The ROI of Getting Off to a Good Start: How to Build a Business Case Based on an AI Assessment
By Sebastián Grattarola — CTO of BigCheese, a Premier AWS Partner
The conversation about artificial intelligence in organizations has taken a new turn. Two years ago, the question was, “Does this work?” Today, almost no one doubts that it does; the question has changed, and it’s a tougher one: “Where do I start, and how do I know it’s worth it?” That’s the question leaders around the world are asking themselves.
It’s a business question, not a technical one. And yet most AI initiatives start on the technical side—a model, a pilot, a test that someone saw in a demo—without first answering how much the problem they want to solve costs or how much it would be worth to solve it. The predictable result: a lot of activity, little conviction, and a budget that no one dares to approve because no one put the numbers on the table.
In the organizations we work with in Latin America—fintech, retail, and payment providers—the message is always the same: there’s no shortage of enthusiasm for AI, but there’s a lack of a business case that can turn that enthusiasm into an investment decision. That is exactly the purpose of an AI Assessment: measure before you build.
What Does an AI Assessment Address?
An AI Assessment is a short, focused project that takes a vague “we want to use AI” and turns it into a prioritized portfolio of use cases, each with its own business case. The deliverable isn’t a technical report—it’s the clarity needed to decide where to invest first, how much it will cost, and what you can expect to get in return.
The focus, then, is not on the method (the phases, workshops, interviews) but on the output: an informed investment decision. This follows the same logic as AWS’s Path-to-Value framework, which maps out the journey of generative AI from idea to value in production. And at the heart of that decision is the ROI model. Below, I’ll share the framework we used to build it, so you can apply the same logic to your own organization.
Where Does the Return Come From: Five Dimensions of Value
The most common mistake when estimating the value of an AI agent is to focus solely on cost savings—that is, how many person-hours are freed up. It’s the easiest dimension to calculate, but it’s rarely the most significant. A solid business case projects value across five dimensions (in line with the AWS Cloud Value Framework, which measures the value of the cloud beyond mere cost savings):
- Cost savings and efficiency — person-hours freed up by tasks that agents can complete independently or more quickly, plus reduced costs associated with legacy tools, errors, rework, and SLA violations.
- Revenue generation and growth — faster responses that drive sales recovery, improved conversion rates through personalized recommendations, new capabilities (24/7, new channels), reduced churn, and the ability to scale volume without incurring a linear increase in costs.
- Productivity and throughput — how many more transactions are processed in the same amount of time, how much resolution time is reduced, and how many hours are freed up for the team to focus on higher-value work.
- Customer and Employee Experience — improved CSAT/NPS, reduced customer effort, and, on the internal side, teams freed from tedious tasks, with the resulting impact on turnover.
- Risk mitigation and compliance — consistent enforcement of policies, real-time detection of fraud and errors, traceability and auditability, and operational continuity that does not depend on specific individuals.
The key is to assess each dimension against a baseline of the current state. Without a credible baseline (how much the process costs today, how often it’s run, how long it takes, how much is lost due to errors), ROI is a guessing game. That’s why the model always starts by quantifying the present before projecting the future.
The Model, Step by Step
Let’s look at how this framework applies to an illustrative case. The numbers that follow are examples, chosen to demonstrate the method (they do not correspond to an actual client), but the structure is the one we use in every engagement.
The scenario: An organization manually processes a high volume of incoming documents—for example, verifying and extracting data from documents that arrive through multiple channels. Currently, a team reads, classifies, and uploads these documents by hand.
Step 1 — Baseline of the current state. Let’s assume 4,000 documents per month, an average of 15 minutes per document, with an hourly labor cost of US$18. That amounts to 1,000 person-hours per month, about US$18,000 per month, or US$216,000 per year in processing costs alone. Added to that is an 8% rework rate due to data entry errors.
Step 2 — Value Projection (the five dimensions). A GenAI agent (built on Amazon Bedrock for extraction and classification, with guardrails and traceability built into the design) can autonomously handle a conservative percentage of the documents (say, 60%) and accelerate the rest. We project the value as follows:
- Efficiency: ~US$130,000/year in freed-up hours (redeployed to higher-value tasks; this does not necessarily mean a reduction in headcount).
- Error reduction: Reducing rework from 8% to 3% yields another round of savings.
- Throughput: The same capacity can process 2–3 times the volume at no additional cost, enabling growth without hiring.
- Experience and Risk: Faster responses and a complete, auditable trail of every agent decision.
Step 3 — Investment. It is estimated to consist of two parts: the one-time construction cost (agent design, integration development, data preparation, governance and guardrails configuration, and team enablement) and the recurring annual operating cost (AWS services: model inference on Bedrock, orchestration, storage, support and maintenance, and internal time spent on monitoring and continuous improvement; an agent is not a “set-and-forget” solution).
Step 4 — ROI, payback period, and time horizon. With the annual value and the investment on the table, the calculation is straightforward:
ROI % (a N años) = Valor neto a N años / Inversión total a N años × 100
Payback (meses) = Costo de construcción / Valor neto mensual
We always model 12-, 24-, and 36-month time horizons, with one key caveat: Year 1 typically captures only 50% to 75% of the full-scale value, due to the adoption curve and phased rollout. Presenting Year 1 as if it were the full-scale program is the quickest way to lose credibility with a CFO.
Step 5 — Sensitivity Analysis. No figure stands up to the scrutiny of an investment committee if it is presented as a certainty. That is why the model always runs three scenarios (conservative, base, and optimistic) by varying the critical assumptions: what percentage of tasks the agent handles, how long adoption takes, and how much construction costs might rise. A business case that shows its own lower bound is much more convincing than one that shows only the upper bound.
Why Measuring First Changes the Conversation
There’s a common objection: “All this talk of modeling ROI before building sounds like bureaucracy that slows down innovation.” Experience shows the opposite. The organizations that put AI to work the fastest are the ones that first took the time to understand where the return on investment lay. Not because the model is a mere formality, but because it’s what unlocks the investment—it’s the same principle of governance as an enabler that we apply to AI governance: it gives decision-makers the economic conviction to say yes, and prevents wasting effort on use cases that don’t move the needle.
Starting with measurement also changes what gets built. When you prioritize a portfolio based on value and impact (rather than which demo excited whom), the first use case to go into production is usually the one that pays for the ones that follow. That’s the compounding effect that turns an isolated AI initiative into an organizational capability.
The Point
AI has evolved from a technological gamble into a business decision, and business decisions are made based on numbers. An AI Assessment isn’t just a preliminary step before “the important part”—it is the important part, because that’s where you determine whether what you’re going to build is worth the cost and where it makes sense to start.
If your organization is exploring AI and the question on everyone’s mind is, “Where do I start, and how do I know it’s worth it?”, that’s exactly the question a good business case answers—before writing a single line of code.
The best way to make a big impact with AI is to think big and start with something achievable: a first use case with a clear return on investment that will pay for the next ones.