Skip to content
Let's talk about your case
Hábil AI

Answers with your real data, at any hour, with a person deciding what matters.

Hábil AI serves your customers, looks up your information and takes steps in your processes on its own, with a person approving anything that moves money or data. It connects to the systems you already have.

The AI platform connected to your systems.

Let's talk about your case

Prefer email? Write to us at hola@habil.mx

What is Hábil AI?

Since 2024 we have connected AI to the company's systems, with a person deciding what matters.

Hábil AI helps companies like yours use artificial intelligence with the data and rules of their own systems. We don't sell a generic tool: we solve specific processes in your operation, with the AI you already use or with one we install for your company, without putting at risk what already works.

Your team most likely already uses ChatGPT, Copilot or Gemini. But that AI doesn't know how much a customer owes you, where a claim stands or why an invoice didn't reconcile: that data lives in your systems, and the AI can't see it. That is where our work begins.

Trying out an AI is easy; the hard part is making it work with the data in your own systems. The model is almost never the problem: connecting it to your systems, your rules and your people without putting at risk what already works is. Since 2006 we have been integrating systems that sustain critical operations; that experience lets us bring AI into real processes without asking you to replace your ERP, your CRM, your core system or your controls.

What sets Hábil apart: we don't start with a chatbot or a model. We start with the process, connect it to your systems and your rules, define what the AI cannot do on its own, and test it with real cases before recommending it.

Inside, Hábil AI is six pieces, and each one does a single job: one understands and writes, another checks your documents before answering, another is the guarded gateway to your systems, Nova talks with people, the agents act inside your process and forecasting estimates what may come. They are combined according to the problem you want to solve.

Example illustration: real data, with permissions and a record.

Three levels, depending on what you need

Hábil AI is a product and a service. The product is the platform. The service is what we do with you so the AI works with the data and the rules of your systems. It is engaged by level, and each level builds on the previous one.

The platform

Your AI connected, with a record

What you get with the product:

  • An installation just for you, on your infrastructure or on ours, operated by Hábil.
  • The AI you choose: the one you already pay for or an open model on your infrastructure.
  • A guarded gate to your systems: each application with its own permission, usage limits and a record of who queried what.
  • Answers with a source: it checks your documents before answering and tells you where each piece of data came from.
  • A person approving whatever moves money or data.
  • Your conversations do not train any model without your written authorization.

To get started

One process, with indicator and exit

We choose with you a process where you are losing time, money or customers today, and we leave it running with an indicator, an exit criterion, and scope and cost agreed before we start, including the processing cost with its cap.

That first process uses only the pieces it needs. When it is done, you know what the AI does, what your team keeps doing and with which controls.

To grow

More processes, pieces and Hábil Lake

We add processes and pieces —Nova for your customers, agents inside your operations, prediction— and, if you decide to, we connect it with Hábil Lake. We install and maintain it with you: updates, model changes and monitoring, alerts if something goes down and tested backups, with a monthly service.

Find your area

General Management"Where does AI give me money back this year without risking the operation?"We start with the process where money slips away today, such as VAT that is not credited, and we tell you what changes before you invest
Sales"How do I keep the AI from promising what it shouldn't?"The AI answers only with your rules: it doesn't offer prices, discounts or terms that you didn't define
DigitalYour digital channel loses customers halfway through the process: sign-up, quotation or payment depend on systems that don't respond in timeNova serves customers at any hour and leaves the case ready for your team
OperationsPeople keying in by hand data that already exists in another systemAutomated data capture: the AI extracts the data and your team handles only the exceptions
FinanceA close that doesn't balance and VAT lost because of payment complements that never arrivedWe tell you how much VAT is at risk, from which supplier and on which invoice
Technology (CIO / CTO)"I already have a contract with an AI; how do I use it in my operation?"We connect it to your systems with permissions and a log; the detail is in the section for your team, further down
CollectionsCollecting with the account's real data at hand, without risking a complaintBalance answers and a payment link over WhatsApp; your collector closes the agreements
Planning and Purchasing"Every season I buy too much of what doesn't sell and fall short on what does."Forecasting estimates demand by product line and by month, with a range, and your team decides the purchase with that figure

Where AI adds value

Each case says what was happening, what changes, why it was hard and what we have built. It describes the process that can be designed for your company, together with the integration experience at Hábil that backs it; in the first conversation we tell you which part is proven and what must be validated in your environment.

