711 239 085 info@gilsys.com
es ca en
Abstract background

We integrate AI into applications to offer new features

Custom software with AI integration

We add artificial intelligence to existing software and develop new products that incorporate it.

Where does an AI integration project start?

When it is not clear where to start, AI consulting answers that, with an assessment and a roadmap.

The company's custom software

AI is added as one more feature of a custom software development project, for example an internal tool, in both web applications and mobile apps.

Software from another vendor

It can be integrated without necessarily modifying the existing software, by connecting tools that have APIs, such as Odoo, through custom e-commerce modules and in document management systems such as ownCloud.

A new product with AI at its core

When a new tool is created in which AI does not improve something that already exists but is the main function of the product, such as a SaaS product or an internal tool. For example, our public tender finder and our SEO monitoring tool.

Smart classification systems

Classifying products automatically when they are added

An example based on a real case that, in an e-commerce store (PrestaShop), finds the already classified products that are most similar to the one being added and suggests the ideal category.

Smart category suggestions

The system compares the new product with the ones already classified and proposes the most likely categories, together with the estimated match percentage.

Reorganizing the existing catalog

The system can review the database and suggest restructuring, such as moving certain products to another category or even changing the category structure.

Criteria tailored to each e-commerce store

The solution is integrated into each e-commerce store as a custom module, with the decision criteria defined with the client according to its catalog. With few categories, the AI decides which one fits best based on the product's name, description and attributes. If there are many categories, the system looks for similar products that are already classified and the AI decides based on them.

Automatic or supervised classification

The system can be set up so that every suggestion is always reviewed by a person, or so that products are always classified automatically.

The same concept could be applied to many other areas where classification is needed, such as documents, data from a unified product management system or any other information that needs to be organized and tagged.

Custom-built chatbot assistants

Custom AI chatbot for business

Connected to the data of internal or external applications, limited by user.

No predefined flow

A traditional chatbot has predefined answers and redirects to a form when the question was not anticipated. We can develop chatbots that look for the answer in manuals, procedures, the incident history and the application's own database, with restricted permissions.

Citable sources in every answer

Each answer can link to the documents or records used as context for it, so that the person asking can check it. If no related document is found, the chatbot says so instead of making up text.

Querying data and proposing actions

It is possible to answer questions such as "which case files have had no activity for more than a month?" with real data, and then apply actions to them, such as marking them as archived.

It can take the current screen into account

In help systems for tools, chatbots can be set up to answer with the application's own documentation, taking into account the information being displayed on the screen.

With permissions per user

Offering personalization and segmented permissions requires identifying the user first.

On the channels needed

In addition to the application itself, a chatbot can be integrated as a custom component on a website, an e-commerce store or an online platform, or connected to WhatsApp, Telegram or any other data input channel of the company.

Using a vector database

It can rely on preprocessed data stored in a vector database to make it easier to search for information by intent.

Sometimes it is hard to search using a single concept

Semantic search

We find results by intent, not by exact words.

A conventional search engine searches based on specific words or concepts. With AI, the search can be interpreted: a search for "problems with the invoice" can return matches such as "duplicate charge on the receipt".

There are different ways to implement semantic search: it depends on whether the database in use can be modified.

Without modifying the database: synonyms with AI

Before searching, the AI turns what has been typed into different related terms ("invoice", "receipt", "charge", "payment") and converts them into the filters that the application's search engine understands. This way, the database is not modified, but each search takes a little longer and has a small cost for the query to the AI model that calculates the synonyms (although precomputed results can be stored).

With a vector index: search by meaning

Each record is converted in advance into a vector that represents its meaning and stored in a vector database. When searching, the query is converted in the same way and the most similar records are retrieved, with no need to generate synonyms. This technique can be combined with traditional text search. With semantic search, the system responds quickly and with no cost per search, but the index has to be created and all the records processed. We have implemented this solution in, for example, document management systems.

