What an AI automation project includes
We shape the engagement around the actual brief. The proposal identifies which of these deliverables are included, who supplies the inputs and how each one is accepted.
- Process review and a pilot with a success measure agreed in advance
- Integrations with your CRM, email, forms, spreadsheets and databases through their APIs
- Triage, extraction and drafting workflows with a review queue
- WhatsApp or website assistant with handoff to a person
- Internal knowledge assistant over your own documents, citing its sources
- MCP server so assistants such as Claude or ChatGPT can use your tools with set permissions
- Data map, call logs and a test set rerun before every change
Where to start with AI automation
Start here when one repetitive task takes hours every week, has clear inputs and someone who can check the output. If you need a full product with accounts, roles and billing, start with web app development; if your clients need a place to see their own records, look at client portal development. If the task follows fixed rules, we will recommend a form or a short script instead of a model.
From one measured pilot to production
The pilot covers one process. We collect real past examples with the correct outcome, agree the measure before building (time per case, error rate or response time) and run the automation beside your team: it proposes, a person decides, we compare. If it does not beat the measure, we stop and you keep the findings and the test set.
If it does, we take it to production: permissions limited to the task, a review queue for uncertain cases, logs of every model call, alerts, a fallback when the model fails, spending limits and documentation your team can follow. Each month we review errors and costs, and we rerun the test set before any change of model or prompt.
Where does AI automation pay for itself?
Where the work repeats, the inputs are clear and a person can check the output quickly: sorting incoming messages, reading the same kinds of documents, drafting answers that already exist in your files, assembling routine reports. Decisions with legal, financial or relationship weight stay with people; the automation prepares them.
Some tasks are cheaper without a model. If a dropdown on your form can route an inquiry, it beats asking AI to guess from free text. If a report always pulls the same numbers, a scheduled query is faster and never invents a figure. Dardo will tell you when that is the case.
Most companies sit between trying AI and scaling it. A chatbot trial is easy; a process that runs every day with data from your systems, a person reviewing the exceptions and a number that shows it works is not. Closing that gap takes integration, review and measurement, which is what this service covers.
| Task | Good fit when | Keep a human when |
|---|---|---|
| Inquiry triage and routing | Messages arrive by form, email or WhatsApp and must reach the right person by service, location or urgency | The message is a complaint, a legal matter or from a key account |
| Document and form extraction | Invoices, orders or applications follow a few known formats and the fields go into a system | A wrong field changes a payment, an eligibility decision or a contract |
| Drafting replies for review | Most questions repeat and the answers exist in your documents | The reply commits to a price, a deadline or an exception to policy |
| Internal knowledge assistant | Staff search manuals, policies and past projects for answers | No document supports the answer, or two documents disagree |
| Reporting | The data sits in systems we can query and the format is fixed | The numbers go into a tax, financial or regulatory filing |
| Customer assistant | Questions are about hours, published prices, availability or order status | The person asks for a human, is upset, or the case falls outside published rules |
Can a WhatsApp or website assistant handle your customers?
Yes, for a defined set of questions, passing everything else to a person. The assistant answers from your published information and connected systems, such as hours, prices, availability and order status, says when it does not know, and hands the conversation to your team with its history.
WhatsApp sets its own rules. Meta's WhatsApp Business Messaging Policy allows automated replies within the 24-hour customer service window but requires prompt, clear and direct escalation paths, such as transfer to a human agent, a phone number or email. The assistant serves your business and its customers, not general questions. Every assistant we build includes:
- A clear label that the customer is talking to an automated assistant.
- Handoff to a person on request, on signs of frustration and on topics outside the agreed list.
- Answers grounded in your own content, with no commitments on price, dates or exceptions.
- Conversation logs your team can review, with personal data limited to what the case needs.
- A regular review of unanswered questions, which often reveal content your website lacks.
How do you connect AI to your systems without lock-in?