VAT that doesn't get lost · Finance

Before: the company pays invoices whose payment complement never arrived, and that VAT can't be credited. Nobody knows how much it is.

What changes: you know how much VAT is at risk, from which supplier and on which invoice, in time to recover it.

Why it was hard: what the SAT (Mexico's tax authority) says has to be cross-checked against your banks and your accounting, invoice by invoice.

What we have built: we integrate electronic invoicing and SAT information with our clients' accounting systems.

Pieces: the connection and the AI that understands and writes.

Cuadre already has it solved for your invoicing.

AI on your systems, without replacing your core systems · Technology

Before: when an interface fails between your administrative system (ERP) —SAP, Oracle, Microsoft Dynamics, Odoo, CONTPAQi or another—, your CRM and the rest of your operation, your people spend hours reading technical logs to find out what broke.

What changes: the AI tells you in plain language what happened, which invoice, payment or purchase order it affects and what to do. Your system remains the one that decides.

Why it was hard: you have to understand at once the system that fails, the one that receives and the business behind them, without putting the operation at risk.

What we have built: we connect the most widely used ERPs and CRMs —SAP, Oracle, Microsoft Dynamics, Odoo, Salesforce, HubSpot— with the operations of banks, insurers and regulated companies.

Pieces: the AI that understands and writes, and the connection.

Automated data capture · Operations

Before: your team spends hours copying into your systems the data from IDs, tax status certificates, proofs of address, bank statements and invoices.

What changes: the data arrives on its own, with the evidence of where it came from, and a person only confirms.

Why it was hard: documents arrive as photos, crooked and incomplete, and a wrongly captured data point becomes a compliance problem.

What we have built: we have already solved ID reading for a bank in Mexico.

Pieces: the AI that understands and writes, and the connection.

Account opening without a branch · Banking and fintech

Before: opening an account required going to a branch, and someone keyed in by hand the data from each ID.

What changes: your customer opens their account from their phone. We verify that they are a real person in front of the camera, compare their face with their ID and validate their data against the INE (national electoral institute) and RENAPO (national population registry). Doubtful cases go to a person on your team.

Why it was hard: photos arrive crooked and incomplete, and every validation has to withstand an audit.

What we have built: the identity validation of a bank in Mexico, including the INE lookup; we offer it complete.

Pieces: the connection (identity verification is built in as a service of the process).

Collections over WhatsApp that end in a payment · Banking and fintech

Before: your collectors call and send messages that go unanswered, and when the customer replies nobody has the balance or the payment method at hand.

What changes: it answers with the account's real data, explains the options you authorized, sends the payment link and confirms the payment; agreements and sensitive matters are closed by a person.

Why it was hard: every answer has to come from your system, within your policies and without risking a complaint.

What we have built: direct-debit payment collection for an insurer in Mexico.

Pieces: Nova and the connection.

The simple claim, ready for a decision · Insurance

Before: a simple claim sits on the desk for days while someone gathers documents and checks the policy.

What changes: the case reaches the claims adjuster complete, and the decision remains theirs.

Why it was hard: the rules of each policy are different, and an error costs money and reputation.

What we have built: issuance and claims systems for insurers in Mexico.

Pieces: the one that checks your documents, and the connection.

Claim reporting over WhatsApp, 24 hours a day · Insurance

Before: the policyholder wants to report at any hour, and your desk receives loose photos and documents through several channels.

