Each client has unique needs, leading the product to accumulate edge cases rather than enhancing its intelligence.
Every custom request serves as a data point, and every workaround created by an experienced engineer highlights a recurring issue that hasn’t yet been clearly defined. Each correction to an AI output indicates where institutional knowledge resides and what it would take to encode it.
The platform should emerge from this foundational work, not precede it.
Successful builders begin by closely engaging with the problem, working consultatively, and understanding how the same objective can lead to different execution paths across various clients. This crucial phase is where the patterns reside; it cannot be skipped in the pursuit of scalability.
Once you can identify the patterns, you can extract the underlying principles. The platform crystallises the insights gained from this work.
The same reasoning applies to firms on the buying side of AI solutions.
When an AEC firm acquires an AI tool, it’s not merely licensing software; it’s contributing its institutional knowledge to a system it does not control. Every correction made by an estimator, and every workflow adjustment applied by a project manager, serves as training data. This information leaves the firm and accumulates elsewhere.
The agent layer introduces a more structured approach. Currently, when two agents exchange information, here’s what typically happens: Agent A completes a task and sends the result to Agent B. Agent B processes this information based on its own optimisation criteria, not according to the recipient’s preferences. The human user often learns what transpired only after the fact, if at all.
What an agent transmits is never just raw data; it’s always a distilled and opinionated interpretation of that information.
A2A, MCP, and ACP are working on developing the communication channels, but without meaningful semantics, these channels are merely faster versions of email. The real question is not how quickly agents can communicate, but whether the receiving human actually needed the information at that moment and in that context.
It’s not just about the “what,” but also the “when” and the “why.”
Firms that build this architectural layer early won’t just become better AI users; they will also be the ones unwittingly training other firms in the industry.
I wrote about the buy-vs-build decision for Last Week In ConTech.
