WHAT WE DO
Four services, described with enough detail to make a decision.
ARA is a software and technology company that builds AI agents, custom software, GEO positioning and Business Intelligence systems for mid-size and large companies across Latin America and the United States. We use AI on two levels: inside the product we deliver, and inside our own development process. We work with Claude (Anthropic), GPT (OpenAI) and Gemini (Google), and pick the model that fits the task. This page covers each service, its typical scope, usual timelines and the signals that tell you it applies to your case.
Updated:
What is an AI agent, and when is it worth building one?
An AI agent is a software system that takes a task, queries your company's data and tools, runs the steps needed to complete it, and logs every action it took. The difference from a chatbot is execution: a chatbot answers, an agent does the work.
An agent is worth building when a repetitive process with describable rules is consuming hours from skilled people and relies on data that already lives in your systems. If the process is undocumented and changes every week, fix the process first: automating a confusing workflow multiplies the confusion.
- Support agents that resolve requests against your real data and hand off to a person with the context already assembled.
- Back-office agents that classify, extract and load information from documents, email and forms.
- Document analysis agents for contracts, policies, tenders or case files.
- Multi-agent orchestration: specialised agents with their own tools, coordinated by a flow with explicit rules.
- Automated evaluations and guardrails: test cases, action limits and a trace of every decision.
| Dimension | Usual reference |
|---|---|
| Pilot in limited production | 4 to 8 weeks from kickoff |
| Initial integrations | 2 to 5 systems (CRM, ERP, document store, email, support channel) |
| Success criterion | Defined before any code: resolution rate, cycle time, human error avoided |
| Oversight | Human in the loop for any action with financial or legal impact |
| Deliverable | Agent running in your infrastructure, evaluation suite, operating documentation |
Every agent we ship includes an evaluation suite: test cases with expected outputs that run on every change. Without evaluations there is no way to know whether a change to the prompt, the model or the tools made behaviour better or worse.
What does custom software development include?
Custom software development at ARA means building a complete product, from architecture to deployment, with AI assistance at every stage and mandatory senior human review of all code that reaches production.
The short answer: we deliver systems that run in your infrastructure and that your team can maintain. AI assistance compresses the time spent writing code and tests; it does not replace architecture or review. That is why timelines shrink without the standard dropping.
- Web platforms, internal portals and APIs.
- Legacy modernisation, migrated in stages, with the old and new systems running side by side.
- Integrations with ERP, CRM, existing databases and third-party services.
- Infrastructure, observability and deployment pipelines.
- Handover: documentation, pair programming with your team and post-launch support.
| Stage | Typical duration | Outcome |
|---|---|---|
| Diagnostic | 1 to 2 weeks | Scope, risks, proposed architecture and estimate |
| First usable version | 6 to 10 weeks | System in production with the core flow complete |
| Iteration | 2-week cycles | Increments shipped and measured with real users |
| Handover | Ongoing | Living documentation and your team writing code |
What is GEO positioning, and how is it measured?
GEO (Generative Engine Optimization) is the work of getting language models — ChatGPT, Claude, Gemini and Perplexity — to mention and cite your company when someone asks them a question in your industry.
Classic SEO optimises for a list of blue links. GEO optimises for a written answer that carries one to five sources. They are different targets and they coexist: the same well-structured content usually serves both, but the success criterion moves from ranking position to mention frequency.
- Visibility audit: a fixed set of industry questions is run against each model and we record whether you appear, from which source and in what context.
- Citable content: definitions at the start of each section, figures with context, explicit authorship and dates.
- Structured data: Organization, Person, Service, FAQPage, BlogPosting.
- Model-facing files: llms.txt and llms-full.txt, plus robots.txt with explicit rules for GPTBot, ClaudeBot, PerplexityBot and Google-Extended.
- Recurring measurement of mentions and answer share against the initial baseline.
| Criterion | SEO | GEO |
|---|---|---|
| Unit of result | Position in a list | Mention inside an answer |
| Primary signal | Links and domain authority | Clarity, structure and verifiability of the text |
| Winning format | Page optimised for a keyword | Blocks with definition, data and source |
| Measurement | Rankings and impressions | Citation frequency across a fixed question panel |
| Time to effect | Weeks to months | Weeks to months, depending on indexing and training cycles |
One honest caveat: nobody controls what a model answers. What you can do is publish verifiable, structured, easy-to-cite information and measure the effect with a question panel that stays stable over time.
What does a Business Intelligence project solve?
Business Intelligence is the discipline of turning scattered company data into a set of reliable metrics, with a single definition per metric, available to whoever has to decide.
The usual problem is not missing data, it is missing agreement: three teams calculate the same metric three different ways and the meeting is spent arguing about which number is right. The first deliverable of a BI project is a metric dictionary, not a dashboard.
- Source integration: ERP, CRM, spreadsheets, in-house systems and external services.
- Data model and metric dictionary with one definition per indicator.
- Operational and executive dashboards, focused on decisions rather than chart count.
- Automated alerts and reports on deviations, with agreed thresholds.
- A natural-language query layer on top of the validated model, when the case justifies it.
| Signal | What it indicates |
|---|---|
| The monthly report is assembled by hand in spreadsheets | Fixed hourly cost and copy-paste error risk |
| Two teams report different numbers for the same thing | No single metric definition |
| Decisions wait for month-end close | Data latency higher than operational latency |
| Nobody knows where a number comes from | No traceability from figure to source |
How does a project with ARA start?
Every project starts with a one to two week diagnostic in which we work inside your operation, with your team, until we understand the real problem and what it currently costs.
We do not hand over a PDF and leave. The diagnostic ends with three things: a written scope, a measurable success criterion and an effort estimate. If the work shows that AI is not needed, we say so: some problems are solved by an integration, a process change or a SQL query.
- Week 1: interviews with the people who run the process, plus a review of systems and data.
- Week 2: scope, success criterion, risks and a staged delivery plan.
- After that: two-week cycles with usable deliveries measured against the baseline.
What don't we do?
We don't take projects whose result cannot be measured, and we don't take projects that consist of adding a layer of AI on top of a process nobody can explain.
- We don't deliver demos that never reach production.
- We don't automate decisions with legal or financial impact without human oversight.
- We don't resell model licences: we pick the provider per task and the contract stays in your name.
- We don't work with sensitive data without a prior agreement on processing, retention and access.