Through the APIs your tools already have, and through MCP where an assistant needs to use them. The Model Context Protocol is an open standard for connecting AI models to tools and data, under the Linux Foundation's Agentic AI Foundation since December 2025. The Linux Foundation reports it has been adopted by Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor and VS Code, with more than 10,000 published servers, so a tool you expose once can serve different assistants.
We keep the model replaceable. Prompts, test sets and integration code live in your repository and each step names its model, so switching provider means rerunning tests, not rebuilding. Tools start read-only; sending a message, changing a record or issuing a refund waits for a person's approval.
dardo.studio is our working example: its public MCP server and A2A interface let assistants read our services and prepare a project brief in English or Spanish. Both are read-only, and the person decides whether to send anything.
What happens to your data?
Before building, we map which data each step sends, to which provider, in which country and under what retention terms, and we send only the fields the task needs. Identification numbers, health details and payment data are masked unless the task cannot work without them.
In Colombia, Ley 1581 de 2012 requires prior, informed authorization to process personal data (Article 9), a legitimate purpose communicated to the person, and security measures against unauthorized access. Article 26 prohibits transferring personal data to countries without adequate protection levels, with exceptions such as express consent. Many model providers process data abroad, so your legal adviser confirms the authorization text and the basis for any transfer, and we build the records that show it:
- A data map: fields, systems, providers, countries and retention.
- Logs of every model call: input, output, model version, reviewer and decision.
- Access by role, with credentials kept out of prompts and code.
- Provider plans whose written terms state whether your data is used for training and how long it is kept.
- A way to delete a person's data when they ask or when the purpose ends.
Questions before you choose
What drives the cost of an AI automation project?
The number of systems to connect and the quality of their APIs, how varied the inputs are, how strict the review must be, and how many languages and channels are involved. Model usage is billed by the provider per use, usually to an account in your name, so volume matters too. Dardo scopes the pilot separately, so you decide on production with real numbers.
Which AI models do you use?
Dardo is provider-neutral. For each step we compare models on your test set and weigh accuracy, privacy terms, data location, speed and cost, so one workflow may use different models for extraction and for drafting. Changing provider later means rerunning the tests, not starting over.
Will AI replace our staff?
The aim is to remove repetitive work, not people. Your team reviews uncertain cases, handles exceptions and keeps customer relationships, while the automation prepares drafts and sorted queues. Dardo measures hours saved and errors caught; what you do with that time is your decision.
Is our data kept private?
Only the fields a task needs leave your systems, and we document which provider receives them, where and under what terms. Every model call is logged with its input, output and reviewer, and access follows roles. For personal data in Colombia, Ley 1581 de 2012 applies, and Dardo builds the authorization and transfer records your legal adviser approves.
What do you need from us to start?
One process with an owner, real past examples with the right outcome, access to the systems involved (a test environment if you have one) and the measure that would make the pilot worth continuing. Access to systems and examples usually sets the pace of a Dardo pilot more than the AI work does.
How do you maintain it as AI models change?
Providers update and retire models on their own schedule, so each workflow pins a model version and keeps a test set. Before any switch Dardo reruns the tests and compares accuracy and cost, and each month we review errors, unanswered cases and spending.
Who owns what you build, and what is not included?
You own the code, prompts, test sets and logs, kept in your repository and infrastructure, and the model provider accounts are in your name. Model usage fees, licences for your existing software and legal advice are not included. Dardo does not build automations that send messages or change records without review unless you decide in writing that a task is safe to run alone.
Sources & further reading
Sources behind this page, with further detail from the original publishers.
- Linux Foundation: formation of the Agentic AI Foundation (MCP)linuxfoundation.org
- Model Context Protocolmodelcontextprotocol.io
- Colombia, Ley 1581 de 2012 (personal data protection)cancilleria.gov.co
- WhatsApp Business Messaging Policywhatsappbusiness.com
- A2A Protocol specificationa2a-protocol.org