An organisation can hold extensive information and still lack the data required for a particular AI decision. Records may describe the right activity but arrive too late. Fields may be populated consistently but mean different things across systems. Outcomes may be available only for the cases selected by an earlier process. A well-maintained repository does not resolve those problems by itself.
Data readiness is therefore a property of the intended use. The organisation needs evidence that the information represents what the decision requires, is available when action remains possible, and can be used within the relevant permissions and controls. It also needs outcomes that support credible evaluation. Without those conditions, model development can produce confidence in a capability that the operation cannot reproduce.
The governance article established that permitted use determines the evidence and authority required. This article examines the information on which that use depends. The leadership task is to identify the data gaps that change the investment decision, resolve them proportionately, and avoid making either a promising model or a broad data programme a substitute for that assessment.
Start with the decision and its information requirements
A data-readiness assessment needs a defined decision, the action it supports, and the conditions under which it will operate. Asking whether the enterprise’s data is ready for AI is too broad to produce a useful answer. A capability that helps employees retrieve internal guidance and a capability that assesses affordable credit require different information and evidence.
For the employee assistant, readiness includes authoritative guidance, current versions, and retrieval that respects each user’s access rights. A repository containing stale and conflicting procedures can produce plausible but inappropriate answers. The team needs to identify which source governs, remove or mark superseded material, and test whether retrieval exposes information to staff who are not permitted to see it. The same readiness principle applies, but the evidence concerns reliable, authorised guidance rather than credit performance.
The lending case from the strategic-advantage article provides a concrete test. The bank wants to serve small businesses with seasonal cash flows that its existing process struggles to assess economically. The proposed decision concerns the amount and timing of affordable credit, within the bank’s risk appetite. Its information requirements follow from that purpose, rather than from the columns available in a convenient extract.
The team needs to distinguish recurring trading receipts from transfers, identify relevant obligations, and understand variation over time. It also needs reliable connections between the business, its accounts, and repayment outcomes. A model cannot compensate for treating money moved between a firm’s accounts as additional sales or for assigning one entity’s obligations to another.
The assessment must consider the bank’s visibility. Account activity describes only what passes through the accounts it can observe. A borrower using several institutions may have income or obligations elsewhere. Missing visibility is different from an absence of activity. The investment case needs to establish whether the available information supports the target population or whether additional evidence is required.
The bank begins with a bounded investigation of this segment’s information requirements. It does not approve enterprise-wide data remediation on the assumption that every improvement will support the lending proposal. Each identified gap needs a clear consequence for the decision and a proportionate response.
Test meaning and availability at the point of action
Data quality needs to be assessed against business meaning. Completeness and valid formats are useful, but a populated field can still misrepresent the concept required. A payment reference may be unreliable evidence of a receipt’s purpose. A transaction date may represent processing rather than the point at which information became available to the bank.
In the illustrative investigation, the team’s first extract includes daily balances and receipts that are finalised overnight. That extract is suitable for examining historical patterns. It does not establish that the same information is available during an application submitted earlier in the day. Training and evaluation must reconstruct what the decision would actually know at the time, rather than using later corrections or outcomes.
The bank tests the proposition using information available at the end of the previous business day, with clear treatment of subsequent activity. This aligns the assessment with a next-day decision service. An instant lending promise would require a different availability design and evaluation. The team narrows the intended service instead of assuming that historical access proves operational readiness.
The investigation also finds transfers between linked accounts counted among trading receipts. In this illustrative extract, the misclassification overstates trading receipts by more than 15% for roughly one in eight applicants. The team corrects the classification, validates the entity links, and repeats evaluation before interpreting model performance. The earlier result described the quality of a convenient dataset, rather than the intended decision.
Technical improvements can then target evidenced needs. A reliable classification or account-linking service may be reusable across other decisions, but its value begins with demonstrated requirements. Leadership can fund that capability with a clear explanation of what it enables and how its quality will be maintained.
Establish permission and traceability for the intended use
Accessible information is not automatically available for every purpose. The assessment needs to establish the basis for use, applicable restrictions, retention requirements, and who can access the data or derived outputs. Those conditions depend on the information, the proposed use, and the relevant jurisdiction and agreements.
The lending team works with the bank’s legal, privacy, and control functions to establish which account information can support the proposed assessment. It also examines any externally supplied information and the conditions attached to it. Permission for one service or analysis does not establish permission for a different decision. Unresolved restrictions affect scope and investment, rather than becoming a release detail left until model development is complete.
