AI Readiness Assessment

An AI readiness assessment should end with a decision that an accountable person can act on: proceed within a defined boundary, test a specific uncertainty, repair a prerequisite, or stop the proposal. A score alone cannot tell a business whether an agent should draft an answer, recommend a payment or change a customer record. Those activities demand different evidence.

For an operations director, the useful question is whether a named workflow is ready for its next investment decision. For a CIO, it is whether the organisation can support that workflow after the initial demonstration. Assess both, and keep the findings separate. An enterprise AI maturity model describes repeatable capability across the business; readiness is conditional on the use case, users, data and proposed authority.

Define what “ready” would permit

Write the decision boundary before sending a questionnaire. Identify the work, the people affected, the systems involved and the action being considered. “Ready for AI” is too broad. “Ready to test an assistant that drafts equipment-return summaries for a supervisor, without assigning charges” can be assessed. A later proposal to assign charges automatically requires a different review even if it uses the same model.

Record the current service standard and the failure that matters most. A return summary might be useful when it is slower but more complete; a charge recommendation might be unacceptable when the supporting photograph is ambiguous. The assessment should expose those trade-offs before anyone sets an overall accuracy target. Where the main question is whether the proposed approach can work at all, use a focused AI feasibility study alongside readiness.

Build an evidence ledger, not an opinion survey

Interviews reveal where to look. Evidence determines what is known. Ask each respondent to show a recent example, the actual source system and the person who can explain an exception. Record whether the evidence is observed, reported but unverified, missing or conflicting. Missing evidence is an uncertainty to resolve; it should not silently receive a middle score.

Readiness evidence for a bounded agentic workflow

Area	Evidence to inspect	Decision it supports
Work and demand	Recent completed cases, backlog, rework and reasons for escalation	Whether the proposed intervention addresses the real constraint
Knowledge and data	Authoritative documents, access rules, update ownership and representative examples	Whether the system can reach reliable information for the intended users
Action authority	Current approval limits, exceptions and records of who approved changes	Which actions may be assisted, recommended or executed
Evaluation	Examples of acceptable and unacceptable outcomes, with named reviewers	Whether performance can be judged before expansion
People and capacity	Reviewer availability, training needs and ownership after the trial	Whether the proposed oversight can actually happen
Technology and support	System access, integration constraints, support arrangements and recovery procedures	Whether a bounded trial is practical and supportable
Business value	Measured workload and a credible route from improvement to business benefit	Whether further assessment is worth funding

Use the ledger to connect every conclusion to its source. “Data is accessible” needs more than an administrator confirming that a database exists. Check whether the intended role can access the required records, whether records can be linked correctly, and whether the permitted use includes the proposed activity. Any privacy, security or sector-specific applicability questions need the relevant specialist; an assessment score does not settle them.

A rental-return example: one workflow, different readiness answers

Consider an illustrative equipment-rental company. Staff receive return photographs, inspection notes and signed hire conditions. Management wants an agent to reduce disputes about damage. A workshop initially describes the problem as slow inspection reporting. Following actual returns reveals that the larger delay occurs when an old contract version is used to interpret a damage charge.

The assessment finds that photographs are available, but some cannot be reliably matched to the equipment identifier. Inspectors disagree on how to describe ordinary wear. Supervisors can review draft summaries, but they cannot review every proposed charge during peak returns. These are three separate findings: record linkage, decision consistency and review capacity. Buying a stronger model would not, by itself, resolve them.

A defensible next step is a limited experiment using approved, correctly matched historical cases. The system may organise evidence and draft summaries. It may not decide liability, send a charge or change the rental account. The operations owner must agree examples of ordinary wear; the records owner must establish identifier checks; finance retains charge approval. The business can learn about summarisation value while withholding a capability for which the evidence is inadequate.

The example is a proposed assessment method, not a customer result. It also shows why the assessor should revisit use-case discovery when the stated problem differs from the actual bottleneck. An assessment that recommends a different intervention can be more valuable than one that approves the original AI idea.

Separate blockers from improvements

A blocker prevents the next proposed action from being evaluated or controlled. An improvement makes the operating environment better without necessarily blocking a narrow trial. The distinction depends on scope. Missing access to any authorised evidence is a blocker for an evidence-based assistant. An untidy archive outside the selected trial scope may be an improvement item, provided the trial does not claim to cover it.

For each gap, record the consequence, owner, required evidence of closure and the decision that remains suspended. Avoid tasks such as “improve governance” with no observable endpoint. A useful task is “operations lead approves the escalation categories and demonstrates who receives each category during the trial”. Put consequential uncertainties into the AI risk register, rather than hiding them in a presentation appendix.

• Proceed: the stated next step has adequate evidence and an accountable owner.
• Proceed with conditions: a narrower scope is supportable, with explicit exclusions and monitoring.
• Resolve first: a prerequisite prevents a meaningful or controlled trial.
• Do not pursue: the problem has insufficient value, a better alternative or an unacceptable residual exposure.

Run interviews around conflicting evidence

Invite the process owner, frontline users, data or system owner, technology lead and relevant control specialists. Bring finance when the value depends on labour redeployment, avoided expenditure or revenue assumptions. Ask users and managers about the same recent cases separately before reconciling their explanations. A signed process map may describe intended practice while the case history shows how work is really completed.

Test the handoffs too. If a human must review uncertain outputs, ask where the review appears in the working day and what happens when the reviewer is absent. If a source document changes, ask who updates it and how affected decisions are identified. Readiness often depends on these ordinary operating details more than on the sophistication of the proposed technology.

Place Saudi guidance in the assessment correctly

SDAIA has published an AI Adoption Framework, described in the Saudi Press Agency announcement as guidance for adoption. Use applicable official guidance as a reference; do not present this article’s checklist as an SDAIA certification or official scoring instrument.

For a Saudi organisation, local relevance comes from its actual operating context: Arabic and English records, named decision owners, contractual commitments, the data involved and its sector. Record unresolved applicability questions and refer them for review. Do not infer permission to deploy from a completed readiness workshop.

Require a handover that can change the investment decision

The final pack should contain the assessed boundary, evidence ledger, findings by capability, conditional decision, remediation backlog and a brief for the next test. State what was not inspected. A leadership summary should explain what becomes possible after each prerequisite is resolved, rather than simply reporting that a percentage has increased.

Transfer the funded actions into the transformation roadmap. Keep the evidence available for the later pilot-to-production decision, when broader users, live data and more consequential actions may require another assessment. Readiness has an expiry condition whenever the approved boundary changes.

Open for AI’s agentic AI consulting can help clarify workflow, authority, data readiness and evaluation needs. To discuss a readiness assessment for one workflow, bring a redacted process example, the proposed action and the decision you need to make. Agree the assessment boundary and deliverables before the work starts.