How to make company information readable by an AI agent

The second change is in how information is treated. It stops being something a person looks up when needed. The repository, the tickets, the pull requests, the comments, the reports and the decisions are no longer isolated records. Together they form a living representation of the state of the project.

On top of that representation it is possible to build monitors, routines, reviews and agents, without somebody explaining the whole context again.

The next stage is a persistent environment that holds this operating knowledge: the rules, the context, the processes and the traceability. It is not meant to be "the main agent". Its job is continuity. The interface can change and the model can change, and the project keeps its structure.

The last stage is a common interface. People should not have to learn how the whole agentic system works on the inside in order to use it.

Third-party model or self-hosted model: how to decide

The 2023 framework separates two deployment paths: a third-party model, or an open-source model hosted in-house and trained with proprietary data. Its rules are general guidance, not the record of a deployment.

  • A third-party model can pose more privacy and data security risks, and its cost can escalate over time.
  • A self-hosted model can lead to higher security risks unless you invest in specialists in machine learning, security and privacy.
  • If a unique application of the model is vital to what differentiates the business, consider deploying and training a model in-house.

Where to start with AI in a company: decision rules

These rules are my general advice, drawn from the sources of this guide.

  • If nobody can describe how a piece of work is done today, write the process before you add a model.
  • If the state of the work lives in conversations and in memory, move it to records that a person and a program can both read.
  • If an agent needs the whole context explained on every task, fix the environment before you change the model.

How to integrate AI in an organization step by step

The steps are general advice. They follow the order of the essay Integrating AI in an Organization.

  1. Choose one process that people already run.
  2. Write how it works today: how the work arrives, how evidence is left, who reviews and how follow-up happens.
  3. Make those practices consistent and observable across people.
  4. Put the records where people and programs can read them.
  5. Name the use and the deployment type, and list the risks that apply to them.
  6. Bring in an agent with close supervision and specific instructions.
  7. Build monitors, routines and reviews on top of the records.
  8. Widen the autonomy of the agent as the environment gives it state, rules and valid actions.
  9. Consolidate the operating knowledge in a persistent environment.

Risks of using language models in a company

The 2023 framework lists six categories of risk.

  • Prompt injection. Crafted inputs manipulate the model into unauthorized actions or data exposure.
  • Data leakage. The model exposes sensitive information or proprietary details.
  • Training data poisoning. A model that learns from compromised text can misinform its users.
  • Denial of service. Malicious interaction degrades the service or causes high resource costs.
  • Insecure supply chain. Vulnerabilities in the supply chain of the model lead to bias, security breaches or system failures.
  • Overreliance on generated content. Too much dependence on output that can be wrong leads to misinformation.

The framework names the risks and does not prescribe a control for each one. For the controls on one kind of connection, read the guide on connecting a database securely using MCP.

What goes wrong when AI is added before the process

The first three items describe where that project started. The last one follows from the lean case study.

  • The agent rebuilds the context every time. It needs supervision, specific instructions and a constant transfer of context.
  • Practices differ from person to person. They are spread across people, tools and ways of working, and the agent inherits that ambiguity.
  • Operating knowledge has to be transferred again and again. Each person has to learn how the agentic system works on the inside before using it.
  • An average hides the weak layer. If four layers are strong and one is weak, the company pays for capacity it cannot use.

A worked example: a method that used to travel by meeting

LeanManufacturing.mx is the platform of a lean manufacturing practice based in Monterrey. Its method had existed for decades. A plant manager could not reach it at night, in Spanish, without booking a meeting first.

I built the platform that made it reachable. It shows the same dependency in a small system: the AI rests on written knowledge.

  • The knowledge is written down and open. A documentation corpus covers the method, and it is also the retrieval base of the assistant on the site.
  • People and agents read the same material. A public MCP server lets an AI agent query the documentation that a person reads.
  • The model can change. Generation runs through an OpenAI-compatible interface, so the answering model is a swap and not a rewrite.
  • An instrument says where to invest. A self-assessment asks fifteen questions from the shop floor across five layers: data, logic, interface, orchestration and feedback. The real capacity is the minimum across the five, not the average.

The same five layers appear in Structural Waste in Digital Operations, a paper in revision that is listed on the research page. Its abstract states three propositions: separation lowers waste, flow precedes automation, and capability is bounded by the least mature layer. The second is close to the order this guide follows.

What these sources show and what they do not

  • The essay on process and autonomy reports the order in which things happened on one project, as the work of a small group of people. It gives no measurement of how much supervision was reduced.
  • The persistent environment and the common interface are described there as stages being consolidated, not as finished products.
  • The framework dates from 2023 and does not document a deployment.
  • The paper is in revision and its manuscript is not public yet. According to its abstract, the weakest-layer proposition is proved as a band and shown by simulation to be separable from additive models. A pre-registered study of 500 repositories probed it, and its frozen criterion was not met. The abstract reads that result as a verdict on the proxy and not on the cross-layer claim.
  • The case study describes the public platform of a practice, not an adoption programme run inside a company. It gives no figures on usage or on business results.

Where to go next

Common questions

What is the first step to integrate AI in an organization?

Make one process explicit before choosing a model. Write down how the work arrives, how evidence is left, how a request connects to a change, how it is reviewed and how the real state is known. In the project where I learned this order, agents needed a lot of supervision at first, and that changed as those practices became consistent and observable.

Why should the process come before the autonomy of an agent?

An agent placed in an ambiguous environment has to rebuild the context on every task. When the state of the work, the rules and the valid actions are written where the agent can read them, it needs less supervision. In the project I describe, autonomy was a consequence of that structure and not the starting point.

Should a company use a third-party language model or host its own?

It depends on the use and on the risks the company can manage. A third-party model can pose more privacy and data security risks, and its cost can grow over time. A self-hosted open-source model needs specialists in machine learning, security and privacy to run safely. Consider hosting your own when a unique use of the model is what sets the business apart.

What are the main risks of using language models in a company?

The framework I wrote in 2023 lists six: prompt injection, data leakage, training data poisoning, denial of service, an insecure supply chain and overreliance on generated content. Not all of them weigh the same in every use. Start with the high-risk scenarios that apply to your own use case and deployment type.

How do I know which part of the operation to improve first?

Look for the weakest layer. The self-assessment I built for LeanManufacturing.mx rates five layers and takes the minimum across them, not the average, so the next investment goes to the weakest one. The paper that states this idea is in revision, and its pre-registered study did not meet its frozen criterion. Use the rule as a working instrument and not as a confirmed measurement.

Work and evidence

LeanManufacturing.mx (opens in a new tab)

My contribution: Built the platform of a lean manufacturing practice: an open documentation corpus, two working tools, a retrieval assistant grounded in that corpus and an MCP server.

What this supports: Putting AI on top of a documented method, so that people and agents read the same material.

What it does not establish: It is the public platform of a consultancy, not an adoption programme run inside a client company.

Source reviewed:

Technical guides