The Machine Learning Readiness Audit: A Scoring Framework for Businesses

Why most machine learning projects fail before they start

Industry surveys repeatedly find that a large majority of machine learning projects — commonly cited figures put it above 85% — never reach production. The most frequent cause is not a bad algorithm; it is a business that started building before checking whether it had the data, infrastructure, team and organisational support the project actually required. A typical failure pattern looks like this: a company spends $50,000 building a model, discovers midway through that its underlying data is too sparse or too inconsistent to support it, and has to restart the data-collection phase from scratch — adding six to twelve months and a second round of cost on top of the first.

A readiness audit is a structured way to catch this before the money is spent. It will not guarantee a project succeeds, but it reliably filters out the projects that are certain to fail for reasons that have nothing to do with the modelling itself.

The five dimensions the audit scores

The framework scores a business out of 100 points across five weighted dimensions, reflecting how much each one typically determines project outcomes in practice:

Data carries the heaviest weight for a specific reason: it is the one dimension nothing else can compensate for. A brilliant team with a generous budget and full leadership backing still cannot train a model on data that does not exist or cannot be trusted.

Scoring the data dimension

The data section asks seven concrete questions, each scored on a scale (typically 0/1/3/5 points), covering: whether data is collected systematically and digitally rather than on paper or scattered across spreadsheets; how much history exists (a churn model realistically needs six to twelve-plus months, a demand forecast benefits from two to three years); measured data quality across completeness, accuracy, consistency and relevance; whether data is structured (database tables) versus unstructured (free text, documents, images); whether a clear target variable exists for the problem (did the customer churn, yes/no; how many units sold); whether the overall volume clears a rough threshold (roughly 1,000+ records is a common minimum for many standard business problems, with fraud and rare-event detection needing proportionally more); and whether enough relevant, informative features exist to support prediction — twenty or more relevant features is a strong position, under five is a serious constraint.

As a rule of thumb worth internalising: a month spent cleaning and structuring data before a project starts typically saves three months of downstream problems during it.

Scoring infrastructure, team, budget and culture

Infrastructure (20 points) checks for CRM/ERP-style business systems already in use, API integrations between systems (versus manual data exports), cloud computing access, available technical staff, and — increasingly relevant for any business exposed to power or connectivity disruption — redundancy in the IT setup itself, such as backup internet and power. A business relying on a single internet line and a single power source is at meaningfully higher risk of a stalled ML deployment than one with basic redundancy in place.

Team (20 points) checks whether leadership and key staff understand what ML realistically can and cannot do, whether the organisation has previously absorbed new technology successfully or shows strong resistance to change, whether any ML specialist or trainable analyst is available in-house or via a clearly identified outsourcing path, and — critically — whether a specific named person will act as product owner: the person who translates the business problem into requirements, makes trade-off decisions, and is the primary point of contact with the technical team. Projects without a named product owner fail at a disproportionately high rate regardless of how good the model turns out to be.

Financial readiness (15 points) checks budget realism against typical project costs (a chatbot commonly runs $2,000–10,000; a recommendation system $10,000–30,000; demand forecasting $5,000–20,000; computer vision for quality control $15,000–50,000), whether the business has actually calculated an expected ROI and payback period rather than just "hoping it helps," and whether it is prepared for ongoing costs — typically $500–2,000 a month for model support, $200–1,500 a month for cloud infrastructure, and $1,000–5,000 a quarter for periodic retraining.

Culture (10 points) checks for active leadership sponsorship (a project without visible executive backing has been observed to fail roughly 90% of the time) and whether the organisation already makes decisions from data rather than from intuition alone — a business that has never used analytics to inform a decision is unlikely to adopt ML outputs smoothly regardless of model quality.

Interpreting your score

Add the five dimension scores for a total out of 100:

• 80–100 — Ready. Start a real project with a $10,000–50,000 budget, expect a first result in two to four months, and move straight to defining the specific business problem (see the companion article on identifying ML-suitable problems).
• 60–79 — Nearly ready. Identify the single weakest dimension and invest in it directly — typically three to six months of targeted work — while running a low-cost pilot ($5,000–15,000) in parallel to build organisational confidence.
• 40–59 — Needs serious preparation. Custom ML development is premature. Focus on foundational systems (a CRM at minimum), consider off-the-shelf SaaS tools with built-in ML (chatbot platforms, email tools with predictive send-time optimisation) to build familiarity, and plan on six to twelve months before reassessing.
• Below 40 — Basic digitisation first. This is not a verdict on the business, just a statement of sequencing: get transactions and customer data into a digital system, build basic reporting habits, and plan on twelve to twenty-four months before an ML project makes sense.

The single most important habit the audit encourages is repetition: reassessing every three to six months turns a one-off diagnostic into a visible measure of progress, and most businesses that eventually run successful ML projects can point to a documented improvement in their readiness score over the preceding year.