Adding generative AI to software you already run: a practical guide for managers

Qasim Ali · Senior Software Engineer (Java, AI), CS2 Technologies
Published · 7 min read

Most companies asking about generative AI don’t want a new product. They want the CRM, the customer portal or the internal tool they already pay for to draft, summarize, sort or search a little better.

That’s a sensible goal, and usually a smaller project than people fear. The risk sits elsewhere: a feature that leaks data, costs more than it saves, or writes something confidently wrong in front of a customer.

This guide covers the decisions that keep those risks small. It’s written for the person who signs off on the project, not for the engineers who build it, and it draws on the AI features we run in production in our own products.

1. Pick one narrow job for the first release

“Add AI to the CRM” isn’t a use case. “Draft a reply to an inbound lead from the details already on the record” is. So is “summarize a long support ticket before an agent opens it” or “pull the invoice number, date and total out of an uploaded PDF.”

A good first job has three things going for it. Someone does it by hand today, so you know what good output looks like. Mistakes are cheap to catch before they reach a customer. And it’s frequent enough that saving a few minutes each time adds up.

Resist the urge to launch a general chat window on day one. An open box invites questions the system was never prepared for, and it’s much harder to test than a feature with one input and one expected output.

For managers: Write the first use case as a single sentence with a clear input and output. If you can’t, it isn’t ready to build.

2. Decide where the data may go before you pick a model

Every request to a language model sends some of your data to whoever runs that model. So the first question is which provider, in which region, under which agreement.

The common options each have a different shape. Amazon Bedrock runs models from several vendors inside your AWS account, in the region you choose. Azure OpenAI Service offers OpenAI’s models inside Microsoft Azure, with its own regional deployments and enterprise terms. The OpenAI API directly is often the quickest place to start. Any of them can be the right answer. What matters is reading each one’s data-use and retention terms, checking that the model you want is offered in the region you need, and confirming you can sign the agreement your sector requires.

When we moved Notermed, our AI medical scribe, onto Amazon Bedrock to meet a HIPAA requirement, the real work wasn’t the model change. It was making it impossible for patient data to reach anything outside the approved boundary. The app now refuses to start if someone points it at an unapproved provider.

For managers: Put the data rule in writing first (which data, which provider, which region) and ask your team how the code enforces it.

3. Keep a person in the loop, and design the screen for it

For a first release, the AI should prepare and a person should decide. In Notermed, the AI drafts the clinical note and the doctor reviews and signs it. In AgentlyLeads, our own CRM, the assistant can draft an email, but nothing it writes goes out without human approval.

That only works if reviewing is easy. Show the draft next to the source it came from. Make editing a single click away. Highlight the fields the AI wasn’t sure about instead of hiding them. If checking the draft takes longer than writing it from scratch, people will either stop using the feature or stop checking, and the second outcome is worse.

It also helps to decide early what the AI must never do. In a medical note, that means never guessing a diagnosis or adding findings nobody mentioned. Your list will look different, but you should have one.

For managers: Ask to see the review screen in the first demo. How fast someone can approve or fix a draft decides whether the feature gets used.

4. Test with your real documents before anyone relies on it

Clean sample data hides problems. Real inputs come as scanned PDFs, forwarded email chains, spreadsheets with merged cells and notes typed in a hurry. When we tested Notermed with real uploads, some files failed outright because they exceeded the model’s input limit. Clean test files never would have shown that.

Before launch, collect a set of real examples, with personal details removed where needed, and write down what a correct answer looks like for each one. Run every version of the feature against that set. When someone changes the instructions or swaps the model, the same set tells you whether quality went up or down before a customer notices.

For managers: Ask for an evaluation set built from your own data, and ask to see its results every time the prompt or model changes.

5. Put a cost limit on every request

Language models are billed by how much text goes in and comes out. A feature that looks cheap in a demo can get expensive once it runs on every record, every night, for every customer. Long documents and chat histories that keep growing are the usual culprits.

The fixes are mostly design choices. Send the model only the part of the record it needs. Use the smallest model that passes your test set. Cache answers that don’t change. Set a monthly budget per feature and a usage cap per customer, so one heavy account can’t blow through the whole budget.

For managers: Ask what one request costs, roughly, and what happens when a single customer uses the feature ten times more than expected.

6. Log everything, ship behind a switch, keep a way back

You’ll want to know what was sent, which model answered, how long it took and what the user did with the result. That log is how you debug complaints, track spending and show an auditor what happened. Mind what goes in it, though: if the inputs contain personal data, the log needs the same protection as the source system.

Release the feature behind a flag. Turn it on for internal users first, then a small group of customers, then everyone. If something goes wrong, you switch it off without a new deployment. Treat model changes the same way. When Notermed changed models, we rebuilt the document index in a fresh store and kept the old one, so we could roll back in minutes.

For managers: Before launch, ask your team to show you how they’d switch the AI feature off for one customer, and how long it takes.

7. Decide what success looks like before you launch

Pick a few numbers that tie back to the job from section 1. How often do people accept the draft as is, edit it or throw it away? How much time does the task take now compared with before? What does the feature cost per month, and per use? How often does the AI produce something a reviewer has to correct for a factual error?

Measure the manual baseline before the feature ships, or you’ll have nothing to compare against. Then review the numbers after a few weeks. A feature that people quietly stop using is telling you something, even if nobody files a complaint.

For managers: Agree on two or three measures and a baseline up front. Without them, the AI feature can’t prove it’s worth what it costs.

The short version

Adding generative AI to existing software works best as a narrow, reversible change: one job, a written data boundary, a person approving the output, tests on real documents, a cost limit and a switch to turn it off. Get those right, and the choice of model becomes one of the easier decisions in the project.

Qasim Ali

Senior Software Engineer (Java, AI), CS2 Technologies

Qasim has 18 years of backend experience, including ten years on the backend platform of Vodafone Kabel Deutschland, and builds AI features that run in production for CS2’s own products and for clients. If you’re weighing a first AI feature for a system you already run, he’s happy to talk it through.