AI can increase the pace at which an organisation produces, analyses and acts. If the underlying work is clear, that can improve capability. If decisions, information and accountability are confused, the same acceleration can multiply inconsistency and risk.
AI readiness due diligence should therefore test the organisational system around the technology, not count licences, pilots or vendor claims.
Executive answer
A target is ready to scale AI where it can define useful outcomes, supply reliable information, understand the workflow, assign decision ownership, preserve human escalation, evaluate performance and learn from adoption. Investors should:
- connect AI use cases to the value-creation plan;
- test data and process conditions at the point of work;
- clarify human decision and escalation boundaries;
- review evaluation, monitoring and incident response; and
- estimate adoption effort as well as technology cost.
Readiness is the organisation’s ability to use AI within a governed work system and improve that system through evidence. The NIST AI Risk Management Framework organises this work around Govern, Map, Measure and Manage. OECD principles add human oversight, transparency, robustness and accountability.
TL;DR
AI readiness is not the number of tools deployed. Test seven conditions: a clear use case, reliable information, an understood process, a named decision owner, human escalation, evaluation and monitoring, and the capacity to learn and adopt. Evidence from field research shows that generative AI can improve productivity in a defined setting, with different effects across workers. That is a reason to test context, not to assume universal returns. The Diagnostic can surface organisational hypotheses, but it does not certify AI readiness.
Seven due-diligence checks
1. Is the outcome clear?
Define the user, task, decision and measurable result. “Use AI to improve efficiency” is not an investable operating assumption.
2. Is the information reliable and permitted?
Examine source quality, access, ownership, privacy, security and freshness. A sophisticated model cannot repair unclear provenance or unauthorised use.
3. Is the workflow understood?
Trace the current work, including exceptions and rework. Automation applied to an unstable process can hide problems and scale variation.
4. Who owns the decision?
Separate model output from accountable judgement. Name who can accept, reject or override an output and who carries the consequence.
5. Where does human escalation remain?
Define confidence thresholds, prohibited uses, material exceptions and the route for review. Human oversight must be designed into the work, not added as a slogan.
6. How is performance evaluated and monitored?
Specify a baseline, test set, quality criteria, harmful failure modes, drift signals and incident response. Vendor benchmarks rarely settle local performance.
7. Can the organisation learn and adopt?
Assess training, workflow redesign, feedback, management attention and whether people can challenge the system. Licences without changed work are not adoption.
Examine the portfolio claim, not only the pilot
A pilot may succeed because it has clean data, expert supervision and protected attention. Scaling changes volume, user diversity, exceptions and integration. Ask which controls and support must grow with usage.
For each use case, record intended value, affected stakeholders, decision owner, information sources, validation evidence, escalation, monitoring and residual uncertainty. Place cost against the full operating change, including process redesign, training, assurance and management time.
What the evidence does and does not say
Brynjolfsson, Li and Raymond found a substantial average productivity improvement from generative AI assistance in a customer-support setting, with larger benefits for less experienced workers. The result is important and bounded. It does not establish the same effect in every function, firm or risk environment.
The automation-augmentation literature also highlights a continuing tension: technology can substitute for tasks while increasing the importance of complementary human judgement. Readiness requires explicit choices about both.
The contrary case
An organisation does not need perfect process maturity before experimenting. Bounded, reversible use cases can help expose ambiguity and generate learning. The danger is treating experimentation as proof of scalable capability. Keep stakes, permissions and monitoring proportionate.
A practical test: one use case, one operating contract
Select the AI use case most material to the thesis. On one page, state the outcome, workflow, information, owner, human boundary, evaluation, monitoring and stop condition. Ask management and frontline users to review it separately. Investigate material differences before underwriting scale.
What investors ask next
Does an AI policy prove readiness?
No. It is evidence of governance intent. Readiness also requires working controls, usable data, ownership, evaluation and adoption.
Should every model be independently validated?
Assurance should be proportionate to consequence, opacity, reversibility and regulatory obligation. High-impact uses need stronger evidence.
How should AI value enter the model?
Separate proven local results, plausible extensions and speculative upside. Include operating-change and assurance costs.
Can the Diagnostic assess AI?
It can surface relevant organisational conditions such as decision clarity and learning. It is not an AI audit, technical evaluation or certification.
Continue the investor diligence cluster
Use the Organisational Genius Diagnostic to open a structured conversation about the operating conditions around AI. Then test the technology, workflow and governance evidence directly. Explore the investor path for deeper review.
© Course Correction Consulting LTD. Evidence-informed practitioner guidance, not an audit, valuation or investment recommendation.
.svg)
.svg)
