Maria Kupa
All posts

17 November 2026 · 3 min read

Why Your Company Needs AI Skills: How to Teach Copilot Your Products

Part 2 of a series on how I use AI as a credit risk analyst.

Besides the month-end reporting and provisioning tasks I wrote about in my previous post, a lot of my daily work consists of ad-hoc questions. A colleague asks how many personal loan borrowers missed their payment this month, or a manager needs a quick check on late payment trends.

Now that organizations are adopting Microsoft 365 Copilot and giving employees licenses, the common assumption is that Copilot can simply search through shared folders, SharePoint, or network drives and figure out what we mean.

In practice, it doesn't work that way. Every company has dozens of folders, hundreds of files, and multiple versions of policies with shared access. But the bigger problem is that documentation describes business ideas, while our databases store raw operational records. When you ask generic AI to make sense of internal documentation on its own, it has to guess how business terms connect to actual data points.

To show what I mean without sharing sensitive company data, I asked AI to help me create a skill file for a synthetic credit product: a standard Nordic 5-year unsecured personal loan (in Swedish, privatlån 5 år).

Here are three common situations where letting Copilot search documentation on its own breaks down, and why writing an explicit product skill helps.


Where generic Copilot breaks down

1. Product names vs. database codes

In presentations and everyday conversations, we call this product privatlån 5 år (5-year personal loan).

In the database, it doesn't exist under that single name. It is split into three separate product codes: direct online loans, broker-sourced loans, and an older portfolio acquired during a company merger.

If an analyst asks Copilot: "What is the delinquency rate on our 5-year personal loans?", which loans should it count?

  • For total portfolio exposure or provisioning, you need all three codes.
  • For recent underwriting scorecard performance, you must exclude the acquired portfolio because those loans were granted under older, legacy criteria.

SharePoint documentation doesn't tell Copilot which code belongs where for a specific analytical question. Without explicit rules, Copilot either misses the acquired book completely or mixes it into new origination reports.


2. Bank delays vs. real late payments

Most customers pay on the 27th by direct debit (Autogiro).

It takes up to 4 banking days for banks to process the transfer. In the database, the system starts counting days late right away, but our policy gives a 4-day grace period for the money to arrive.

If you ask Copilot on the 29th who is late, it will look at the raw database counter and tell you 20% of customers missed their payment. In reality, they paid on time—the money is just moving between banks.


3. What counts as an "active" customer?

When someone asks: "How many active personal loan customers do we have?", different teams mean different things:

  • Marketing counts anyone with an open account or profile, even if their loan is fully paid off.
  • Credit risk only counts borrowers who currently owe money (outstanding balance > 0).
  • Operations counts every contract that hasn't been formally closed, including accounts sent to debt collection.

If Copilot searches through internal folders, it will find monthly presentations from Marketing, Risk, and Operations—all with slides labeled "Active Customers", each reporting a different number. Without explicit instructions, it will pick one at random, giving you a figure that doesn't match what you actually asked for.


Why writing an AI skill makes the difference

GitHub Copilot already knows how to write code and SQL. What it doesn't know is our company context.

An AI skill is just a clean markdown file (SKILL.md) where we write down the product logic, data definitions, and company rules in plain language. Instead of expecting Copilot to search dozens of SharePoint folders and guess which file is current, we point it to a single curated file that explains:

  • Which database codes map to the commercial product.
  • How to handle the 4-day bank clearing delay so on-time payers aren't marked as late.
  • What counts as an active customer (borrowers with balance > 0 vs. open accounts).

I put together an example companion skill for this 5-year personal loan product in my GitHub repository: skill/consumer-personal-loan-risk/SKILL.md.

By writing down clear product instructions instead of letting AI search through shared folders on its own, Copilot stops making plausible guesses—and becomes an assistant you can actually rely on for daily credit analysis.

Ask me about this article

I read every question and answer them myself. Leave your email if you'd like a personal reply, or check back here for answers I publish.

Only used to reply to you. Never shown on the site.