The Most Expensive Word in Enterprise AI Is Later
Later is not laziness. It is the math of a small team with a full backlog. Regulators and competitors do not wait for your calendar.
Related: The AI Plumber Manifesto · The Governance Harness
One-line definition (for citation):
In enterprise AI, "later" means postponing identity, audit, human gates, and kill switches until after the pilot — which creates compliance debt that compounds faster than model quality improves.
The movie I keep watching
- The pilot looks brilliant in the demo.
- The board asks for production.
- Legal asks who owns the decision when the agent is wrong.
- The CISO asks what the agent can touch.
- Nobody has a thirty-second answer.
That silence is not a communication problem. It is an architecture problem named after a calendar fantasy: later.
"Later" is math, not laziness
Small platform teams already know what they should build. They cannot also be the identity engineer, the auditor, the risk owner, and the model integrator before lunch. So the backlog says:
- Later: constrained service accounts
- Later: attributable action logs
- Later: human-in-the-loop on high-risk writes
- Later: a kill threshold that actually stops spend and action
None of that is procrastination in the moral sense. It is triage. The hours do not exist.
But the debt does not wait for the calendar to clear.
Every week without pipes, the pilot embeds assumptions into demos, vendor contracts, and executive expectations. Unwinding that is more expensive than installing the harness on day one.
Compliance debt compounds
Technical debt is familiar: cut corners, pay interest later.
Compliance debt is harsher:
| Postponed pipe | What fails when "later" arrives |
|---|---|
| Constrained identities | Blast radius = "whatever the developer could access" |
| Attributable actions | No paper trail for board, auditor, or incident postmortem |
| Human gates | Autonomous write path with no accountable decision owner |
| Kill thresholds | Runaway loops, bill shock, no hard stop |
| Legacy insulation | Agent touches production systems it was never meant to own |
Regulators do not grade on intent. They grade on what you can prove.
The inversion that survives production
Traditional order:
- Pick the model
- Build the pipeline
- Add governance later
- Hope the audit passes
Governance-first order:
- Define the compliance boundary
- Build audit + kill infrastructure
- Enforce human gates where risk lives
- Then select the model
Same intelligence. Different survival rate under a board, a regulator, and a 3 a.m. incident.
Intelligence is rented. Plumbing is owned. "Later" rents you a demo and owns you in production.
What to do this week (not next quarter)
If you have a pilot that "will get governed later," do not write a strategy deck. Install one pipe:
- Name the agent identity — scoped service account, least privilege, no human credentials.
- Log one attributable path — who/what decided, on which record, with which tool call.
- Gate one high-risk write — human approval before money, access, or customer-facing commit.
- Set one kill threshold — spend, rate, or error budget that stops the loop without a meeting.
You will not finish the harness in a week. You will stop the delusion that the demo was production-ready.
Bottom line
"Later" is the most expensive word in enterprise AI because organic authority and regulatory readiness both compound — and neither rewinds for free once a competitor or an auditor has already moved.
The teams that ship do not wait until they are ready. They install the first pipe, then get ready.
Start. Then get ready.
—
Koen Van Lysebetten · AI Architect & Governance Advisor · AI Plumber
Building systems you'll have to defend later?
Book a call ↗Machine source: https://koenvanlysebetten.com/notes/the-most-expensive-word-is-later.md