Internal tools with AI means adding a language model to the systems a company already runs (the dashboard, the CRM, the spreadsheet, the back-office application) so that people can ask instead of search and approve instead of retype. A company needs it when its data lives in tools that work, but getting an answer out of them takes a person who knows where to click, and when its repeated tasks are held together by scripts that nobody wants to touch.
What I build
- Assistants over company data. A chat inside the tool, with the context of the screen, suggested actions and alerts, that answers from the records of the company.
- Questions in plain language. A connection between an AI assistant and a database or a base of tables, so that a person asks for the overdue tasks or the last order of a customer and gets the answer.
- Analysis that explains itself. Answers that arrive with the queries, the checks and the figures behind them.
- Automations with approval. Flows where the model proposes and a person confirms the steps that change a record, with a trial mode before production.
- One layer in place of scattered scripts. A single service for synchronization, with change detection, retries, logs and a history of the data.
- A safe door to the model. One service that holds the keys and that every tool calls, so that no key sits in a script or a browser.
Symptoms in the tools you already have
- People leave their work to look something up in another tab, and lose the thread.
- The same report is rebuilt by hand from exports, and reconciling it takes hours.
- A sync fails at night and the company finds out from a customer.
- A key for a model is pasted into a spreadsheet script or a front end.
- A dashboard shows a number and nobody can say which records produced it.
Each symptom points to a different piece of work. The first asks for an assistant. The second and the last are data work. The third is integration. The fourth is security, and it comes before all the others.
How the work runs
- Inventory. I list the tools, who uses each one, what data it holds and how it can be reached: an interface, an export or a database.
- Put the data in order. If records are scattered, or lost between syncs, they go first into one database with a model that keeps history. An assistant over inconsistent data repeats the inconsistency with confidence.
- Define the terms once. What margin, overdue or active mean in this company, written where both people and the model read it.
- Read before write. The first version only answers. Actions that change records come next, each one behind an approval, a flag that the business can turn off, or a trial mode.
- Watch and correct. Corrections made by people go back into the rules, and failures show in one place.
What you receive
The code in your repositories, the service that holds the keys in your cloud account, the definitions of your business terms as a document, and the controls: flags, trial mode and logs.
What this does not include
I do not replace a system that works to make room for AI, and I do not build a chat window with no data behind it. I do not let a model write to your records without a step where a person can stop it. If the data is not in order, that is the first project, and I say so before starting.
Where the evidence is
For Talisman, an accounting platform from Texas, I built a project management module inside the product as an AI engineering contractor. That included the full experience of its assistant: streaming chat, suggested actions and alerts, staged for the connection to a model. For a company in the warranty management industry, I moved scattered records into one relational database and built the dashboard on top. I turned an inherited integration between a property management system and a field service system from a bridge that kept no history into a data warehouse with history, duplicate detection and flags that the business controls. And Airtable MCP, which is open source, lets a person talk to an Airtable base in everyday language through an AI assistant.