The greatest value of generative AI is not always in creating new products and agents. Modernizing legacy systems may be the key to bringing forward the return on your AI investment.
Someone on your team sees a new way for customers to restock through the ordering portal. Or a new flow to increase conversion in the app. Or a way to communicate delivery delays that reduces complaints. In two weeks, a prototype is ready. The leadership team sees it, likes it, approves it. The architecture committee approves it. From then on, the opportunity enters the IT queue. Integration with the system that already runs the business. Security review. User acceptance testing. The deployment window. Each step exists for a good reason, and none of them is fast. Months later, the opportunity has still not reached a single customer. When it does, your focus and energy are already somewhere else.
If opportunities like these take more than two weeks to reach your customers' hands today, your company is not an exception. But it is one or two steps behind those leading the market.
What if your AI investments were channeled so that your ideas, however simple or complex, were always one week away from working in your customers' hands?
Where does AI come in: in the new project or in the system that already runs?
Today, these investments have a well-known destination. Ask any executive what the company is doing with AI and the answer comes ready: an assistant in customer service, an agent that reads contracts, a copilot for the sales team. What the market announces every week pushes any leadership to have projects like these, and they need to exist. The problem is when they move forward at the cost of not accelerating, optimizing or modernizing the systems that already run the business, the legacy that sustains that operation. Each new project takes months to discover the business rules nobody documented, the integrations nobody mapped, the adjustments the system needs to work well. And, because they put generative AI in front of the customer, they still need to discover which guardrails to build so the model does not say or do what it should not. The queue of AI projects grows, becomes more and more aspirational, and the results keep getting pushed to the next quarter.
Meanwhile, the rest of the operation keeps the same pace, with the same cost and the same inefficiencies as always. The e-commerce app that brings in most of the revenue has its backlog moving at the usual speed. The B2B ordering portal that processes millions every month still needs the same squad just to stay up. The sales app that the sales force opens every day gets one improvement per quarter. The loyalty program stays in the background for lack of technical hands. These systems, the legacy that already runs the business, remain as they were: the same code, the same change pipeline, the same steps and the same timelines.
To the board, everything seems fine: the number of AI projects in the portfolio has doubled. But the time between the idea and the customer remains the same. So does the cost of sustaining what already exists, and none of the new projects was designed to bring it down. And AI is being used to keep up with the hype, not to increase results and streamline the operation.
An executive who has worked at three of the largest consumer brands in the world summed this up for us in one sentence: "people say they are innovating, but when you look closely, nothing is actually changing". What is missing is to stop and understand how to apply, in the most effective way, what technology already allows today.
What changed to make this possible now
Now reverse the scene. The business team details what it wants. A pipeline of AI agents updates the system that already runs the business in hours or days. The following week, the customer is already using it. Your team asks, the pipeline delivers. This is already possible today, in most projects, and not only here. At Taqtile, systems that took us three, four or six months to develop are being rewritten in two weeks. Three main factors made this possible.
The first is the leap in how well models generate code. On SWE-bench Verified, the most widely used public test for real software engineering tasks, the best models solved about 2% of the tasks at the end of 2023. At the end of 2024, close to half. At the end of 2025, about 80%. In 2026, they passed 90%, and the test was retired for having become too easy. Other capabilities also advanced. In tests that simulate a customer service agent, such as τ²-bench, models went from less than half of the tasks in 2024 to close to 80% in 2026. That is a lot, but it is not the same leap. The capability that advanced the most in this generation of models is not talking to the customer. It is building and rebuilding software.
[image: chart of the year-by-year evolution of SWE-bench Verified, from 2023 to 2026, with the leading models of each year in the chart and in the alt text]
The second is the pipeline around the agents. Leading engineering teams are rewriting legacy systems with a fraction of the effort it used to take, because agents now work inside a support architecture and good development practices: specification before code, tests the agent itself must pass, adversarial tests that try to break what was done, code review and security checks. This is what we call a harness: the set of rules and checks that tells the agents how to develop, verifies what they did and blocks what does not pass. This validation is more robust than the one that existed when the original app was built. A system that needed a squad of six or seven people to be sustained now needs one or two engineers.
The third is that the legacy system is the best input a code generator can receive. The e-commerce app or the B2B platform that has been running for years has business rules defined and validated by time in use, mapped integrations, known bottlenecks and users who complain the next day when something changes. The new project has none of that. It has hypotheses. And governance weighs less: AI-generated code goes through an engineer, tests and validation before it reaches a customer. Errors show up in the pipeline, not in front of the buyer. It is the lightest form of AI governance that exists today. Generative AI talking to the customer is harder to control and takes longer to generate results. Does it make sense to devote most of the energy to it, when there is a much more predictable path to results in the short term?
Where the rewrite of legacy systems begins
The first step is to choose the system, not the AI project. Whoever wants to shorten the time between the idea and the customer looks at the ordering portal, at the sales app, at the system the next idea will have to go through, and only then at the AI roadmap. It is also the lowest-risk choice. The new AI project is, by definition, the uncertain bet: nobody knows yet whether it will work, or for whom. The legacy system is the opposite. It is validated, has customers using it and generates revenue every day. The assistant and the agent come next, with less effort, because they are born on top of a foundation already rebuilt for it.
The second step is to map and decide with the same AI. Surveying the process the system serves, the rules it applies and the integrations it touches, comparing alternatives and deciding: each of these steps has also become shorter with supporting AI. What used to take weeks of meetings now fits into days.
The third is to set up the pipeline and take care of its quality. The objection is well known, and for years it was true: fast means poorly done. But the trade-off between speed and quality no longer exists in this pipeline, and when the trade-off disappears, so does the reason to protect the old pipeline. This is how we work today. We build the pipeline together with the client's team. The code, the specifications and the architecture belong to them. The following requests, their team asks and the pipeline delivers, with fewer people and less time than before. Our technical team stays where it makes a difference: taking care of the quality of what goes out and evolving the harness as models and practices change, which happens every few months. Nobody is locked in to us, to our harness, to a third-party platform or to an AI model. It is worth asking the same of whoever builds for you.
How do you measure the return on AI today?
AI adoption stops being measured by the number of projects in the portfolio. The measure becomes the time between an idea and a real customer, in the systems that already run. The question worth taking to the team stops being "which AI project will we do this year" and becomes "which of our systems, rewritten, would get the next ideas to the customer faster". For those responsible for the budget, a second question is worth asking: how much will the cost of maintaining and evolving our main systems drop in the next twelve months? If the answer is nothing, or nobody knows, the company will pay next year the same it paid this year to maintain what it already has. In your operation, which would be the first system to rebuild: e-commerce, the B2B portal, the sales force app?