In my last piece, I wrote about the forward deployed engineer (FDE). TCS’s recent announcement that it intends to create up to 8,900 FDEs is one of the more revealing developments in enterprise AI this year.
The number matters less than what it represents. Across consulting firms, hyperscalers and technology services companies, a race is on to create professionals capable of mediation between increasingly sophisticated artificial intelligence (AI) systems and complex enterprises. The last mile, so to speak.
Palantir institutionalized the idea by embedding engineers in client organizations rather than leaving them at the receiving end of a ‘requirements’ document. The premise was simple: enterprise AI projects fail less because of technology than because businesses and technologists rarely define the same problem.
For much of the last few years, enterprise AI has been constrained by understanding. Data was fragmented, processes were undocumented, incentives conflicted and business users described symptoms rather than causes.
The FDE emerged because someone had to bridge those worlds, sitting close enough to the client to hear what was actually being asked, and close enough to the model to know what it could actually deliver.
The bottleneck is already shifting. Enterprises are no longer asking if AI can automate work, but deciding which decisions AI should participate in, how much authority it should receive, where human oversight genuinely matters and where it merely creates the appearance of control. That last distinction is the one that most corporate boards are not yet equipped to draw.
I have seen multiple startup lenders enter the private credit market with the aim of automating large parts of the credit decision itself. These models work.
The unit economics are hard to argue with. But the real question is whether an institution understands how thousands of individually sensible lending decisions might gradually alter the risk profile of an entire portfolio. No single loan looks wrong, but the portfolio might be unrecognizable in a few years, and nobody signed off on that.
An engineering team confirms the architecture. Finance executives demonstrate return on investment. Risk assessors validate the portfolio’s past performance. The legal department signs off on compliance. Every function does its job. Yet, the decision may be wrong.
Perhaps the model gradually shifts the institution’s effective risk appetite. Maybe credit officers stop questioning borderline recommendations because the model is usually right. Profitable borrowers might quietly disappear because the model has optimized lending for an objective no human intended. These ‘failures’ are not technical but arise because the organization optimized one objective while changing another.
We increasingly assume that good decisions are additive. Combine enough expertise, enough data and enough computational power and judgement will emerge. The FDE embodies that philosophy.
The role combines engineering with business understanding. Others are now trying to combine AI expertise with domain knowledge. Each new hybrid role is built on the same assumption: that stacking up a set of competencies produces wisdom.
This creates understanding, but whether it leads to judgement is doubtful. The distinction matters because understanding grows by adding perspectives while judgement improves by minimizing assumptions.
Judgement is subtractive. Good investors outperform the pack because they eliminate poor investments. Experienced physicians diagnose difficult cases by discarding explanations that do not fit the evidence. These professionals succeed in their jobs by refusing to accept an explanation simply because it is coherent.
Enterprise AI is moving rapidly into this territory. As models become more capable, organizations are drowning in plausible answers. AI analyses hang together, financial models tell a convincing story, every recommendation appears defensible and a dashboard that once forced a choice between two explanations now offers five, all internally consistent and with confident fluency, leaving a manager under deadline pressure with no structural reason to prefer one over the others.
In an earlier column, I argued that enterprises would increasingly require ‘forensic plumbers,’ or people capable of tracing failures through increasingly complex AI-enabled systems. Their job would be to get rid of plausible explanations till clarity is attained.
The same capability will be valuable before AI deployment. Waiting for the postmortem is an expensive way to discover a lack of judgement. The next scarce skill in enterprise AI will probably be the ability to clip out bad assumptions before organizations commit to a digital rejig.
As AI gets embedded in the operating fabric of organizations, acquiring a competitive advantage will depend less on building systems than on deciding which systems must be built.
Companies that consistently discard weak assumptions before they become expensive commitments are likely to fare better, as would those that reward sceptics who ask inconvenient questions as readily as they reward engineers who deliver working models.
While understanding grows abundant, judgement is becoming scarce. This is the reality businesses must confront.
The author is a technology consultant and venture capitalist.