Agentic AI is an amplifier, not a remedy
Deloitte's warning is unusually direct: there is a risk of layering agents onto broken processes — doing so doesn't fix those processes; it amplifies the challenges. An agent inherits the process it's dropped onto. Point it at an ungoverned workflow — side doors, undocumented exceptions, tribal judgment — and it will execute every one of those flaws autonomously, at volume, without the human who used to quietly catch them.
This is why "we deployed agents" and "the work got better" are two different events. Automation multiplies whatever it is given. A standardized, traceable process multiplied is leverage; an ungoverned one multiplied is risk — now moving faster than anyone can see, and harder to trace after the fact.
The pilots are everywhere. Production is rare.
The enthusiasm is near-universal; the operating-model readiness underneath it is not. The gap between the two is where agentic programs stall.
of companies haven't redesigned work to fit AI — even as automation expectations run high. They're automating jobs built for humans, unchanged.
are actually running agentic systems in production — 30% exploring, 38% piloting, just 14% with deployable solutions.
The gap isn't the models — it's the model underneath. 93% of AI budgets go to technology and just 7% to workforce and readiness; roughly 80% still lack a mature governance model for autonomous agents. Deloitte's metaphor lands it: bolting agents onto an operating model built for humans is "like fitting a jet engine to a bicycle."
The trap isn't new with agents. The process-automation field — RPA, BPM, intelligent automation — has long found that automating an unstandardized process scales its errors rather than removing them; agentic only raises the cost, because it takes out the human who used to catch them.
Automatable in parts
The trail exists but the flow varies — automate the stable steps, standardize the rest first.
Ready for agents
Governed and provable — automation and agentic AI compound a model that works.
Scales the dysfunction
Where most work lives today. An agent here runs the side doors faster and hides the flaw.
Fast, but unprovable
Consistent yet opaque — outputs no one can trace to a governed origin. Make it followable.
Automation only pays in the top-right — standardized and traceable. Drop an agent on the bottom-left, where most work actually sits, and it scales a flaw no dashboard will show. Standardization moves work rightward; the Traceability Ratio moves it up.
Standardize first — then point the agents at it
Deloitte's guidance to leaders is to map how work should run before automating how it does run — "don't simply pave the cow path." An agent laid over an ungoverned process paves it at machine speed. The order is the whole safeguard: standardize and make the process traceable, then let automation amplify a model that works.
- The agent inherits every side door
- Exceptions once caught by judgment are missed
- Actions leave no traceable audit trail
- Governance is retrofitted after an incident
- Dysfunction now runs at machine speed
- One governed path the agent can follow
- Exceptions codified as explicit rules
- Every action traces to a governed origin
- Oversight designed in, not bolted on
- Automation compounds a model that works
Deloitte's own read of the winners: the organizations succeeding with agents took a process-first approach, redesigning the work around AI rather than dropping agents onto outdated processes. Standardization and traceability are the prerequisite — and, for agents that act without a human in every loop, they are also the safeguard.
GSDPI, applied to an agentic program
The same five stages that read the operating model also make it ready to automate — putting the standardization and traceability work before the agents, not after an incident.
-
G
Get Intake
Standardize the front door before you put an agent on it. An agent at an ungoverned intake just automates the side doors — so the governed path has to exist first.
-
S
Sort Decision rules
Codify the decisions a human used to make by judgment into explicit, governed rules. Otherwise the agent invents its own — and the tribal knowledge that caught edge cases is exactly what it lacks.
-
D
Do Execution
Give the agent a single governed execution path, with every action logged. Autonomy is only safe when the route it runs is the one you standardized — not a shortcut it discovered.
-
P
Prove Audit trail
The Traceability Ratio becomes the agent's audit trail: any output can be followed back to the work and the rule behind it. Without it, agent results are reported, not proven — at machine speed.
-
I
Improve Oversight loop
Govern how agents learn and where humans stay in the loop — oversight as a standing role, not an afterthought. The model improves from a governed baseline instead of compounding an ungoverned one.
New Insights, the day they publish.
The operating-model reads behind ERP, AI & services — a short note when one’s worth your time.
New posts & Insights only. Unsubscribe anytime.
The readiness gaps — point by point
What executives experience as "the agents didn't deliver" resolves, on inspection, into specific gaps in the operating model underneath. The six that carry the most weight:
The process was never standardizedGet
The workflow the agent is dropped onto still runs through side doors and undocumented exceptions. Automation doesn't clean that up — it executes it faithfully, at volume. The first requirement of automation is a process worth automating.
Exceptions lived in human judgmentSort
The edge cases were handled by an experienced person applying unwritten judgment. That tribal knowledge is precisely what the agent doesn't have — so unless the rules are codified first, the agent mishandles exactly the cases that mattered most.
No traceable audit trailProve
You can't follow an agent's output back to a governed origin — which work, which rule, which input produced it. That's "reported, not proven" at machine speed. When something goes wrong, no one can reconstruct why, and trust in the whole program erodes.
Governance retrofitted, not designedProve
Roughly 80% of organizations lack a mature governance model for autonomous agents — clear decision boundaries, real-time monitoring, audit trails. Retrofitting oversight after an incident is slower and costlier than designing it in from the start.
Budget skewed to technologyImprove
93% of AI budgets go to technology and just 7% to workforce readiness. The model-work — standardizing, governing, redesigning roles — is the part that decides success, and it is the part that goes unfunded.
The work was never redesignedAll
84% of companies automate jobs designed for humans without rethinking how the work should be done — Deloitte's "jet engine on a bicycle." The frame was built for people; bolting an agent onto it stresses every joint that was never meant to run at that speed.
How ETEGY reads it
Every gap above shares one root: the operating model wasn't standardized or traceable before the agents arrived. We make it ready — read the actual state, standardize and govern the process, make it traceable — so automation amplifies a model that works instead of scaling one that doesn't.
A Zero-Based read standardizes and governs the process across the GSDPI lifecycle — before a single agent is deployed. The Traceability Ratio then becomes the agent's audit trail: every autonomous action follows back to a governed origin. Standardize, make it traceable, then automate — the order is the safeguard. See the ZBT Discovery →
The Automation Readiness Scorecard
Rate your operating model against the five things that must be true before you automate — scored by GSDPI stage, with a readiness verdict. Built to circulate: take it into your AI steering committee.
Zero-Based Transformation — be first to read it.
Be first to know when Zero-Based Transformation™ publishes.
Occasional launch updates only.