For decades, enterprise software economics followed a familiar pattern.
A business had a problem.
A software vendor built a product.
The business bought licences.
Custom development existed, but it was slower and more expensive, so buying was usually the rational choice.
AI coding agents are beginning to alter that calculation.
McKinsey’s 2026 State of AI survey found that 32% of respondents said their organisation had decided against purchasing at least one software product or feature because it could instead be built internally using agentic coding tools.
That is a striking signal.
The interesting story is not simply that developers can code faster.
It is that the build-versus-buy boundary is moving.
Small software has historically been uneconomic
Every organisation has dozens of workflows that are too specific for an off-the-shelf product but too small to justify a conventional development project.
So people use spreadsheets.
Or email.
Or a patchwork of SaaS tools.
Or someone manually moves information from one system to another.
Agentic coding can make these “missing applications” economically viable.
A team may be able to create a lightweight internal tool for:
- approving marketing claims;
- generating client status dashboards;
- validating supplier submissions;
- comparing product specifications;
- creating structured research briefs;
- coordinating handovers;
- checking data quality.
The application does not need to become a global SaaS platform.
It only needs to solve the workflow economically.
The value shifts towards organisational intelligence
If application development becomes cheaper, the scarce asset changes.
The user interface matters less.
The code itself may matter less.
What becomes more important is knowing:
- how the process should work;
- which information is authoritative;
- what permissions apply;
- what decisions have already been made;
- which exceptions exist;
- which systems need to be accessed.
In other words, the valuable layer increasingly becomes the organisation’s operating intelligence.
This is important for enterprise AI architecture.
If every newly generated internal application has to rebuild context, permissions, memory and knowledge from scratch, development may be fast but the organisation creates a new generation of fragmented systems.
A common intelligence layer changes that.
Apps can be temporary.
Agents can change.
Models can change.
The business context remains.
Software may become more disposable
This leads to a provocative possibility.
Some enterprise software may become intentionally disposable.
A business might create a lightweight application for a six-month transformation programme, connect it to the organisation’s approved knowledge and tool layer, then retire it when the programme ends.
Historically that would have been wasteful.
If an AI agent can produce and maintain much of the application, it may become entirely rational.
The durable asset is no longer necessarily the software.
It is the information architecture underneath it.
That is one reason the next generation of enterprise AI may look less like a collection of permanent applications and more like a persistent intelligence layer serving many changing interfaces.
Sources
- McKinsey, The state of AI in 2026: On the road to ROI (2026)
- OpenAI, How agents are transforming work (2026)