The budget for an AI agent often covers the build, while the data work becomes clear later, usually after the business case is approved and the ROI clock has started.
A pilot works with sample data. Then the team tries to connect it to the records the business actually uses. Some are incomplete. Others are restricted or spread across applications. Getting them ready takes work that was never included in the plan.
The team has funded the experiment without accounting for everything needed to put it to work. In our customer conversations, executives are asking for measurable results tied to the KPIs they already report on: revenue growth, cost to serve, cycle time, working capital, and customer retention. To judge that investment, they need to understand what it takes to deliver the outcome and keep it working. The data problems were already there; AI brings them into the business case.
There is a second half to that story that executives should hear just as clearly. The same technology that exposes the data problem can shorten the time and cost of fixing it. AI and agents are now practical tools for the data work itself: profiling sources, matching records, drafting definitions, and monitoring quality. Used well, they pull forward the date when an AI investment starts paying back.
The work behind the first usable result
Consider an example I use to explain the data problem: the same customer appears in several business applications, but the records are not linked. Each system may look fine to the team maintaining it, but a business user trying to understand that customer has to move between systems and assemble the context. That friction shows up in the numbers: longer sales cycles, slower case resolution, and cross-sell revenue that never materializes.
For an agent to perform that work reliably, somebody has to make the relevant data usable together. That can mean resolving conflicting records, agreeing on which source is authoritative, establishing access, and supplying the business context the task requires. Connecting the applications alone does not settle those decisions.
This initial remediation work belongs in the estimate for the business use case, alongside the agent build and integration. It is often the largest variable in time to value, which makes it the line item most worth compressing.
The amount of work varies. Volume affects how much data must be processed and maintained. Variety affects the effort required to reconcile sources, formats, and definitions. Velocity determines how frequently the data must change to remain useful. Veracity concerns whether it can be trusted for the intended action.
Take financial statement preparation, an example I use to explain the difference. An agent can read a defined set of financial documents and perform specified calculations, with a person verifying the result. The underlying figures still have to be correct. The task has known inputs and a defined calculation, which makes its data requirements easier to scope than a workflow that needs changing customer information from several applications. For a CFO, the payoff is also easy to measure: days to close and hours spent on manual reconciliation.
Both require preparation, but the work involved is different, and the estimate should reflect that. I would start with the business problem, identify the KPI it should move and the data it requires, and scope the preparation against that requirement. That gives the business owner a more credible basis for deciding what to fund and a baseline to measure the return against.
Where AI and agents shorten the path
Much of the remediation described above has historically been manual: analysts sampling tables, data stewards reconciling records by hand, and subject-matter experts writing definitions in workshops. That labor is a large part of why data work runs long and over budget. AI changes the economics of several of those steps.
Discovery and profiling. Agents can scan source systems, catalog what exists, flag gaps and anomalies, and propose which fields a given use case depends on. Scoping becomes an informed estimate instead of a weeks-long round of interviews and sampling.
Record matching. Models can match records that refer to the same customer, supplier, or product across systems, score their confidence, and route only the ambiguous cases to people. Stewards spend their time on exceptions rather than the whole population.
Business context. Language models can draft data definitions, glossary entries, and lineage documentation from existing schemas, reports, and policies. The business owner reviews and approves a draft rather than writing one from scratch.
Quality monitoring. Agents can watch data as it changes, detect drift or broken feeds, and open a ticket or propose a fix before a bad record reaches a customer interaction or a financial report.
None of this removes the decisions that belong to people. Someone still has to decide which source is authoritative, who may access what, and how accurate the data must be for the action it drives. What AI removes is much of the repetitive labor between those decisions, which is where most of the time and cost sits. These tools need the same discipline as any other agent: human review for high-impact changes, an audit trail of what was altered, and evidence that the data actually improved.
For an executive, the effect shows up in three numbers: a shorter time to first production result, a lower remediation cost per use case, and a maintenance effort that grows more slowly than the number of agents it supports.
What remains after the cleanup
Once the initial work is complete, the business keeps changing. Records are added, applications change, and new workflows introduce new data dependencies, so the foundation needs people and processes to keep it usable. That is an ongoing operating cost. If the budget ends when the first dataset is ready, the organization has funded preparation without funding maintenance. Monitoring agents can keep that cost from rising in step with every new use case, but they do not make it zero, and it still belongs in the budget.
The consequences can extend beyond the source system. Where data has been copied into indexes or other retrieval assets, correcting the original record may also require updating what the agent retrieves. As more use cases depend on that data, the scope of a correction grows. Agents that trace lineage can identify every downstream asset a correction touches, which keeps the cost of a fix from multiplying.
Consumption-based pricing adds another cost. If poor data leads to an ineffective campaign or a bad customer interaction, the company still pays for the activity and then has to deal with the result. That spending needs to be judged against the business KPI the agent is meant to move, such as conversion rate, cost per resolved case, or revenue retained, rather than against activity volume.
There is also the investment in pilots that never reach that outcome. Experimentation has value when it answers a defined question. But an agent that works on sample data and then stops indefinitely at the production boundary leaves the organization with development effort, platform spend, and employee time without the intended operating benefit. In ROI terms, the cost keeps growing while the return stays at zero.
I have spoken with companies pursuing many pilots at once, each with its own demands on data and people. That spreads the effort before any one use case has demonstrated a business result. Narrowing the focus gives the team a chance to complete the data work and establish what a useful deployment actually requires. Before funding another pilot, I would want to understand whether the unresolved data work behind an existing one is the better investment. Often it is, and AI-assisted data tooling makes finishing that work cheaper and faster than it used to be.
Fund a path the business can sustain
My recommendation is to build the data foundation incrementally around the use cases the business wants to scale. There is no need to stop every AI initiative for a year-long enterprise data overhaul, but there does need to be a funded path from the selected pilot to a measurable result on a named KPI. AI-assisted data work makes that incremental approach more practical, because the foundation for each use case can be built faster and at lower cost than a traditional data program.
For that use case, the plan should account for the initial preparation, the ongoing work to maintain the data, the AI tooling that will accelerate both, and the resources required to put the agent into daily use. One person needs accountability for delivery across those dependencies. The business owner and the data team still have to agree on the outcome and the work underneath it.
People assigned to deliver it also need time. Asking someone in finance to oversee an invoice-reconciliation agent while carrying their full daily workload divides the attention the initiative needs. Agents that take on routine data checks can free some of that time, but they do not replace the accountable owner. Budget, staffing, and the measure of success have to be part of the same plan.
Expansion can then follow evidence: establish that the first workflow moves its target KPI against a baseline, understand what it takes to sustain it, and extend the foundation as additional agents require it. Each additional use case should cost less than the last, because the cleaned data, the definitions, and the agents that maintain them are reused. That compounding is where the portfolio-level return comes from.
I would ask every sponsor to track a short scorecard: time from approval to first production result, total cost per use case including the data work, the movement in the business KPI the agent was funded to improve, and the ongoing cost to keep it accurate. Those four numbers say far more about whether the investment is working than a count of pilots.
That is the commitment I would put in front of an executive sponsor: what it takes to prepare the data, keep it usable, and deliver the business result. Those costs belong in the decision before the prototype is built. So does the plan to use AI to bring them down. AI priced the data problem; it can also help pay it down.
Check all thoughts