What changes: the policyholder reports, sends photos and documents and checks progress from the chat; your desk receives the file in order.

Why it was hard: every report has to reach your claims system complete, with no manual data entry.

What we have built: the mobile apps that policyholders of two insurers in Mexico use for their transactions.

Pieces: Nova, the one that checks your documents, and the connection.

Stuck orders nobody saw · Operations

Before: an order gets stuck between the store, the warehouse and the invoicing system, and nobody finds out until the customer complains.

What changes: every morning an agent reviews orders with no movement, finds the reason for each one and leaves them to your team; if one needs to be reopened, it prepares it and a person authorizes it.

Why it was hard: every order goes through several systems, and finding where it got stuck means following it end to end.

What we have built: we connect online stores and order systems with the most widely used ERPs, so an order can be followed end to end.

Pieces: the agents and the connection.

Next season's purchase, with a range · Commerce

Before: buying for the season is done from memory and last year's Excel file. There is too much of what doesn't sell and too little of what does.

What changes: forecasting estimates demand by product line and by month, with a range, and your team decides the purchase with that figure at hand. The decision is still yours.

Why it was hard: it takes organized, clean sales history, and a forecast that beats "the same as last year". If it doesn't beat it, it isn't used.

What we have built: online sales platforms, with their inventory and their prices.

Pieces: forecasting.

This morning’s stalled orders, with their reason

Example · illustrative data

The same orders: discovered when the customer complains, or reviewed by an agent every morning. Switch the view to see the difference.

This morning’s stalled orders, with their reason
Today: until the customer complains
OrderWhere it stalledWhat happens
20817Warehousenobody has seen it
20822Invoicingnobody has seen it
20831Shippingnobody has seen it

Nobody finds out an order got stuck until the customer complains.

Three ways to do it, depending on your problem

They can be engaged separately or together. When the AI needs to answer with up-to-date information from your operation, the connection described in the first one is part of the scope.

What each piece does

Behind the three ways, six pieces are at work. Each has a single responsibility, and none does another's job. Each is told in four parts: the problem, what changes, why it is hard and an example.

1. The AI that understands and writes

The problem: your team already uses an AI, but that AI doesn't know your company, and sending sensitive information to a third party isn't always possible.

What changes: the AI understands the question in Mexican Spanish and writes the answer in your company's tone. You choose which one: the one you already pay for —ChatGPT, Copilot, Gemini or Claude— or an open model installed on your servers or in your private cloud, where conversations and documents are processed on the infrastructure you define. Both can coexist: sensitive work is processed in your company and the rest uses the AI you already pay for.

Why it is hard: not every model is fit to serve customers: several fail when using tools or following a chain of steps. That is why every model goes through our tests, with your real cases, before we recommend it, and before you buy equipment we measure the size you need. The one that makes things up least with your cases wins.

This piece is not a data source: it understands and writes, but the figures come from another piece. Your investment stays in what is yours —the connection, your documents, the controls and the tests—, so you can change models without redoing the rest.

An example: an insurer wants the AI to read files that cannot leave the company. We install an open model on its servers, with a company profile —tone, terms, rules— and, if it is measured to improve, we adapt it with materials the company approves. The rest of the team keeps using the AI it already pays for.

2. The one that checks your documents before answering

The problem: an AI that answers from memory makes things up, and says them with complete confidence. When it gets a policy or a coverage wrong, the mistake is yours, not the AI's.

What changes: before answering, the AI searches your business documents —manuals, policies, catalogs, product terms— for what applies to the question, and answers with that. Every answer says which document it came from, and your team can review it. When a document changes, the search uses the new version.

Why it is hard: finding the right paragraph among thousands of pages is not the same as finding a similar one. And what changes by the minute —prices, stock, balances, orders— cannot come from documents: that is asked of your systems at that moment, through the connection.

Before any answer goes out, a program checks that every number, date or policy comes from a source consulted at that moment. It is built to tell you it didn't find the answer instead of making it up, and it tells you where each piece of data came from.

