Five Questions to Ask Before Starting an AI Project

The research consensus is now overwhelming: most AI projects fail not because of the technology, but because of the decisions made before the technology is even selected. RAND Corporation identifies problem misunderstanding as the primary cause of failure. Prosci finds 63% of challenges are human. BCG reports that 60% of companies derive zero value from AI investments.[1][2][3]

Five questions, asked rigorously before any project begins, would prevent the majority of these failures. They are deceptively simple. Answering them rigorously is not.

The diagnostic is straightforward. Can you articulate the specific business problem in measurable terms? Do you know the current cost, error rate, or time requirement of the manual process? Is AI genuinely the right tool for this problem, or would something simpler work just as well? Better training, process redesign, conventional automation. Any of these could achieve the same result at lower cost and risk.

If your answer begins with "we want to use AI to..." rather than "we need to solve..." you are starting from the wrong end. The technology should follow the problem. When it leads, the failure rate approaches 95%.[5]

Data readiness isn't binary. It's a spectrum that should be assessed across five dimensions: completeness (do you have enough historical coverage?), accuracy (is the data clean and consistent?), accessibility (can you technically get to it, or is it trapped in silos?), governance (do you have classification, lineage tracking, and privacy controls?), and sustainability (can you continue collecting and maintaining it over time?).

The red flag is when an organisation says they'll figure out data quality once the project starts. Data readiness should be assessed in the first two weeks, before any building begins. Discovering that your data is insufficient in week eight is far more expensive than discovering it in week one.

McKinsey reinforces this: 48% of US employees would use generative AI more frequently if they received formal training.[9] And the trust gap between executives (who tend to be enthusiastic about AI) and frontline workers (who tend to be sceptical) means that even technically excellent deployments can fail at the adoption stage.

Prosci identifies five factors that predict AI adoption success: democratised AI expertise (building capabilities across the entire workforce, not just IT), individual choice (letting employees choose their own AI tools), internal skills development (which consistently outperforms reliance on external consultants), an experimentation culture (the single most significant factor for adoption), and comprehensive scale (larger, well-resourced initiatives paradoxically succeed more often than cautious, small-scale tests).[2]

Here's the uncomfortable truth: if your organisation has a history of failed technology implementations, poor training programmes, or resistance to process change, AI will not be different. The same organisational dynamics that killed your last digital transformation will kill your AI initiative, unless you address them explicitly before the project begins.

EY's February 2026 research shows that 90% of successful financial services firms have established C-suite or board-level accountability for AI outcomes. Eighty-four percent prioritise responsible, ethical AI deployment even at the cost of short-term speed.[11] Governance is not optional. The organisations that treat it as an enabler rather than an obstacle deploy AI faster and more profitably.

Three frameworks should be on every regulated organisation's radar before any AI project begins:

The red flag is any AI initiative that plans to "add governance later." In regulated industries, governance is a design constraint. Building a system and then trying to make it compliant is like building a house and then trying to add the foundation.

The traditional build-versus-buy binary is also increasingly obsolete. HatchWorks, Gartner, and Forrester now advocate an "own versus orchestrate" model: buy the compliance-heavy foundation (the system of record), and build the differentiating intelligence layer on top.[14]

When evaluating a potential partner, the critical differentiators are not the methodology steps. Every firm claims to start with the problem, focus on ROI, and handle compliance. The real differentiators are domain expertise (deep knowledge of your specific regulated industry), track record (production systems, not just pilots), speed (delivery in weeks, not months), and ownership (you own the IP and can operate independently when the engagement ends).

If your potential partner has never deployed a production AI system in your industry, they will learn on your budget. If they cannot point to systems that are running today, not pilots or proofs of concept, their claims about production deployment are theoretical.

Not every organisation needs custom AI. Not every organisation needs it now. The organisations in the 5% who succeed are not the ones who moved fastest. They're the ones who moved most deliberately. With clear problems, ready data, prepared people, embedded governance, and the right partners.