How to implement AI in a mid-sized company: choose the first role

A mid-sized company can have a good ERP and still depend on people who chase documents, copy data from one system to another and ask where each request stands. At a dairy in southern Chile, more than four thousand invoices a year reached the tax authority and nobody could say what had been bought, from whom or at what price: one person classified them by hand and purchase orders were built in Excel. The ERP was there. The work was still human.

That is the first decision when you implement AI in a mid-sized company. Not which tool to buy. Which role to digitize first.

Look for an assignment with a beginning and a delivery

"We need to use artificial intelligence" does not describe a job. "We need purchase requests to reach the approver complete" does, and it lets you talk about a role: what it receives, what it delivers, where the delivery lands and what stays out.

A good first role has three things: an assignment you can explain in one sentence, known systems where that work lives, and clear decisions that remain with a person. You do not need to automate the company to start. You need to know what result the first digital collaborator delivers and what is not its job.

Write it with the person who knows the exceptions

A job description almost never contains what it takes to operate the job. The experienced person knows when two names are the same supplier, which request is incomplete even if it looks whole, and which case should not go down the usual path. That knowledge lives in her head and leaves with her.

The role manual is written from that knowledge, not from a template. A useful manual says:

• What it receives and through which channels.

• Which documents and systems it may consult.

• What it delivers and where that delivery lands.

• What information is missing to continue, and who to ask.

• Which decisions it prepares but does not execute without a signature.

• Who resolves exceptions and with what context.

At Yunt the manual is built by sitting down with the person who does that work best today. The company contributes what it knows and names the manager; we build and operate the collaborator. A role, not software.

Define who signs before granting access

Sharing systems is not granting access to everything. A purchasing collaborator needs suppliers and orders; it does not need to see personnel files. Each role has its own permissions and limits, written before anything is connected.

Approval cannot be left at "someone will review it" either. There is a named person who signs what matters and a rule that says when she steps in. And if that person is unavailable, the expected behaviour is written too: wait, notify or escalate. With a manager and a record, from day one. Otherwise you are handing the AI the same ambiguities that slow your people down today.

Test it with the hard work, not the perfect example

Before putting the role into operation, test it with the ugly cases: contradictory information, duplicate documents, empty fields, requests out of scope. Sometimes the right answer is to stop and ask.

The review looks at the delivery and at the path. If the collaborator says it verified an order, the record shows what it checked and what it found. A well-written answer does not replace that evidence. That is what the glassbox is for: every action is written down with its reason and its backing, and before anything goes out, each claim is checked against the record. What has no backing is not sent, not published and not signed until a person resolves it. The manager reviews a case without rebuilding it from five emails.

Leave the next department ready

The first role does not have to cover four areas. But from the start it pays to know which connections and which information will serve when another department brings in its collaborator. At that dairy, the purchasing role started by classifying invoices against the chart of accounts and answering by email how much was spent and with whom; today it handles purchase requests and orders, and what comes next is the inventory of the five warehouses, on the same connections.

Expansion keeps what works and adds different responsibilities. Purchasing, finance and operations can share background without becoming a single role or losing their access limits. Each manager supervises her own; the company keeps a common base.

Frequently asked questions

Do we have to replace the ERP?

No. The collaborator is built on the systems you already have: your ERP, your email, your WhatsApp. The feasibility of each connection is reviewed when the role is defined.

Does my team have to learn to configure anything?

No. Building and operating are Yunt's job. Your team explains how the work is done, enables the agreed access and signs what requires approval.

How long does it take?

With the information ready, the collaborator is working in under a month. What sets the timeline is not the technology: it is how long the company takes to hand over the process, the access and the cases for the exam.

The next step

If your company has several departments, start with the assignment that today demands more coordination than it is worth. We can look at how to turn it into the first role and which other collaborators could work on that same base.

See how we define a digital collaborator: https://www.yunt.ai/en/digital-collaborator