Five Things to Know Before You Build an AI System

95% of generative AI pilots never reach production scale.^[1] In most cases, the technology works. The problem is that the problem was not understood well enough before the build started. The wrong process or the wrong version of the process gets automated. The architecture does not fit the workflow. The interface requires people to change how they operate. Six months later, adoption is low and the investment has not paid off. In 2025 alone, 42% of companies abandoned most of their AI initiatives, with the average organization scrapping nearly half of its proofs-of-concept before they reached production.^[2]

Before writing a line of code or selecting a model, there are five dimensions that determine whether an AI system will succeed in a business environment. These are business decisions that shape every technical choice downstream.

1. Problem Selection

The first step is choosing the right process, because not every business problem benefits from AI. The problems worth pursuing share three characteristics: they are repetitive, they are labor or cost intensive, and they are information or content heavy. A process that meets all three is a strong candidate. A process that meets one is probably better served by traditional automation or workflow improvements.

This is where most organizations go wrong first. Gartner's analysis of cancelled agentic AI projects found that the majority were pursued based on technological fascination rather than concrete business value.^[3] McKinsey's data supports the same conclusion: only 6% of companies achieved greater than 5% EBIT impact from AI, and those were concentrated entirely among firms that redesigned workflows before selecting their modeling approach.^[4] The companies that started with the problem outperformed the companies that started with the technology.

Survey the business for candidate problems. Score each one against these three criteria. Then rank them by value with the people who own the processes and the executives who own the budget. The goal is the problem where AI will produce measurable impact, whatever it sounds like in a board deck.

2. Process Understanding

The anatomy of the process determines the shape of the solution. A system designed without this understanding will be built to solve the wrong version of the problem.

Map the process end-to-end with the people who run it, the process owner and the subject matter experts, and capture the workflow as it operates: the workarounds, the exceptions, and the manual steps no one has documented. Document the inputs and outputs. What format are they in today? What format do they need to be in? Understand where this process sits in the larger ecosystem: what feeds into it and what depends on its output. Quantify the current cost in time, labor, and error rate. This baseline is what the AI system will be measured against.

High performers are nearly 3x more likely to have fundamentally redesigned workflows as part of their AI efforts.^[5] Mapping is also where integration constraints, security reviews, compliance requirements, and organizational friction surface, while they are still cheap to address.

3. Architecture Decision

With the process mapped, the next question is what AI should do in this specific workflow. Determine the type of system: is this a chatbot assisting an individual, a reasoning agent embedded at a specific step, or a deterministic pipeline where code handles the gathering and the model handles the judgment? Identify which steps require intelligence and which steps are deterministic and can be handled by code. Understand the file types and systems the solution will touch, how intensive the audit trail needs to be, what maintaining the system looks like, and the complexity of the task the AI is being asked to perform.

The architecture must be shaped by the process. If the process is well-defined and the inputs are structured, a deterministic approach will be cheaper, faster, and more consistent than giving an agent open-ended autonomy. By late 2025, Salesforce reached the same conclusion, shifting its Agentforce platform away from probabilistic execution and back toward guardrails, deterministic automation, and data-first design.^[6] Safe and reliable AI depends on deterministic foundations.^[7]

4. ROI and Timeline

Not all value shows up on the same schedule. Some AI systems deliver immediate impact: the process runs faster or cheaper from day one because the system slots directly into an existing workflow. Others require behavior change, new interfaces, or new ways of working, and the value takes months to materialize as people adopt the tool.

Know which category your project falls into. Adoption-dependent value is the hardest to capture, but projects with dedicated change management resources achieve 2.9x the success rate, and user-centered design approaches produce 64% higher adoption.^[8] If the system requires people to change how they operate, if it requires training, if it introduces friction, then the ROI timeline extends significantly and the risk of failure increases. The highest-value AI systems are the ones that are invisible to the end user: same interface, same workflow, faster and cheaper results.

5. Interface and Change Management

How people interact with the system determines whether they use it. Trust in fully autonomous AI agents dropped from 43% to 27% in a single year.^[9] That decline happened because organizations gave agents broad autonomy and the results did not meet expectations. AI is a tool that operates within a process. When organizations treat it as a replacement for the process rather than a component within it, the result is friction, inconsistency, and declining confidence across the business.

The ideal system runs behind the scenes: the user selects a client account or uploads a document, the system handles the rest, and the AI never becomes something the user has to learn or think about. Every new interface, every new workflow, every prompt window is friction that erodes adoption. If the process is well understood and the inputs are known, the system should plug directly into existing infrastructure so that adoption is a non-event.

These five dimensions are the first steps of any project plan. They are the questions that need clear answers before any building begins. If they cannot be answered with confidence, the project is not ready. If they can, the architecture, interface, and implementation decisions follow naturally.

The checklist below is designed to be used at the start of any AI initiative. Fill it out with the process owner, the subject matter experts, and the technical team before committing to an architecture or a timeline. The clarity it produces is worth more than any prototype.

AI Readiness Checklist

1. Problem Selection
Process being evaluated
Is it repetitive?
Is it labor/cost intensive?
Is it information/content heavy?
Business value if improved
Stakeholder alignment (C-suite, process owner, AI lead)
2. Process Understanding
Process mapped end-to-end?
Inputs (current format and shape)
Outputs (required format and shape)
Upstream dependencies (what feeds this process)
Downstream dependencies (what this process feeds)
Current cost (time, labor, error rate)
3. Architecture Decision
System type (chat, single agent, deterministic)
Where does AI reason vs. where does code execute?
File types and systems involved
Audit trail requirements
Maintenance and update complexity
Task complexity for the AI component
4. ROI and Timeline
Expected time to impact
Immediate value or adoption-dependent?
Behavior change required? (Y/N)
Training requirements
Baseline metrics for comparison
5. Interface and Change Management
Can the AI be backend/invisible?
Existing interfaces it can plug into
New interfaces required
Steps that need automation
Steps that stay human-owned
Adoption risk (low / medium / high)

If any section cannot be completed with confidence, the project needs more discovery before committing to a build.

Sources

1. MIT / RAND. 95% of generative AI pilots fail to scale (MIT). 80.3% of AI projects deliver no measurable business value (RAND). ↩

2. S&P Global, 2025 Enterprise AI Survey. 42% of companies abandoned most AI initiatives in 2025. Average organization scrapped 46% of POCs before production. ↩

3. Gartner, June 2025. Predicts more than 40% of agentic AI projects will be cancelled by end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. ↩

4. McKinsey, State of AI 2025. Organizations with significant financial returns are 2x more likely to have redesigned workflows before selecting techniques. ↩

5. McKinsey, State of AI 2025. High performers 3x more likely to have fundamentally redesigned workflows. ↩

6. Sweep, "Why Enterprise AI Stalled in 2025." Salesforce shifted Agentforce narrative away from probabilistic execution toward guardrails, deterministic automation, and data-first design. ↩

7. Sweep / NVIDIA. "Deterministic logic and AI are not opposites; safe AI depends on deterministic foundations." ↩

8. Pertama Partners, AI Project Failure Statistics 2026. Change management resources achieve 2.9x success rate. User-centered design produces 64% higher adoption. ↩

9. Shakudo, "Why 80% of Enterprise AI Agents Fail in Production." Trust in fully autonomous AI agents dropped from 43% to 27% year-over-year. ↩