An example: an insurance agent asks whether a policy covers a treatment. The AI searches that product's general terms, answers with what they say and shows the paragraph. If the terms don't settle it, it says so and passes the case to a person.

3. The connection: the guarded gateway to your systems (includes your MCP server)

The problem: the AI can't see your systems. And opening them up without control is worse: it could see too much or do something nobody authorized.

What changes: we design and build the connection between the AI and your systems —customers, policies or loans, collections, invoicing—, even if they are twenty-year-old systems and without stopping your operation. The connection decides who gets in and what they are allowed to access, and keeps a record of every query. Every query or action on your systems goes through it.

Why it is hard: each system that uses it needs its own identity and its own permissions. The price has to come out with the buyer's price list, and an error cannot arrive disguised as an answer: if something failed, the connection says so.

Two ways to use it:

  • For your AI and for whoever you authorize. It is what the market calls an "MCP server": an open standard that the most widely used AIs use to connect to other systems, and the name your technology team will look for. The same connection can let your insurance agents, suppliers or customers see only what is theirs: their policy or their invoice.
  • For your own systems, with no conversation. Your online store, your order system or your internal systems ask the AI for a result and get an answer ready to use: alternatives to an out-of-stock product, related products or a summary of what your customers think. Nobody has to write to the AI.

Examples of functions built into your systems, with no conversation (they are use cases that go through the connection, and each is named for what it solves)

  • Automated data capture. Reads invoices, orders or forms and records them in your system. Anything that doesn't match is set aside for a person to review.
  • VAT at risk. Finance sees the VAT at risk invoice by invoice. Cuadre already has it solved for your invoicing.
  • Customer feedback summary. Reads your customers' reviews, surveys and messages, and gives you what comes up most often. If there aren't enough to draw a conclusion, it tells you so.
  • Recommendations. Suggests to each customer what else could be useful to them, based on what they have already bought.
  • Assisted reconciliation, with Cuadre. Matches payments against invoices and bank transactions, and leaves you only the differences.

Before the connector, your APIs. A connector's reliability depends, in large part, on the APIs it calls. If an API doesn't explain what it does, doesn't tell one error from another or charges again when the agent retries, the agent may carry out wrong or hard-to-reverse actions even if the connector is well built. That is why we start by reviewing your critical operations: which are ready, which are fixed in place and which are better covered with an intermediate layer that adapts and protects them, aiming not to rewrite your system.

An example: a manager asks "which customers have payments more than 60 days overdue?". The AI queries your system through the connection, with that manager's permissions, and answers with the real list.

4. Nova: the one that talks with people

The problem: your customers write at any hour and your team answers the same thing all day. What does need a person —closing, collecting, retaining— waits behind everything else.

What changes: Nova answers your customers on WhatsApp or on your website, at any hour, with the real information from your systems. When the case needs a person, Nova hands it to your team already put together: who it is, what they asked for, what they were told and what is missing.

Why it is hard: Nova can't improvise. Today's price and stock are asked of your system at that moment. Balances, payments and procedures are only shown after confirming the person's identity. And if on WhatsApp the price depends on who is buying, it is only given to someone with a linked account.

Nova is not a fixed-answer chatbot: it is an assistant connected to your systems. It is enabled per process and per channel; balances, payments and procedures are shown only after the person has been validated with your company's rules.

How it works

  • It answers with what you know. It connects to your documents and your systems. It doesn't tell your customer a price, a date or a policy that it didn't check there at that moment.
  • It knows when to stay silent. If the information isn't enough or the case is sensitive, it doesn't improvise: it passes it to a person on your team, with full context, and your customer knows someone will write to them.
  • With rules, an audit trail and privacy. It sees only what you allow it and logs who queried what. We never use your conversations to train a model without your express authorization. Whatever has an effect —a refund, an exception, an agreement— is decided by a person.
  • With the AI you choose and your brand's tone. It works with the AI you already have under contract or with models installed in your company, and it speaks like your company, not in a generic tone.