Relating data and detecting duplicates by their meaning

If the meaning of each product record, user profile or document in the application is stored, the tool can use it to offer new AI features.

Relating and matching

With the information classified by meaning, it would be possible, for example, to search a network of contacts for profiles that can carry out a given task, or to find a supplier's products based on their function.

Detecting duplicates

Another interesting feature is detecting records that say the same thing in other words, even if they are not identical: two records for the same product, or two contacts who are the same person but whose data are not exactly the same.

Adjusting criteria without changing code

With AI we can interact with data naturally, without having to add every adjustable parameter as a configuration option.

Without AI: programmed rules for each condition

With traditional rule-based logic, every new condition the user needs to configure requires custom development. If a tool has to filter by category, sort by price and apply conditional styles (such as highlighting offers or private-label products), specific panels have to be programmed to configure how each option works.

The accumulation of special cases drives up what we call technical debt, making the tool more complicated and expensive to develop and maintain.

With AI: the criteria explained

With AI we can configure tools so that the criteria are expressed naturally, as text: "Seasonal items go first, items low in stock are not shown on the home page, and own-brand items come before resale items." Written like that, in one sentence, by someone who knows the business but not how to program. To change the criteria, the explanation is changed, not the code.

Required is not the same as correct

Smart validation with AI

An efficient input data validation process helps maintain the quality of the stored information.

With programming: the format

Strict validation can check that a field is not empty, that it is numeric, that it has the format of an email address, a tax ID (such as the Spanish NIF) or a date, or that it is within a range. None of these checks prevents the user from typing "no description" as text. The field is in fact filled in, and the record would be stored and processed without any problem being detected.

With AI: the meaning

With AI we can understand the meaning of what has been entered, so we can validate data logically, for example, that a description really describes the product, that one field does not contradict another in the same record or that the text is in the right language. In addition, by integrating it with custom development, we can decide what to do: block, warn or allow.

Cleaning data and keeping it clean is considered a fundamental process for data governance.

More AI features that can be integrated into applications

AI handles the points where a tool used to need a person to intervene.

Writing, summaries and reports

Descriptions based on the data in a record; automatic drafts for replying to requests, reviewed by a person; summaries of documents and meetings; plain-language business intelligence reports based on the system's data.

Detection of errors and unusual behavior

Detection of inconsistent data or unusual behavior, by comparison with the stored data. For example, an order with a quantity far higher than usual for a given product.

Enrichment and normalization

Completing missing attributes in product records, standardizing contact addresses or normalizing supplier catalogs that come in different formats.

Automatic prioritization

Tasks and incidents ranked by urgency and impact, by analyzing their description and estimating times based on previously recorded experience. For example, an incident that prevents invoicing takes priority over answering an informal query.

Image generation

Product images, functional mockups or illustrations created from a description or a reference photo.

Transcription and translation

Voice notes, calls or meetings converted into text, and content translated into other languages. For example, a technician could describe from a mobile phone the work done for a customer and have it recorded in writing in the customer's file.

Feedback analysis

Automatic classification and sentiment identification, based on reviews and messages. For example, separating complaints about shipping from complaints about the product in an online store's reviews.

Prediction and recommendations

Analyzing content based on criteria defined in natural language, generating purchase forecasts to anticipate future demand from historical data, or personalizing product recommendations for each user on e-commerce platforms.

Data extraction from documents is covered in automation, and image recognition in AI for manufacturing.

Even without access to the source code

How AI is integrated into an existing application

We have been developing custom software since 2004. AI is one more tool that we integrate with that experience and that broadens what each project can offer.

We adapt the source code

If the source code is available, AI can be integrated at code level, as new functions in the application that are used when needed.

We connect directly to the database

If the code is not available, AI can work directly with the information stored in the application's database, reading and even, if permissions are granted, writing information. For example, reading a product catalog and offering reclassification options, running data quality checks or even answering questions about the stored information, as a chatbot.

