Raft showed how to prioritize AI initiatives and build a realistic roadmap

Raft has proposed a practical way to transform a fragmented list of AI ideas into a manageable transformation program: first calculate the potential value of each initiative, then match it against achievability, select an implementation approach, and only then build the roadmap. Such an approach is needed by companies that already see dozens of possible AI scenarios but don't want to drown in pilots without measurable results. The material is based on the following thesis: you can implement almost anything, but you cannot do everything simultaneously.

This is why the author proposes abandoning a chaotic set of experiments and moving toward portfolio management of AI initiatives. The logic starts with the gap-analysis phase: the organization gets a list of tasks where it becomes clear which use cases it is ready for now, and which ones lack data, infrastructure, competencies, or security frameworks. But assessing feasibility alone is insufficient.

To decide what truly deserves budget and team resources, you need to separately evaluate expected business value. For this, Raft offers a simple formula for potential value: problem scale, frequency of occurrence, and complexity of manual solution. Each parameter is scored on a scale from 1 to 10, and the final score falls in a range from 1 to 1000.

Scale shows how many people or processes are affected by the problem; frequency shows how often it occurs; complexity of manual solution shows how much time, money, and effort is required without AI automation. The authors separately note a practical threshold: if an initiative scores fewer than 200 points, in-house development may take too long to pay off. In such cases, it is more sensible to look at ready-made off-the-shelf solutions and consciously accept their limitations rather than launch an expensive custom development for weak results.

The next step is the "value—achievability" matrix. It helps separate initiatives that deliver maximum returns with acceptable complexity from those that are premature or economically questionable. The proposed approach is to select projects from the upper left quadrant of a 2×2 matrix: quick wins, major bets, incremental innovations, and other scenarios where there is noticeable impact and a clear path to implementation.

After this, the company assembles what is called an AI Tech Gartner's Sandwich: for each initiative, it chooses an implementation model—buy, build, or partner—and then adds a layer of risks and security. The idea is to view AI not as a standalone service, but as a multi-layered system of applied solutions, platform components, and protective mechanisms, including a TRiSM approach to trust, risk, and security management. The final result of this selection is not just a list of priorities, but a phased transformation roadmap.

In the first phase, which the authors call the foundation, it is recommended to take one or two high-value, low-complexity initiatives, launch pilots with quick results, and in parallel close technical gaps: prepare data, build integrations, strengthen infrastructure, hire missing roles, and train employees. For this phase, preference is often given to the cloud to reduce the cost of hypothesis verification and shorten time-to-market. The recommended duration of this phase is six to twelve months.

Next comes scaling: successful pilots are expanded to new teams and processes, and medium-complexity initiatives are moved into active execution if critical gaps have already been addressed. In parallel, the layer of threat monitoring and protection of AI systems is strengthened. The third phase is optimization, when AI tools already become part of operational activities, and focus shifts to SLA, quality control, combating model degradation, supporting the MLOps pipeline, and reducing total cost of ownership, including a possible transition from cloud to on-premise infrastructure.

The main conclusion of the material is that AI transformation requires not inspiration but discipline. If a company first calculates value, then honestly checks achievability, and then distributes initiatives by phases, it reduces the risk of getting stuck in endless prototypes and begins to manage AI as a mature investment portfolio. For the market, this is an important signal: the winners will not be those who launched demos fastest, but those who connected AI projects to business metrics, security, and a realistic implementation plan.