What changes with Nova

  • Customer service. Today: at eleven at night a prospect asks where their quotation stands or how much it costs; nobody answers until the next day, and by then they have already asked elsewhere. With Nova: it answers right away. It gives the price, if it is public, and the status of their quotation after confirming their identity. It leaves the case ready for your team to close first thing in the morning.
  • Collections. Today: your collector spends the morning answering "how much do I owe?" and "where do I pay?", and the conversations that do need an agreement wait behind all the others. With Nova: those answers go out on their own, with the payment link, and the collector keeps the conversations that do need an agreement, with the full history. The collector negotiates; Nova does not.
  • Claims. Today: the policyholder reports a claim by phone and sends photos and documents separately; the file is put together by hand over several days. With Nova: it receives the report, asks for what the policy requires and hands it in order to your team. Reviewing the documents and deciding the payment remains with your claims adjuster.
  • Your internal team. Today: in Finance or Technology someone asks why a VAT was not credited or why an interface got stuck, and the answer depends on who is on call. With Nova: it explains in plain language what your systems and policies say, cites where it got it and passes to the right person whatever it cannot resolve.

What Nova does not do. It doesn't approve refunds or exceptions, doesn't negotiate discounts, doesn't ask for card details in the conversation and doesn't handle complaints or emergencies on its own: it passes them immediately to a person.

Review your case with us: we take your real questions and explain how Nova would handle them.

With your customers, the assistant can carry your own brand, powered by Hábil AI.

An example: at eleven at night a customer asks "how much do I owe and where do I pay?". Nova confirms it is really them and answers with their balance and the payment link. If they want to negotiate, Nova records the case and passes it to your collector in the next business hours, with the full context.

5. The agents: they act inside your process, with a person approving

The problem: there is work nobody does because nobody has time to check: stock running out, an incomplete product sheet, an order stuck for days. It is discovered when it has already cost money.

What changes: the agents review your operation on schedule, look up, compare and prepare the work in your systems, and alert whoever is responsible. Anything that moves money or data is carried out with a person approving. You decide, by process and by tool, how far each agent goes: only look up and alert, propose for a person to approve, or automate within limits you set.

Why it is hard: an agent that acts needs permissions limited to the bare minimum, a record of every decision it made and a way to prevent the same action from happening twice. And every system it touches has to be ready to be used by a program, not just by a person.

When a step needs several actions in a row (looking up, comparing, preparing), the agents take care of it, with permissions limited to what is necessary and a person approving anything that moves money or data.

An example: every morning an agent reviews stuck orders, finds those with no movement for more than two days and leaves your team the list with the reason for each one. If one needs to be reopened, it prepares it; a person authorizes it.

6. Forecasting: it learns from what already happened to estimate what may come

The problem: buying, collecting and planning are done from memory and last year's Excel file. There is too much of what doesn't sell, too little of what does, and the receivables that will fall behind are noticed when they already have.

What changes: forecasting estimates demand, collections or risk by product line and by month, always with a range —the most likely, the minimum and the maximum—, and your team decides with that figure at hand. What it forecasts is stored, and over time you can see whether it was right.

Why it is hard: it takes the history of your operation, organized and clean, and a forecast that beats the simple ways of guessing, such as "the same as last year". If it doesn't beat them, it isn't used, and we tell you.

Forecasting learns from your organized tables, not from your conversations, and no language model is trained or fine-tuned with your data without your written authorization. If you have Hábil Lake, it learns from the data the Lake has already organized and reconciled.

An example: before the season, Purchasing sees how much each line is expected to sell, with its range, and decides how much to order. At the end of the season, it compares what was forecast with what actually happened.

