Home/AI Resource Kit/Foundations
Foundations

AI Readiness: What Has to Be True Before You Start

Roughly 95% of enterprise GenAI pilots produce no measurable P&L impact. The two prerequisites that separate the rest are free to diagnose and absent from every vendor deck.

DATAQUALITYPROCESSDEFGOVERNANCEINTEGRATIONADOPTIONMEASUREMENTREADINESS GATESFLOW STOPS AT THE FIRST FAILED GATEPASS · FAIL
What this actually means

AI readiness is not a technology question. It is whether the operational foundations exist for a model's output to be trusted and acted on. A model reasoning over inconsistent records produces confident wrong answers faster than a person would — speed becomes the risk rather than the benefit.

What the evidence says about AI readiness

FIG. 1 — THE GAP BETWEEN ADOPTION AND IMPACT
of organisations reported using AI in 2024, up from 55% the year before 78%
Source: Stanford HAI, AI Index Report, 2025
of enterprise GenAI pilots showed no measurable P&L impact 95%
Source: MIT NANDA, The GenAI Divide: State of AI in Business, 2025
of integrated pilots were extracting significant value 5%
Source: MIT NANDA, The GenAI Divide: State of AI in Business, 2025

Adoption is near-universal; impact is not. MIT's authors attribute the divide to organisational factors — data readiness, integration depth, and the capacity to learn from deployment — rather than to model capability.

What readiness actually requires

01

Data quality

Export the data the workflow depends on. Count duplicates, records dormant beyond twelve months, and free-text fields that should be structured. That percentage predicts your outcome better than any model benchmark.

This is the most common single reason a pilot stalls short of production. Remediation is measured in months, not weeks — which is why it belongs before the pilot, not after it.

02

Process definition

Have two people who own the process independently write its steps and decision rules. Where the two lists diverge is your actual specification.

A morning of calendar time. Skipping it is why automations get switched off — the model faithfully reproduces an ambiguity that was always there.

03

Governance and ownership

Name one person accountable for what the system is permitted to do, and one for consumption spend. If either answer is "the team", you do not have governance.

Gartner cites inadequate risk controls as one of three named causes behind its cancellation forecast.

04

Integration access

List every system the workflow must read from or write to, and confirm you can actually get programmatic access to each. Permission delays are a common and entirely predictable source of slippage.

05

Adoption capacity

Check whether the team already trusts the system the AI will sit on top of. If people keep shadow spreadsheets today, an AI layer inherits the gap between the official data and the real data.

06

Measurement baseline

Write down the number the system is meant to move, and its value today. Without a before number, "it helped" is unfalsifiable — and unfundable at the next budget review.

When not to do this

When you are not ready, and should wait

  • You cannot state the number the system is meant to move. Start there instead; it costs nothing and it decides everything else.
  • The process exists only as convention. Automating an undocumented process encodes one person's version of it.
  • The candidate workflow runs a few times a month. You need volume to learn from and to demonstrate impact — a quarterly process is a poor first target however important it is.
  • Nobody has authority to stop the project. An unstoppable project is an unmeasurable one, and it will consume budget long after the evidence says otherwise.

Green light, red light

AI readiness: when to start, and when to wait
Dimension Go ahead when Do not when
Data quality Records the workflow depends on are consistent, and you can say who owns each field More than roughly one record in ten has missing, stale or contradictory fields — an AI system will inherit every one and act on it faster than a person would
Process definition The workflow can be written down as steps a new hire could follow Nobody has written the process down; you cannot automate a process that exists only in people's heads
Measurement You have a current cost-in-hours and error rate for the workflow There is no baseline, so the pilot cannot be shown to have worked and will not get funded into production
Governance Someone owns the decision about what the system may and may not do Ownership is unassigned and expected to emerge later
Motivation A specific number needs to move The goal is to demonstrate AI capability — that is a demo, and demos do not survive production

How to assess readiness — without hiring anyone

  1. 01
    Run the free readiness assessment on our AI Transformation page. Six dimensions, four minutes, no email required — it names your weakest dimension.
  2. 02
    Export the data behind one candidate workflow and count duplicates, dormant records and unstructured fields.
  3. 03
    Ask two process owners to write the steps independently, then compare the two lists.
  4. 04
    Pick a high-volume, low-variance workflow — resist the interesting one, because interesting means variable and variance is where AI fails publicly.
  5. 05
    Baseline it numerically before building anything.

Questions buyers ask about AI readiness

It comes from MIT NANDA's 2025 study of 300 public deployments, 52 structured interviews and 153 survey responses across $30-40bn of tracked spend. It measures pilots with no measurable P&L impact — not pilots that failed technically, which is a different and smaller number.

The two checks that matter most — data export and independent process write-up — take an afternoon and a morning respectively. Neither needs a vendor, a budget or a procurement cycle.

Then data is the project, and knowing that before spending on AI tooling is worth the exercise on its own. Data cleanup returns value independently of AI: better reporting, better forecasting, less manual reconciliation.

Usually not. Most early value comes from applying existing models to well-defined problems, which is an integration and process discipline more than a research one.

Want an honest read on your readiness?

Six evidence tests, four minutes, no email required. It will tell you plainly if the foundations are not there yet.

Run the assessment

Read the rest of the kit.

Five capability guides, every figure cited to a named source and year.

Open the AI Resource Kit
Book a Strategy Session