Traceability connects the source to the decision. The bank needs to know which records, transformations, and versions produced an assessment, including what information was missing. That supports investigation of errors, evaluation of changes, and the ability to explain the basis of a consequential action. A derived feature does not lose its significance simply because it no longer resembles the original record.
Responsibilities also need to remain visible as information moves. A source owner maintains the underlying records, a data-product owner maintains the transformations and delivery contract, and the lending-service owner determines whether the resulting information remains suitable for the approved use. These responsibilities can sit within existing teams, but the hand-offs require explicit ownership.
The linked service record introduced in the production article connects these conditions to permitted use. It records the input requirements, restrictions, provenance, and known limitations, with references to the authoritative data and control records. The objective is a dependable connection between evidence and use, rather than another collection of documents that drift apart.
Examine what the outcomes allow you to learn
Outcome data determines what the organisation can establish about performance. For lending, repayment outcomes are observed for funded borrowers. Applicants declined under the previous process generally do not provide the same repayment evidence. A model evaluated only on previously approved borrowers cannot establish how the new approach will perform across the applicants it proposes to add.
The seasonal-business proposal makes this limitation consequential. Its purpose is to reach a segment the bank previously struggled to serve. Strong results on familiar borrowers provide useful evidence, but leave the expansion claim unresolved. The team must describe that boundary plainly instead of treating the observed population as representative of every intended applicant.
Other outcome limitations also matter. Recent loans have not had enough time to reveal their full performance. Repayment difficulty can reflect servicing actions, changed terms, or wider economic conditions. An account closed after refinancing creates a different interpretation from a loan repaid from normal trading activity. Outcome definitions need enough precision to support the decision being evaluated.
A bounded evaluation within existing lending criteria would test a narrower claim. It could establish whether cash-flow assessment reduces the cost and time required to serve eligible seasonal businesses, with credit performance monitored against the current process. It would not establish performance among applicants previously declined or otherwise outside those criteria. Lower assessment costs could make an eligible segment more economical to serve without supporting a claim that the bank has expanded its credit-risk boundary.
The learning plan records which outcomes will be captured, when they become meaningful, and how intervention affects their interpretation. Synthetic data can test software and selected scenarios, but cannot establish the target population’s actual repayment behaviour. Additional information and careful evaluation are needed where missing outcomes limit the business claim.
Resolve the gaps that change the commitment
The investigation now supports a more specific decision. The bank has corrected material receipt classification, validated relevant account links, and established a service using previous-day information. It has confirmed the conditions for data use and identified the limits of its repayment evidence. These findings change the intended proposition from a broad lending ambition to a bounded next-day assessment for an explicitly defined population.
The committee approves one bounded evaluation among existing business account holders who meet current lending criteria, together with the data work necessary to support it. The evaluation tests whether eligible seasonal borrowers can be served more quickly and economically, with acceptable credit and customer outcomes. Using the existing relationship provides a defined acquisition route, and the evaluation must measure invitation costs, customer response, completed applications, and contribution after acquisition and servicing. Those costs sit alongside the proposed assessment savings in the revised economics, addressing the condition established in the investment-priorities article.
The committee withholds funding for an instant service and for broader eligibility until their availability and outcome-evidence requirements are met. Existing account relationships improve visibility, but do not establish that the bank observes every account or obligation. The conclusion is neither that the bank’s data is ready nor that it is unready. The information supports a particular commitment, with recorded limitations and a plan to test what remains uncertain.
Operating readiness also requires a maintained data contract. It should specify meaning, timing, expected coverage, quality thresholds, and behaviour when conditions fail. Monitoring must detect changes that affect the decision, including altered transaction classification, broken entity links, or late delivery. A source can remain technically available while becoming unsuitable for the service that depends on it.
Remediation priorities follow the consequences. A gap that prevents an affordable-credit assessment deserves attention before a cosmetic improvement to an unused field. Work that supports several evidenced opportunities can justify shared investment, with benefits and ownership made explicit. Speculative future uses need their own justification rather than being counted as guaranteed returns from a data programme.
Before committing to an AI capability, leadership needs answers to four questions. Does the information represent what the decision requires? Will it be available and permitted at the point of use? Which outcomes support evaluation, and whose outcomes remain unknown? Which unresolved gap changes the scope, economics, or safety of the next commitment?
Those answers give data investment a business purpose. In this bank, readiness supports a next-day decision for a defined population and leaves wider ambitions conditional on further evidence. The organisation can proceed with what its information genuinely supports, without asking the model to conceal what it does not know.