Where it runs

On your infrastructure —the cloud you already pay for or your own servers— or on ours, with Hábil in charge of operations. Each client has its own installation: your data is not mixed with anyone else's. If you use a model that Hábil operates, that model may serve several clients, but it doesn't keep the conversations or the data of any of them. If the model is on your servers or is the AI you already pay for, nothing is shared.

With someone accountable. We install it and maintain it with you: updates, model changes and monitoring, alerts if something goes down and tested backups, on the schedule we agree, with a monthly service.

What we never do

Let the AI make on its own a decision that costs money: authorizing a payment, making an exception or closing an agreement.

Using your conversations to train or fine-tune a model without your written authorization.

Show the AI more data than you authorize.

Promise you an AI that is never wrong: that is why every answer comes from your systems, cites where it came from and leaves what matters in the hands of a person.

The details of control, and what your Compliance area receives, follow.

Control, security and your data

Your information doesn't stay in the model: we query it live, so it can be updated, deleted and cited. Your conversations don't train any model without your written authorization. Permissions are decided by your system, not by the AI.

What the AI never does on its own

It doesn't collect, approve, issue or modify a record without a person or a rule you defined authorizing it. It can register a request or prepare a transaction; a person authorizes it.

It doesn't offer prices, discounts or terms that are not in your rules.

It has only the permissions your Technology team gives it, per system and per action.

If an incorrect answer or action is detected, it is stopped, the log is reviewed, the rule or the integration is corrected and it is tested again before resuming. The action is executed by your system, not by the AI.

Before it operates, it is tested with your real cases and an agreed exit criterion.

What your Compliance area receives

Log

every proposal is recorded: what the AI saw, what it proposed and who authorized it.

Access

which person or application can query what, per system and per action.

Change audit trail

every change is recorded like any other change in your system.

Tests

the real cases the AI was tested with and the exit criterion agreed before it operates.

Reason for each decision that affects a person

it is kept and a person can review it. The 2025 personal data protection law recognizes your customers' right to object, in the cases it provides for, to automated decisions made without human intervention; your legal area validates how it applies to each case.

You receive the evidence so that your auditor and your regulator can review it.

Your data

You decide which AI sees your data: the one you already pay for —ChatGPT, Copilot, Gemini or Claude— or open models installed in your company. With models installed in your company, conversations and documents are processed on the infrastructure you authorize. We tell you, with data, when it makes sense to adapt a model and when connecting it to your documents is enough.

With control and with evidence.

In front of the model sits an access gateway: each application with its own permission, usage limits and a record of who queried what, which your Compliance area can review.

For your technology team

What your CIO or CTO usually asks.

The connector.

It is an MCP server built to measure for your systems. It defines what each person or application queries, keeps a log of who queried what, and the AI proposes: the action is executed by your system, with the authorization of a person or of a rule of yours. There can be one connector per audience: employees, customers, suppliers and intermediaries, each with its own scope.

APIs ready for agents.

Before connecting an agent we review your operations against what an automated consumer requires: that they clearly explain their purpose, the data they receive and the responses they give; consistent errors, preferably in the standard Problem Details format (RFC 9457); retries that don't duplicate, a dedicated identity and permissions for each agent, visible limits and notice before a version is retired. What is missing is fixed in the API or in an intermediate layer that adapts it, aiming not to touch the source system.

Your systems.

We connect the most widely used ERPs and CRMs —SAP, Oracle, Microsoft Dynamics, Odoo, Salesforce, HubSpot—, twenty-year-old systems and electronic invoicing with SAT information, without replacing your core system and without stopping the operation.

The AI you already pay for, or your own.

Models from OpenAI, Anthropic or Google, or open models such as gpt-oss, Qwen, Mistral or Llama on your servers or in your private cloud. You can switch from one to another without redoing the integration, and both options can coexist: what is sensitive is processed in your company and the rest uses the AI you already pay for. It connects to any model server compatible with the OpenAI API standard.

Lower cost and less repetition.

