An AI pilot can look promising with a few carefully chosen examples and still struggle with the information a business uses every day. Missing fields, contradictory documents and unclear access rights become part of the system's behaviour. Before selecting a model or automation platform, establish whether the data can support the specific job you want it to do.
Data readiness is a practical assessment of fitness for purpose. A customer enquiry classifier needs representative messages and agreed categories. A knowledge assistant needs current, authoritative documents with suitable permissions. A forecasting system needs reliable historical observations. There is no single file-cleaning exercise that makes a business ready for every AI application.
Define the task and the decision it supports
Write down what the system should produce, who will use it and what happens afterwards. For example, a support assistant might suggest a relevant procedure for an employee to review. That is a more bounded task than letting it answer any question about the business and perform the resulting action.
Identify the information required for a good answer. A delivery question may depend on the current order, the carrier status and the applicable customer policy. If one of those sources is unavailable, the system needs a defined way to handle the gap.
Separate the desired outcome from the technology. Ask whether improved search, a structured form or an explicit rule would meet the need. Review data foundations alongside business strategy, governance and operational capability: a prepared dataset alone does not establish a workable service. Our AI readiness and assurance guide provides the related checks.
Inventory the sources and identify an owner
List the systems, files and documents that might support the task. Include their location, format, responsible person and update process. A shared folder may contain several generations of the same policy, while an apparently complete CRM export may omit important conversation history.
Identify the authoritative source for each type of information. If a spreadsheet and a customer system disagree about an account status, the AI should not be expected to invent the resolution. The business must decide which record governs the task and how conflicts are investigated.
Give each source an owner who can answer questions about meaning and quality. Technical access to a database does not provide the operational knowledge needed to interpret it. A developer can explain how a field is stored; a business owner may need to explain what the field actually represents.
Inspect a representative sample, including difficult cases
Review ordinary records, exceptions and material from different periods or channels. A sample consisting only of clean recent entries can conceal the problems the live system will encounter. Include incomplete requests, unusual formats, old terminology and cases where staff currently disagree.
For structured data, check completeness, consistency and permitted values. For documents, review legibility, headings, version information and the relationship between text and tables. A scanned page that a person can read may still be difficult for an extraction process to interpret accurately.
Measure problems in terms of the task. A missing telephone number may be irrelevant to categorising an enquiry but essential to a callback workflow. Avoid spending time repairing every field equally when only some defects affect the proposed use.
Record findings with examples and an estimated impact. This turns “the data is messy” into specific work, such as resolving duplicate policies, standardising service categories or making current order information accessible through a controlled connection.
Check coverage rather than quantity alone
A large dataset can still omit the cases that matter most. For a routing task, examine whether each category has realistic examples. For a knowledge assistant, ask whether the sources answer the questions employees actually raise. Thousands of unrelated files will not repair a missing operational procedure.
Compare coverage across customer groups, service types and enquiry channels where those distinctions matter. If the examples come mostly from one team, the pilot may perform poorly for another. Treat that as a sampling question rather than assuming the model will generalise correctly.
Pay attention to rare but consequential cases. An urgent request may appear infrequently, yet missing it could be more costly than mislabelling an ordinary message. Agree how such cases are identified, tested and escalated. Overall accuracy alone can hide a serious gap.
Resolve freshness and conflicting information
Mark current documents and establish how superseded material will be handled. A policy with no version date or owner creates uncertainty even before AI is introduced. If several files describe different arrangements, ask the responsible person to resolve them.
Specify how changes reach the proposed system. A daily export may be adequate for a stable reference list but unsuitable for live availability. Define the acceptable delay in business terms and identify what the system should do when information is older than that limit.
Include removal as well as addition. Deleting an obsolete document from its original folder does not automatically prove that every copy, index or cache has been updated. Ask the implementation team how changes will be propagated and verified.
Establish permitted use and access boundaries
Check who is authorised to use the information for the proposed purpose. Consider confidentiality, personal information, supplier terms and any restrictions attached to documents. Owning an account that can download a file does not settle every question about using that file in another system.
Involve the people responsible for information governance and security before transferring material to a new service. Record what the provider receives, where information is processed and the relevant retention and access arrangements. Assess the actual configuration, rather than treating every AI service as having identical terms.
Design access around the user and task. A system serving several teams should not expose a restricted document merely because it shares keywords with a permitted one. Access checks need to operate before restricted information becomes available to the answer-generation process.
Use a limited, appropriately handled dataset for the pilot. Reducing unnecessary information makes the assessment easier and reduces avoidable exposure. It also helps the team identify which source contributes to an answer.
Create an evaluation set before improving the system
Write representative questions or inputs and the expected behaviour. Include cases where the correct response is to request more information, decline to answer or send the task to a person. A system that always produces something may look productive while concealing unreliable decisions.
For a knowledge task, record the source that supports the answer and what would make an answer misleading. For a classification task, agree categories and how ambiguous examples should be judged. Resolve disagreements in the reference answers before using them to assess a model.
Keep some examples separate from development. Repeatedly tuning against every test item can produce a system that handles the test set without demonstrating broader usefulness. Review new real-world cases during the pilot and add them through an agreed evaluation process.
Turn the checklist into a readiness decision
Classify findings as ready for a bounded pilot, requiring preparation or unsuitable for the proposed use at present. Assign an owner and action to each unresolved issue. A useful assessment ends with decisions rather than a vague recommendation to improve data quality.
Distinguish essential preparation from optional refinement. Missing permissions, unresolved authoritative sources and absent evaluation examples may prevent a meaningful pilot. Cosmetic formatting differences may be manageable through the selected extraction process. Make that judgement with evidence from the actual task.
Plan how readiness will be maintained after launch. Changes to services, documents, categories and supplier connections can alter the system's reliability. Allocate responsibility for data updates, exception review and periodic evaluation.
Giraffe Digital's digital strategy service connects business tools, customer journeys and measurement. That planning is a useful starting point for an AI proposal: define the job, identify the information it needs, check the boundaries and prove a bounded result before expanding the scope. Data readiness gives the pilot a question it can answer and gives the business evidence it can trust.
