AI integration for business: chatbots and AI agents

ALODEV AIO is the internal operations system that ALODEV wrote and uses every day. Its clearest AI feature builds an acceptance record from a contract: a manager pastes the contract text and starts the analysis, and the server sends the text to a language model that returns a two-level list of deliverables. The client can then agree or request changes on the client portal.

These design decisions are what we bring to client projects:

• The instruction to the model asks only for what the contract actually states, forbids adding items, and returns an empty list when the contract states no scope.
• The model is forced to return JSON in a fixed structure, with temperature set to 0 so results are stable between runs.
• The model name is stored with each record, so after a model change it is still known which model produced which record.
• If there is no API key or the AI fails, the software does not stop: the user is moved to preparing the record by hand.
• A human decides: a record locks only after both sides agree, and a client's change request becomes a change ticket.

On 13 September 2026 we found that the API key behind this feature had been disabled on the live system. The code was correct and the error message was clear, but the error only appears when someone clicks the button. If nobody clicks, nobody knows.

The fix is a checker that runs on a schedule and does not wait for users. It sends two sample contracts to the running process:

1. Positive control: a contract with a clear scope must return at least one item. If not, the AI is broken.
2. Negative control: text with no scope must return exactly zero items. If items appear, the AI is inventing.
3. If the process cannot be reached, the checker reports it as not measured, never as passed.

ALODEV's software is written by ALODEV's engineers, and an engineer is accountable for every change. A chatbot or an AI agent is software like any other: someone reads it, someone reviews it, and someone answers for it when it runs on real data. For clients this means every change passes engineer code review, automatic tests and a test build before it reaches real users.

The work below uses the same building blocks already running in ALODEV AIO: calling a model through an API, forcing structured output, a human approver and a checker. Anything that needs new technology, for example semantic search over a large document store, is stated up front.

Reading and extracting data from documents

Contracts, invoices, purchase orders, candidate resumes and customer emails: the AI reads and fills the right fields in the software, and the data-entry person only approves. We recommend starting here because it can be measured immediately: what share of fields are extracted correctly, and how much time is saved per document.

Chatbots over internal documents

Staff or customers ask questions and the chatbot answers from the company's procedures, policies and product documents, citing the passage it used. It says so when the documents do not contain the answer, and it is checked on a schedule with questions whose answers are known.

AI agents for repeatable processes

An agent lets a model call tools step by step: look up an order, draft a reply, create a ticket, update a status. We design agents with a short list of tools, minimum permissions, a log of every step, and a stop for human approval before any action that cannot be undone, such as sending to a customer, moving money or deleting data.

• The rule is clear and fixed, such as pay by formula or approval by limit: rules written in code run faster, cost less and always give the same result.
• A wrong result has legal or financial consequences and there is no human approver: add an approval step, or do not use AI.
• There is no data to measure right and wrong: until it can be measured, nobody knows whether AI helps.
• Data may not leave the company and there is no budget for running a model on your own servers.

We agree the metric before we build, for example the share of fields extracted correctly, the share of chatbot questions answered with a cited source, or the time to process a file before and after.