Agentic AI Feels New. The Pattern Is Old: Lessons from Canals for RegulatoryWriting

A breakthrough technology rarely wins because the invention looks impressive. A breakthrough wins
when it makes a costly form of coordination cheap.

I’m a student of history, especially the transformative effects of disease and technology on how societies and commerce function. I read many posts on LinkedIn discussing Agentic AI and there is considerable handwringing in some of those posts. Taking a look at history, I suggest the adoption pattern for Agentic AI will be similar. Transportation (canals, railroads, cars, fiber optics) provide a clean reference case for what happens when a technology reduces friction in moving value across distance.

Canals illustrate the pattern
Canals mattered because they turned water movement into controlled flow. Water always moved, but canals added engineered controls and governance—locks, channels, rules, and tolls—that enabled steady, predictable freight transport across distance. That reliability reduced friction and lowered the effective cost of coordination. Industry reorganized around the new throughput. Commerce soon moved more per unit of time.

Railroads repeated the pattern
Railroads repeated the pattern with speed and reliability. They moved more per unit of time.
Additionally, railroads had a knock-on effect. Rail wear made durability a requirement, not a preference. That demand helped scale steelmaking into mass production. Steel became available in high volume, on quick command that enabled many other infrastructure developments.

Cars repeated the pattern again
Mass-produced cars repeated the pattern with point-to-point mobility. Cars increased destinations per hour. Cars also changed the map of work because commuting time became a variable people could buy
down.

Fiber-optic cables repeated the pattern again
Fiber-optic cables repeated the pattern with data bandwidth. Fiber increased information moved per unit of time and reduced latency. Early builders laid far more fiber than networks needed, so large stretches sat “dark”—installed but not lit—because demand had not caught up. As data traffic grew, operators lit that capacity and throughput rose without re-digging the ground. Global coordination became routine because distance had little to do with what was delivered per unit of time.

Agentic AI belongs in the same category
Agentic AI reduces time (and cost) of moving work across systems. Agents increase completed work per day by reducing handoff time between intent and action. Coordination work becomes cheaper than the work it supports.

Regulatory writing provides a clear view of the shift.

What becomes cheap
Most regulatory teams do not suffer from a lack of expertise. Many teams suffer from expensive
coordination.
Coordination shows up as time spent:

• managing TLFs and definitions
• reconciling inconsistencies
• tracking cutoffs, pooling rules, and analysis sets
• propagating late changes across documents
• answering internal “where did this come from?” questions
• playing around at the surface level of a document

Agentic AI targets these workloads because agents are designed to do three things:

• plan a sequence of actions
• use tools across systems
• maintain memory of what has been decided and why

Drafting becomes only one small output. Work orchestration becomes the larger shift.

The three-layer framework that predicts outcomes
History suggests that durable transformations depend on three layers.
1) Infrastructure layer
Canals needed locks, surveying, financing mechanisms, and governance.
Agentic AI needs:

• permissions and identity
• controlled tool access
• audit trails and provenance
• secure boundaries for confidential content

2) Operating system layer
Canals needed schedules, toll rules, standards, and dispute handling. Agentic AI needs algorithms to
define:

• workflows and handoffs
• quality gates and human sign-off points
• evaluation harnesses that score accuracy and consistency
• change control for prompts, models, and templates

3) Business-model layer
Canals reorganized commerce and rewarded the builders of complements.
Agentic AI may reorganize regulatory work and reward teams that deliver:

• decision-ready argumentation
• fewer late-cycle revisions driven by inconsistency
• reduced reviewer workload

Where we are right now
Most organizations sit in an early “canal” phase: impressive boats, missing locks. Many teams are
experimenting with generative drafting while the foundation remains weak:

• sources are scattered across systems
• document IDs and data lineage may be unclear
• review practices depend on extended manual effort
• governance varies by team and vendor

Those conditions do not block AI. Those conditions amplify risk.
A strong agentic approach begins with structure: define what the agent may access, what the agent may
produce, and what a reviewer must be able to verify.

What will likely happen next in regulatory writing
Several shifts appear likely.

1) Value will migrate from prose to architecture
Sponsors will pay less for “first drafts” and more for:

• claim registries tied to evidence
• cross-module coherence
• consistent definitions and denominators
• controlled update propagation

2) QA and governance will become differentiators
Auditability will separate helpful systems from unsafe ones.
A defensible workflow will require provenance, versioning, and repeatable checks.

3) Agent design will look like information design

• structured inputs
  
• constrained outputs
  
• explicit quality gates

The practical takeaway
Canals did not replace judgment. Canals rewarded builders who reduced friction with consistent water
flow and rules. Agentic AI will likely do the same. Teams that build the complements—provenance,
workflow control, evaluation, and traceability—will define the next standard for decision-ready
regulatory writing.
Join the conversation on LinkedIn: https://www.linkedin.com/pulse/agentic-ai-feels-new-pattern-old-lessons-from-canals-writing-cuppan-fft1c