By James Proctor, Co-Founder and Managing Director, The Inteq Group
Time to value from agentic AI improves dramatically when organizations redesign processes to work with data where it currently lives, rather than delaying agent initiatives behind major data re-platforming programs.
Well-designed agent-enabled processes unlock existing data assets, fragmented, imperfect, and distributed as they are, because agents consume data in context at the moment of decision rather than requiring it consolidated in advance.
The practical implication is significant for any executive weighing sequencing: the multi-year data consolidation program most AI roadmaps treat as a prerequisite is, in most cases, neither a prerequisite nor the best use of the next two years.
This paper explains why the prerequisite belief exists, why it fails economically, and what disciplined organizations do instead.
Why Do Organizations Believe Data Consolidation Must Come First?
The consolidation-first instinct is not irrational. It is inherited. For thirty years, extracting value from enterprise data genuinely did require gathering it: warehouses, marts, lakes, and master data programs existed because analytics tooling could only produce insight from data that had been collected, cleansed, and conformed in advance. An entire generation of technology leadership learned, correctly for its era, that the single source of truth comes first and the value comes second.
Agents break the assumption underneath that playbook. An analytics platform aggregates data in advance to answer questions in general. An agent retrieves data in context to resolve one case in particular: this invoice, this order, this claim. It can query the system of record where the record actually sits, reconcile what it finds across sources, and act, the same way a capable employee works across four applications today without anyone first merging those applications. The orchestration happens at the decision layer, above the systems, not inside a consolidated data layer beneath them.
This is the structural reason time to value no longer has to wait for the platform program. The warehouse-era playbook solved a real problem for a different consumption model. Applying it to agents is solving yesterday's problem at tomorrow's prices.
Why Is the Re-Platforming Reflex the Most Expensive Assumption in AI?
Walk into most enterprise AI programs and you will find a familiar sequencing chart: eighteen to thirty months of data foundation work, followed by the AI value phases. The foundation phase is where the reflex lives, and its cost is routinely misjudged because only one of its three components - the visible component, program spend - appears in the business case.
The second is delay: every quarter of foundation work is a quarter of forgone cycle-time compression, capacity reallocation, and quality improvement in the processes the agents would have served.
The third is risk of futility: large data consolidation programs have a long, well-documented history of overrunning, descoping, and occasionally never finishing, which means the AI value gated behind them inherits that delivery risk in full.
Add the three together and the re-platforming reflex is frequently the single most expensive line item in the AI strategy, and the only one that produces no direct business outcome.
None of this argues that data infrastructure investment is wasteful. It argues against the word first. Infrastructure should be justified by its own economics and pursued in parallel where warranted, not positioned as the toll gate through which all agent value must pass.
“The most expensive line item in most AI programs is the two years spent waiting to start.”
How Do Agent-Enabled Processes Work, With Imperfect Data?
The re-platforming argument rests on a premise that deserves direct examination: that agents require clean, consolidated data to function. Well-designed agent-enabled processes are built on the opposite premise. They assume data will sometimes be incomplete, inconsistent, or stale, and they treat that condition as information the process uses rather than a defect that halts it.
The design pattern is straightforward. Where the data behind a case is strong, the agent proceeds. Where it is weak, the weakness itself becomes a routing signal: the case is handled more conservatively, supplemented from another source, or surfaced to a person. The placement mechanics of those human touch points are a discipline of their own, addressed in a companion paper in this series. The point that belongs here is the economic consequence: data imperfection stops being a program-level blocker and becomes a case-level condition, managed in the flow of work at the moment it actually matters.
There is a compounding benefit that consolidation-first programs never capture. Agents working across live operational data surface specific, prioritized evidence of which data problems actually impede which decisions. Most organizations discover that a modest subset of their assumed data issues accounts for nearly all the operational friction, which converts the boil-the-ocean cleanup into a short, targeted remediation list ranked by business impact. The process, in effect, produces its own data strategy.
“Data problems you can see and route around are managed. Data problems hiding behind a future platform are not.”
Why Is Process Design the Cheaper Lever?
Executives allocating an AI investment are choosing between two levers that both claim to unlock value from data. The infrastructure lever is denominated in years and millions: platform licensing, migration, integration, conformance, and the organizational change that follows.
The design lever, redesigning a critical process around its decisions and pointing agents at the data that exists, is denominated in weeks and workshops. When both levers plausibly lead to value, elementary capital discipline says exhaust the cheap lever first, and in this domain the cheap lever is also the fast one.
The design lever carries a second property that rarely makes it into the comparison: it is reversible and informative. A process redesign that underdelivers teaches the organization something specific at modest cost.
An infrastructure program that underdelivers is sunk at scale. In portfolio terms, design-first sequencing buys information cheaply before committing capital expensively, which is precisely the option logic executives apply everywhere else and routinely suspend for data programs.
Why Does Faster Time to Value Protect the Initiative Itself?