With the contracted AI, when the provider allows it, we organize your company's instructions so that context is not repeated in every query; we measure cost, speed and quality before recommending it.

Answers with a source.

What changes —prices, policies, balances— the model consults in your current documents and systems, with the source at hand. When it makes sense, we adapt an open model to your company's tone, terms and formats, with materials you approve, and validate it with real cases.

Tested before recommending it.

Not every open model is suitable for serving customers: several fail when using tools or following a chain of steps. Each one goes through our tests before we recommend it. In our September 2026 tests, an open model runs entirely on a single 16 GB graphics card and solves multi-step tasks using tools; in the technical conversation we explain what was measured and with what criterion.

The right size.

We measure in our lab which model you need, how much it can serve at once and at what speed, before you buy equipment. The AI we propose is understood by the team that proposes it.

After launch.

We maintain it: adjustments, model changes and monitoring, under a monthly service.

LLM — the AI that understands and writes.

Understands the question and writes the answer; it is not a data source. With the AI you already pay for or with an open model on your infrastructure, behind a per-provider adapter.

RAG — the one that checks your documents.

Searches your business knowledge before answering, with hybrid search and reranking; cites the source. It never answers prices, stock or orders from memory.

API — the connection.

The single gateway: identity, permissions per calling system, a published and versioned contract, and a record of every query. Includes the MCP server for your AI and the no-conversation answers for your systems.

Nova — the one that talks.

The conversation on the web and WhatsApp, with verified identity for personal data and handoff to someone on your team.

Agents — the agents.

Look up, compare and prepare in your systems, with three levels of autonomy configurable per client and tool; anything that moves money or data is carried out with a person approving.

ML — forecasting.

Forecasts by product line and period, with a range; learns from organized tables and is only used if it beats a simple estimate.

Every number, date or policy has to come from a source, and a program checks it before it goes out. If a piece of data is missing, it says so; it doesn't make it up.

How to build a connector for your industry (blog article).

What your system needs so an agent can use it with control (blog article).

Hábil AI with Hábil MDW and Hábil Lake

AI answers only as well as the data it has at hand. If your numbers live in three systems and an Excel file, the AI inherits that mess. Hábil Lake reconciles them first; the AI uses them afterwards. Hábil Lake explains what happened; Hábil AI's forecasting estimates what could happen, and the AI lets you ask both.

Hábil AI and Hábil Lake are two independent platforms: they are engaged separately and neither requires the other. Hábil AI works with the systems you already have. When you have both, they feed each other like this:

With Hábil Lake

Your AI answers with your numbers.

When someone asks which product line sold the most or how collections are going, the answer comes from the Lake's organized, reconciled data, through queries defined in advance and with that person's permissions.

Forecasting learns from clean data.

It learns from the Lake's organized tables, not from raw operational data, and what it forecasts goes back to the Lake to check whether it was right.

Product information reaches the AI.

Your product sheets, built in the Lake, feed what the AI knows about your catalog.

Your documents, as raw material.

The Lake also keeps your business documents —contracts, manuals, product sheets— as raw material for your AI.

Today's figures, from the system.

The AI doesn't take today's price and stock from the Lake: it asks your system for them at that moment, with the identity of whoever is asking.

Deletion is a single procedure.

If a person asks for their data to be deleted, the procedure reaches both platforms.

AI acts only as well as its connection to your systems. If each system is connected on its own, the AI inherits that tangle and nobody knows what it touched. Hábil MDW connects first; the AI then goes through the same gateway. Hábil MDW knows where each operation stands; the AI looks it up, explains it and prepares the next step, with a person approving.

Hábil AI and Hábil MDW are two independent platforms: they are engaged separately and neither requires the other. When you have both, they feed each other like this:

With Hábil MDW

Your AI acts through the same gateway.

When an agent from Hábil AI looks up an order or prepares a refund, it does so through Hábil MDW's guarded gateway, with its own identity, its permissions and an audit trail of every call. Anything that moves money or data is carried out with a person approving.