We connect through an API

If the application offers an API, we connect to it without modifying the code or accessing the database. It is the safest option with software from another vendor, and the one we usually use in system integration.

Only when the wait is worth it

AI is used for tasks that would take the user time and where waiting a few seconds pays off. Not all tasks performed by an AI model take the same time: a generative AI writes the answer and needs a few seconds, while a decision AI such as Jev, which only classifies, responds much faster.

Interchangeable models, commercial or private

The applications we develop do not connect directly to a specific model but to an intermediate platform, usually LiteLLM for text and reasoning models and Fal.ai for image generation. Switching models is a configuration change, not a programming one. In each project, the model that gives the best results is chosen, whether from Anthropic (Claude), OpenAI (GPT), Google (Gemini) or another provider. In addition, if the information is sensitive, we can work with open-source models on private servers, in a private AI.

Logging and auditing of every operation

It is possible to develop tools in which every AI action is logged and can be audited: what it was asked, what it answered, what data was used and which person validated the result. For processes such as automations or agents, it is very important to have an operational log to determine the origin and details of each action.

We explain different options for working with AI models in which AI model is used in each project.

What we take into account before developing them

AI-based SaaS and internal tools

When AI is the main function of the product, there are decisions that do not come up in a conventional application and that determine its cost and its viability.

The cost of each use, in the business model

Every generated image or processed text has a cost in the model. It is important to calculate it from the start and reflect it in the estimate of usage costs, so that user growth does not make the product run at a loss.

Which model and which provider

Different models are tested and the one that offers the required quality at the lowest cost is chosen, for text or for images. It is advisable to check periodically whether there are models that fit better and to run the corresponding quality tests.

Quality measured with real cases

Before launching the service, it must always be validated with real data, with a limited group of customers, to determine that the result is acceptable. A demo that works with a few specific cases does not guarantee that it will work with all of them.

Limits and abuse control

Controls must be set up for usage limits per user, content filters and request logging, to control the cost and prevent misuse.

Transparency and the EU AI Act

Depending on the type of application, the AI Act requires informing users that they are dealing with an AI or that a piece of content has been generated by an AI. This must be considered at the start of the project, not in later phases.

We recommend working in phases

In large projects focused on offering services with AI as their main feature, it is usually necessary to define a minimum viable product with a fixed scope and price, launch it and decide the next phase with real usage data.

Frequently asked questions about software with AI integration

Does a model have to be trained with the application's data?

No. In custom-developed tools, the model connects to the application's data and queries it at the moment of answering: if a record changes, future answers will change automatically. Sometimes vector databases are prepared with company information to allow natural language queries, but this does not mean training the model, only giving it context.

Can AI be integrated into Odoo, PrestaShop or another ERP or CRM?

Yes, as long as the software to be integrated has an API or database access, as most current ERP, CRM and online store systems do. There is no need to switch platforms: we integrate AI on top of the existing tools, adapting the solution to each case. For example, in Odoo we can connect an MCP server to operate it directly, in PrestaShop we develop custom modules (such as the automatic catalog classification one) and in a CRM we can connect to its API to update customer records or add sales opportunities.

Who owns the source code of custom software with AI?

The client. Once the project is finished, the source code can be provided to the client on request. We only reserve the use of generic, library-type components that are not part of the specific business logic developed for the project.

What maintenance does AI integrated into an application need?

The same as the rest of the software: automated tests and continuous integration in every new version. In addition, a set of real cases that is run again every time the AI model changes, like the one we describe for a private AI.

How much does it cost to integrate AI into an application?

There are two costs: development, which is quoted at the start, and use of the model, which with a commercial model is paid per query. In an existing application, it depends mainly on how the data is accessed (API, database), on the features to be implemented and on the interface. The cost of a tool is not easy to estimate; everything that affects it is explained in this section, but we can provide a free, no-obligation estimate.
Abstract background

We develop tools and services based on artificial intelligence

We create solutions from scratch or connect AI to existing software to turn data into decisions, automations and new features.