There is a political economy to enterprise AI that sequencing charts ignore. Executive sponsorship, board patience, and funding are not fixed endowments. They decay, and they decay faster for AI than for most investments because expectations were set high and skeptics are watching. A program that spends its first two years building foundations produces exactly what skeptics need: a large spend line with no operational result attached to it. Many AI programs do not fail technically. They are defunded mid-foundation, which is recorded, unfairly but permanently, as AI not working here.
Delivering measurable operational outcomes in early quarters inverts the dynamic. Results convert sponsors into advocates, arm the CFO with evidence rather than projections, and earn the standing to fund the longer-horizon work, including whatever data infrastructure genuinely deserves investment. Speed to value is not impatience. It is how a multi-year transformation finances its own credibility, and it is the most reliable insurance against the gap between a promising pilot and a funded production program.
The Data-Readiness Excuse Is the Most Respectable Way to Do Nothing
Here is the observation that will land uncomfortably in some steering committees, and it needs to be said. Our data is not ready has become the most socially acceptable sentence in enterprise AI. It sounds rigorous. It assigns blame to no one in the room. It defers action to a program with a distant horizon and elastic scope. And nobody was ever fired for insisting the data needed more work. It is the perfect institutional alibi: a way to appear prudent while declining to make a single hard design decision.
I want to be precise about the charge. Data quality concerns are real, and I have spent much of my career helping organizations address them. The alibi is not the concern itself. The alibi is using the concern as a gate, because a gate requires no one to specify which data, for which decisions, in which process, matters enough to fix.
Ask a readiness-gate advocate those three questions and the conversation changes instantly, because answering them requires exactly the decision-level process analysis the gate was postponing. Organizations that hide behind the gate are not protecting themselves from AI risk. They are protecting themselves from accountability, and paying a compounding competitive price for the comfort.
The Misconception: Garbage In, Garbage Out
The standard objection arrives with the confidence of an axiom: agents running on imperfect data will simply produce imperfect decisions faster. Garbage in, garbage out. The axiom is true, and it is being applied to the wrong architecture.
GIGO describes systems that act blindly on their inputs. A batch calculation, a report, a rules engine: feed them bad data and they propagate it, because nothing in the design evaluates the input before acting on it. A well-designed agent-enabled process is not blind by construction. It assesses the sufficiency of the data behind each case before acting, behaves differently when the data is weak, and escalates what it cannot support.
The failure mode GIGO warns about is real, but it is a property of undesigned automation, not of agents as such. It is worth noticing that enterprises have always run on imperfect data. Skilled people compensate through judgment: they recognize when a record looks wrong, cross-check another system, and ask. Decision-centric design gives agent-enabled processes that same compensating structure. The organizations that should fear GIGO are the ones bolting agents onto processes with no such structure, and their problem is the absence of design, not the presence of imperfect data.
What Unlocking Existing Data Looks Like in Practice
Consider a pattern we encounter in customer operations at energy utilities. The environment is textbook fragmentation: a decades-old customer information system, a separate meter data platform, a work and asset management system, and a document repository of service agreements, none of them integrated beyond nightly interfaces. The standing assumption inside the organization is that meaningful AI must wait for the CIS replacement program, a multi-year effort that has already been rescoped twice.
The design-first alternative starts with one high-friction process: billing exception resolution, where disputed and anomalous bills wait days while representatives manually assemble usage history, meter events, rate details, and account context from four systems. Redesigned, agents retrieve and reconcile that context at the moment each exception is created. Cases where the sources agree and the explanation is evident are resolved and communicated immediately. Cases where sources conflict, which is precisely where the data is imperfect, are packaged with the conflict made explicit and routed to a specialist. Resolution time collapses from days to hours, and the utility acquires something its data governance program had never produced: a ranked, evidence-based list of which specific data discrepancies actually drive customer-facing pain, drawn from live cases rather than profiling exercises.
The CIS replacement may still proceed on its own merits. But the value did not wait for it, and the replacement program itself is now better informed than any requirements workshop would have made it.
The Bottom Line: Design Is the Unlock, Infrastructure Is a Choice
The sequencing question at the heart of this paper has a clear answer once its components are priced honestly. Consolidation-first sequencing inherits the cost, delay, and delivery risk of a mega-program and gates all agent value behind it, on the strength of an assumption borrowed from the analytics era. Design-first sequencing puts a cheap, fast, reversible lever ahead of an expensive, slow, committed one, produces operational results while patience and funding are still abundant, and generates the evidence that makes any subsequent infrastructure investment targeted rather than total.
Executing design-first sequencing is a matter of method, not improvisation: selecting the right process, identifying its decisions, mapping what data each decision actually requires, and designing the handling for the cases where that data falls short. This is the core of our agentic AI consulting practice, and it is the difference between working with existing data deliberately and merely hoping the agents cope.
It is also a repeatable analytical skill. Business analysts and process owners who can trace decisions to their true data requirements become the people who unlock process after process without waiting on anyone's platform roadmap. Our business analysis and agentic AI training courses develop exactly that capability.
Perfect data is a destination worth moving toward. It was never the prerequisite. The prerequisite is a process designed around its decisions, and that is available to you in weeks, with the data you already have.
