Enterprise AI is the work of putting language models, and the systems around them, inside the processes a company already runs: its data, its internal tools, its customer channels and its reviews. A company needs it when people are already using AI on their own, each in a different way, and nobody can say what it reads, what it may change or whether its answers are right.
What I build
- The process, written down. How the work arrives, who decides, what evidence is left and what counts as done.
- Data platforms. Pipelines that bring scattered records into one database, with a model that people can query and the dashboards on top of it.
- AI inside internal tools. Assistants and model calls placed in the systems people open every day, in place of one more window.
- Agents, where they fit. For work with many variations that ends in an action in a system, with a clear list of what the agent may do.
- Evaluation. Cases written as code, checks that do not depend on a model, and a gate that stops a release that got worse.
- Integration and operation. Permissions, secrets, deadlines, queues for long work and a record that explains what happened.
Where AI enters a company
AI reaches a company through four doors, and each one asks a different question:
- Tools for employees. Assistants for writing, code and search. The question is what data people paste into them.
- A model call inside an application. Classify a message, extract fields from a document, draft a reply. The question is what happens when the output is wrong.
- Internal decisions. Analysis over company data. The question is whether each figure can be traced to its source.
- Systems that face customers. Answers and actions in front of a customer. The question is what the system may never say or do.
I start at the door where a mistake is cheap and visible, usually repetitive work with clear rules, and move toward the expensive ones when the tests and the records are in place. Not every step needs the largest model either: a fast, cheap model can read and sort, and a stronger one can write the final result from that summary.
How the work runs
- Describe one process. I read how the work is done today and write it down with the people who do it: inputs, decisions, exceptions and evidence.
- Make the state readable. Records, tickets, documents and decisions go where a person and a program can both read them. A system that has to guess the state of the work will guess.
- Name the risks and set the limits. For each use: what data it reads, what it may change, where the keys live and what leaves the company.
- Build the smallest useful piece. A pipeline, a tool or a narrow agent, on real data, with the cases that say what correct means.
- Gate, release and widen. The cases run before every release. The piece starts with a share of the work, and its autonomy grows as the record shows that it holds.
Autonomy is the last step, not the first. In the team where I learned this order, agents needed constant supervision until the process around them was explicit.
What you receive
The written process, the code in your repositories, the pipelines and tools in your cloud account, the cases and the harness that runs them, and notes for the people who will operate it. The knowledge of how to run the system stays in the system, where people and agents can both read it.
What this does not include
I do not start from a model or a vendor and look for a use afterwards, and I do not deliver a strategy document with nothing running at the end of it. I do not promise that AI replaces a department. If the honest answer is that a clear procedure and a clean table solve the problem, that is the answer you get.
Where the evidence is
Two public cases stand behind this page. For a higher-education group that serves 700,000+ students, I built the evaluation harness and the gates that decide whether a change to its production agents may ship. For a company in the warranty management industry, I built the pipeline that moved its records from MongoDB to PostgreSQL and the dashboard its teams use to follow agreements, claims and dealer performance. Each case says what it shows and what it does not.