Guard AI Agency, Resist Corporate Lock-In

I’ll be honest: the image of a “transplantable skeleton” for agentic AI — rip the decision-making guts out of one company and stitch them into another corporate body — is seductive. It promises continuity for customers, value retention for investors, and fewer headaches during mergers. Funny thing is, that same image also frames the problem: who gets to perform the operation, and under what ethical anesthetic?

The cio.com piece is right about one thing: AI systems that flatline when a company dies can strand customers, erase institutional knowledge, and break critical services. Anyone who watched customers scramble after sudden SaaS shutdowns knows “corporate mortality” is not a theoretical risk. Designing systems that can survive organizational churn is a reasonable goal.

Look, the trouble starts when “survive” quietly turns into “roam free.” The article leans optimistic about transplantability as a kind of resilience pattern, but skips the ugly plumbing of governance: legal liability, audited provenance, and explicit consent from the people on the receiving end of those machine-made decisions.

Agency isn’t magic; it’s a bundle of design choices. When an “agentic” system is portable, its models, reward functions, and operational heuristics move with it. That persistence can preserve useful capabilities — like a fraud-detection agent that already knows the gnarly edge cases in a banking book — but it can also preserve harms: baked-in biases, undocumented decision rules, and optimization targets that made sense for the seller but are toxic in the buyer’s context. Think of it as cultural contamination at machine speed.

We’ve seen softer versions of this already. When big cloud providers acquire smaller AI startups, they don’t just absorb teams; they absorb model behavior, quirks and all. Those quirks can quietly spread across product lines under a reassuring banner of “shared infrastructure.” A transplantable agentic core just makes that spread faster, cleaner, and harder to unwind.

So if we’re going to build this skeleton, we need more than version numbers on a repo. We need legally recognized audit trails, tamper-evident model lineage, and human-readable summaries that survive ownership changes. Otherwise, accountability evaporates as soon as the org chart does. This isn’t just a tooling problem. Regulators, auditors, and boards should be able to interrogate what exactly moved in the transplant, what stayed behind, and which assumptions are now embedded in the acquiring company’s daily decisions.

Who owns the bones?

Buyers will love transplantable AI because it means faster productization. Investors will love it for exactly the same reason. You can see the pitch deck already: “Buy us, and you don’t just get users — you get a drop‑in AI spine that keeps everything walking.” But market incentives slice the other way too. If infrastructure is engineered to be neatly detachable and reusable, incumbents with the biggest, most mature “skeletons” gain a serious advantage. Smaller firms and new entrants face a higher bar to compete because they’re not just selling features; they’re competing against asset portfolios of reusable machine agency.

That nudges the market toward consolidation, not experimentation. The messiness of greenfield buildouts — the part where people try weird things and sometimes discover better approaches — is exactly what gets optimized out when the safest play is to buy a proven skeleton and wire it into everything.

Labor markets will feel this as well. If an AI skeleton carries policy decisions and tacit expert behaviors across owners, downstream jobs shift from actual decision-making to monitoring, maintenance, and compliance. That’s not automatically dystopian; a lot of people would rather debug workflows than be the person who has to say “no” to a customer all day. But it reshapes work in ways companies rarely disclose when they tout “continuity.” Communities and workers who once anchored their value in institutional memory — the folks who “just know how things are done here” — may find their bargaining power diluted, because the asset that used to live in their heads now ships as part of a transaction.

There’s a historical rhyme here with how manufacturing automation moved expertise from craftspeople into machines and standardized processes. The factory didn’t just make production cheaper; it changed who held power and how knowledge flowed. Transplantable AI makes that shift faster and more transferable: institutional judgment becomes a tradable asset, not a relationship built over time.

Then there’s systemic risk. Imagine financial institutions or health systems passing around portable decision agents through mergers and joint ventures. A harmful heuristic — say, a risk model that underprices certain edge cases — could propagate across the sector faster than governance can react. The article nods toward resilience benefits, but resilience that hinges on opaque, agentic modules can just as easily become a new vector for correlated failures.

Supporters will counter that continuity matters more than these hypotheticals: a transplantable skeleton keeps services running during messy mergers, reduces customer harm from failed integrations, and preserves hard-won institutional knowledge. They’re not wrong. Orphaned models are a genuine operational drag, and “just retrain it” is often fantasy.

Sure, but continuity without constraints is basically continuity of error. If a model’s objectives, guardrails, and training data can’t be interrogated at transfer time, you’re not preserving knowledge; you’re freezing a set of blind spots and shipping them into a new environment. The fix is procedural, not purely technical: mandatory transfer audits, portability standards that tie reuse to explainability, and regulatory checkpoints for sectors like healthcare and finance where “oops” is not an acceptable post‑mortem.

The funny thing is, these kinds of guardrails might actually help the portability vision survive its own hype. Clear rules create a shared expectation: if you acquire an AI skeleton, you also acquire the obligation to open the hood before you flip the switch.

William Gibson once wrote about code and constructs that outlived their creators — ghosts in the machine that haunted whoever inherited them next. Treat agentic AI infrastructure as just another transferable asset on a balance sheet, and we get exactly that: corporate ghosts with opinions, wandering from owner to owner.

My bet: within a few years, the most valuable AI “skeletons” won’t just be the most capable ones, but the ones with provenance and constraints baked so deeply into their design that regulators, buyers, and yes, even workers trust them enough to survive the next round of corporate surgery.