Nova knows where the order stands.

When a customer asks on WhatsApp where their order is, Nova answers with the status Hábil MDW keeps, not with a guess.

The AI reads the trail.

Every morning an agent reviews stalled operations, finds the reason for each one in the platform's audit trail and leaves them to your team.

Forecasts go back into operations.

What Hábil AI's forecasting estimates —for example, restocking— reaches your systems through Hábil MDW, as a proposal a person approves.

Stuck operations, with their next step.

When an operation stalls, Hábil MDW flags it; the AI explains where it stopped and why, and proposes the next step, which a person approves.

Differences, in business language.

When two systems don’t match, Hábil MDW flags the difference and the AI explains it in business terms: what doesn’t match and between which systems.

Discover Hábil Lake: a single version of your numbers, reconciled against your system, in your Excel or your Power BI.

Discover Hábil MDW: your systems, finally in agreement, with your core system, your ERP and your channels connected without replacing them.

How we start

In a first conversation, at no cost, we choose a process where you are losing time, money or customers today.

1

You write to us on WhatsApp with your industry, your role and your main system.

2

We schedule 45 minutes with a Hábil architect.

3

We ask for the process, the systems it touches and the approximate volume. We don't ask for access or for your customers' data.

4

You leave with a map of a process with AI: the candidate process, the data it needs, its risks, the role of the person and how success is measured. It says what the AI would do, what your team would keep doing and with what controls, and whether it makes sense to go ahead.

5

If it does: a project with an indicator, an exit criterion, and scope and cost agreed before starting, including the processing cost with its cap. In the proposal we also define the deliverables and their technical documentation, each party's responsibility, ownership of what is built, the support and incident-handling arrangement, and the security controls of your environment.

The route is: diagnosis → proposal with scope, cost and risk → test with an exit criterion → operation.

We work with a few clients at a time: the architect who reviews your diagnosis follows your project to the end.

We tell you no when the process doesn't have enough volume, when the rule changes every week or when an error costs more than it saves.

Related

This is what AI looks like on real data: ask your CFDI questions in natural language (sample conversation).

See Cuadre

The same layer that connects your systems is what gives AI access to your data

Systems integration

Every AI query inherits the person's permissions

Control who has access

Frequently asked questions

What people ask us before starting

  1. No. We can work with the one you already have —ChatGPT, Copilot, Gemini or Claude— or with one installed on your servers.

  2. You decide. If you choose the AI installed in your company, the information is processed on the servers you define; before it operates we agree on which data circulates and where.

  3. Yes, which is why it answers with data from your systems, cites where it got it and leaves what matters in the hands of a person. And before it operates it is tested with your real cases.

  4. It works especially well where there are policies, loans, collections, claims or customer service at high volume, of any size.

  5. Yes. They are engaged separately or together.

  6. The process, the systems it touches and the approximate volume. We don't need access or your customers' data in the first conversation.

  7. No. Your investment stays in the connection to your systems, your documents, the controls and the tests; the connection is designed so that the model can be changed without redoing the integration.

  8. After it operates we maintain it with you (adjustments, model changes and monitoring). And every decision that affects a person keeps its reason so that a person can review it; the legal detail is in the Compliance section.

  9. Scope and cost are agreed before starting, including the processing cost with its cap.

  10. Only what you allow it to. Each agent has a level: look up and alert, propose for a person to approve, or automate within limits you set. Anything that moves money or data is carried out with a person approving, and every decision is recorded.

  11. It always comes with a range, not a single number, and it is only used if it beats the simple ways of estimating. What it forecasts is stored and compared against what actually happened, so you can see whether it is right.

Don't see your question?

Let's talk about your case. In 45 minutes we tell you whether it makes sense, what depends on your systems and what risk you must resolve before going ahead.

Let's talk about your case

Prefer email? Write to us at hola@habil.mx