For many years, that's how business software was developed: Spot a problem, find a piece of software that would solve it, and add it into your tech stack.
It is no longer that easy for AI agents to do that.
An agent can be used across applications, extract information, reason, and complete multi-step tasks. That's led to a question many technology teams are asking themselves right now:
Should we develop an AI agent? Should we put an AI agent into our current software? Or completely replace some of our software with an AI?
There is no one way. The answer will differ depending on a business's starting point, what work the agent has to do and how AI can make a valuable difference.
**
AI Is Already Changing How Software Gets Built
**
Organizations are no longer experimenting with AI. In McKinsey's 2026 State of AI survey, 40 percent of respondents from companies with revenue of more than $1 billion reported that they were deploying AI agents, up from 27% a year earlier.
And there's one more change taking place along with that adoption.
According to McKinsey, 32% of those surveyed said their organizations had chosen not to buy at least one software product or feature due to the availability of agentic coding tools that can be used to develop the functionality in-house.
That doesn't mean traditional software is disappearing.
It is another way for companies to look at things as far as technology goes: Can we do this ourselves with AI?
**
When Building an Agent Makes Sense
**
It's most logical to build an agent from scratch when the entire workflow is tightly coupled with the business.
For instance, a firm might discover that it has an internal procedure that crosses several systems, contains confidential information, and requires special rules that can't be accommodated by a package product.
In that case, creating can provide more command over:
Business logic
Data access
Permissions
Integration with internal systems
Agent behaviour
User experience
But building doesn't mean building from scratch. On top of the ability to build your own AI you should know that: A modern approach is to take advantage of other models, APIs, databases, tools, and internal systems instead of having to build your entire AI stack from scratch.
That's where the skilled AI Agent development teams can assist you and guide you in deciding what actually needs to be built from scratch and what can be reused.
**
Integration Is Often the More Practical Starting Point
**
You don't need an AI solution for every business.
Every once in a while the software that's already out there covers virtually everything the business wants. The pièce de resistance, in this case, is known as intelligence or automation.
In that case, perhaps it's a better choice.
You could have an AI agent as a wrapper around your existing CRM, ERP, support, analytics, and knowledge-management systems communicating via approved tools and APIs.
This method may be less disruptive than forcing employees to leave behind their familiar software.
The agent added another layer that helps people connect with those systems.
However, integrations aren't without their own challenges. Merely an agent having connectivity to 5 different business systems doesn't make it immensely helpful.
Bad data, fuzzy roles and not-so-clean APIs and workflows will only make the experience worse.
According to McKinsey, two-thirds of organizations across the globe have already tried out AI agents; however, less than 10% have successfully rolled out and scaled them. Eight out of 10 organizations in McKinsey's research pointed to lack of data as a key challenge to scaling agentic AI.
**
When Replacing Software Starts to Make Sense
**
Replacement has the highest potential for disruption so it should have the highest bar.
There is the possibility that an outdated process could have been built on manual steps that no longer are sensible. An AI agent added at every step may just add a layer of complexity around legacy.
Other times, the optimal option is to reconsider the flow.
Rather than require the agent to open three different apps, then copy from one and paste into another, then check a database, and then update a record, the new system could enable an agent to orchestrate much of that from a single interface.
This is where AI takes over the software problem.
The question is not always "Where can we get an agent?"
Or is it: "Do we still want our workflow to function like it does now?"
**
The Decision Should Start With the Workflow
**
Teams should not be asking: Build or buy an AI agent?
They should be asking: How do I do this?
Examine your workflow.
Ask yourself:
What's getting in the way?
Where does judgment come into play?
What data lives in what systems?
What can be automagically done?
Where does it still make sense to have human input?
What do we do when the agent is mistaken?
These are the questions that tell you whether you need a custom agent, an integration, or an overhaul of your software. Technology architecture should flow from that assessment-not the other way around.
**
AI Agents Don't Automatically Make Existing Software Better
**
It is tempting to view an AI agent as a one-size-fits-all-upgrade for your current apps.
But increasing autonomy to a badly-designed process doesn't solve the problem.
In some cases, it can amplify it.
An agent can shift data between systems more quickly, but if the data itself is inconsistent, the business will end up with wrong answers more quickly. If an agent has too many privileges, it can lead to security and governance issues.
According to a 2026 survey by IBM, just 11% of tech executives reported they were very confident about how well they were preparing for the extent of AI-agent deployment they anticipated over the next year. The IBM survey revealed that 77% said they were falling behind AI adoption.
This is one of the reasons why architecture and governance are features of the AI itself, not something to be dealt with after deployment.
**
Build, Integrate, or Replace?
**
There is no clear winner among the 3 methods.
Build when the workflow is very specific, strategically important, or if existing products cannot deliver needed capabilities.
Integrate when existing software is largely functional and the addition of an AI agent brings value in automation or intelligence.
Replace when the current application is becoming more restrictive than useful and AI offers a chance to rearchitect how the work is done.
The key is don't just implement technology because it's new.
AI agents are revolutionizing what software can do, but not everything that is built can (or should) be reimagined for them.
The better question is simpler:
What would be the software experience if an AI agent could do more of the work?
For companies that need to outsource work to develop their AI, answering that question before settling on an architecture will save a lot of time and investment in software that won't meet people's needs.
Top comments (0)