Most enterprises have an AI agent running somewhere. Ask what changed and you'll usually hear about a step that got faster: a summary drafted, a ticket triaged. Then ask what happened next. The agent waits for a system it can't call, a record someone reconciles by hand, an approval sitting in email.
The research backs this up. MIT NANDA found that 95% of organizations saw zero return from their generative AI pilots (Virtualization Review). Deloitte found that 34% of organizations are using AI to redesign how they operate at a deep level (Deloitte).
In a recent Techstrong webinar, Futurum's Alastair Cooke asked Mactores CEO Bal Heroor why. Bal compared it to cloud in 2011, when companies ran the cloud as a remote data center and paid more for the same thing. AI is in that phase now. Teams are using agents to speed up the work they already do, on systems built decades ago. Alastair put it simply: "People are looking for a faster horse rather than trying to invent a car."
You can't modernize what you can't see. In the session's poll, fewer than one in five attendees said they could produce a trustworthy dependency map of their most critical application in the next 30 days. The people who knew how those systems worked have often left.
The instinct is to collect every document and hand it to an AI. Bal's advice is the opposite: start with the code that actually runs. That means stored procedures, application code, configuration and infrastructure code. Documentation still goes in, at a much lower priority, because outdated documents mislead the model.
The result is a knowledge graph that traces every business rule from the user interface to the database tables it touches. It stays current as the code changes, so the target doesn't move during the project.
At The New York Times, that approach found rules in the PLATO subscription engine that the team didn't know existed. The Times had tried to move PLATO off Oracle five times in six years. Mactores delivered it in 12 weeks against a 12–15 month manual estimate, across roughly 6 million lines of code, with 100% row-level validation and zero revenue impact.
Mactores deploys Aedeon, the agent platform built by Mactores' founders' sister company, across most of its engagements. Aedeon absorbs approximately 60–70% of the hours a traditional consulting proposal of the same scope would staff. That share is pattern work: discovery, dependency mapping, code and schema conversion, test generation and validation.
Everything that needs judgment stays with Mactores' forward-deployed engineers: target architecture, course corrections, security posture and the cutover decision. On the New York Times program, Aedeon carried codebase assessment, schema conversion, SQL remediation and validation across about 8,400 files. Engineers owned the complex logic, cutover and high-availability readiness.
The audience drew the same line in the polls. Many were comfortable letting AI handle discovery or test generation unsupervised. Almost nobody would hand over production cutover.
Bal named a common blocker: teams trying to describe their whole estate to a model in page after page of written context. No context window holds an enterprise, and models get less reliable as context grows.
The better approach stores business rules in a knowledge graph as hard constraints the model can't improvise around. Automated tests then close the loop. When an Oracle-to-PostgreSQL conversion maps a number to the wrong type, a known pitfall AWS also documents (AWS Database Blog), the failing test sends the agent back to fix it before an engineer ever sees it.
Bal estimates that development, testing and validation through user acceptance make up about 80% of a typical modernization program's cost. Agents now carry most of that work. The budget that's left goes to the human-owned work where the risk sits: security posture, landing zones, integration and cutover.
The economics show up in Mactores' published work. For Seagate, Mactores cut the factory data platform's annual operating cost from $2.4 million to $900,000 and engineering effort from 21,000 hours to 4,500.
One more rule: check before you spend. Once an AI run starts, the compute cost is gone. Review the context, the direction and the rules before every large run, not after.
Agents aren't the problem. The systems they have to reach through are. Alastair closed on the standard that matters: "The most important thing is what's actually happening in production."
Watch the full session on demand.