Your data is not ready for AI, and for one workflow that is fixable

In a Gartner survey of 1,203 data management leaders, run in July 2024, 63% of organisations said they either do not have, or are unsure whether they have, the right data management practices for AI.¹ The respondents are data management leaders reporting on their own organisations, so the figure is a self-assessment and no one’s actual data was tested.

Gartner’s own conclusion is sharper. It predicts that through 2026, organisations will abandon 60% of AI projects that are unsupported by AI-ready data.¹ That is a forecast and not a measured result. Other sources point the same way. Gartner’s July 2024 forecast for generative AI listed poor data quality first among the causes of projects being dropped after proof of concept,² and RAND’s interviews with 65 practitioners named a lack of data as one of five root causes of failure in AI projects.³

For a COO or CFO the practical question is narrower than “is our data ready for AI”. It is whether the data behind one workflow you want to automate is good enough to run it, and what it would cost to make it so. That is a question a team can answer in weeks, and it does not need a data-lake programme.

What “AI-ready” means in Gartner’s account

Gartner’s position is that readiness depends on the use case. The data has to be representative of the use case, including its patterns, errors and outliers, and it needs governance, metadata and monitoring around it. It also says AI-ready data is not a one-off task: it is an ongoing practice that covers metadata management, observability and governance.¹

If the data has issues, then the data is not ready for AI.

Gartner sets out five steps: align data to the use cases, define governance requirements, use active metadata, build the pipelines, and assure and enhance the data with practices such as DataOps and observability.¹ The list is written for a data management team. A manager who owns a workflow needs a version of it in business terms.

Four kinds of data one workflow needs

The following breakdown is our reading, built for workflow automation. It is not a Gartner framework. For a single process such as invoice handling, claims intake or purchase-order matching, the data that decides whether automation works falls into four groups:

• Process data. Records of what happened in the workflow: timestamps, statuses, who touched each case, how often it was reworked. Without it you cannot measure a baseline or see where cases get stuck.
• Documents. The invoices, contracts, emails and forms the process runs on. They need to be findable, in a readable format and current. A scanned image with no text layer, or three conflicting versions of the same contract, will produce wrong answers.
• Master data. The reference lists the workflow depends on: suppliers, customers, products, cost centres. Duplicates and out-of-date entries cause exceptions that a person then has to resolve.
• Access rights. Who may see and change which records. An automated workflow needs a defined identity, and it should have no more access than the person it replaces. Unclear rights are a risk and a reason for projects to stop.

Each of the four also has to be representative of the use case, which is Gartner’s test applied to a business process. If the training or test examples contain only clean invoices, the workflow will fail on the messy ones, and the messy ones are where the manual hours go.

A simple test shows how this plays out. Take twenty recent cases from the workflow, chosen at random, and ask a person to complete each one using only the data the automation would see. Every case where they have to open another system, ask a colleague or guess is a case where a machine would also fail. The share of cases that pass is a rough readiness score for that workflow, and the reasons for the failures become the repair list.

Why starting with the whole estate tends to stall

Gartner says traditional data management is too slow, too structured and too rigid for AI teams, and that the uses of data are poorly documented and sit in silos.¹ Its advice is to build on existing data management practices and add AI-specific capabilities, such as vector stores, chunking and retrieval-augmented generation.¹

Our reading is that this supports a narrow start. A programme to clean everything has a long list of owners, no single business case and no point at which anyone can say it has worked. Data for one workflow has one owner, a defined set of records and a result you can measure: cases handled, exceptions, hours. Older systems, which often hold the master data and the documents, can also be where integration costs surface; our technical debt benchmark looks at that side of the question.

What this does not prove

Gartner sells research on data management and AI, and its 60% figure is a prediction, not an outcome it has measured. The 63% comes from a survey of data management leaders about practices, not about results. Neither tells you how much a given clean-up costs, how long it takes, or how much of a pilot’s failure it would have prevented.¹

The narrow route has its own risks. Fixing data for one workflow can create local fixes that do not carry over, such as a duplicate-supplier rule that works for invoices and breaks for contracts. It can also leave the next workflow with its own silo. Some defects cannot be repaired at all: a document archive with no history, or master data that nobody has maintained for years, may need a decision to start the record again. And clean-up is often the largest line in the budget, so it should be priced before the model is.

Finally, no source we found shows that organisations starting with one workflow’s data succeed more often than those that run a wider programme. The case for starting narrow is that it limits the cost of finding out you are wrong.

What to do next

1. Choose one workflow and name the business owner who will accept the result, before anyone discusses models or vendors.
2. Map where its process records, documents, master data and access rights sit, and who is responsible for each. Gaps in that map are findings.
3. Sample recent real cases, including exceptions, and count the defects in each data group rather than describing the data as good or bad.
4. Price the fixes and the integration with older systems alongside the build, and compare the payback with and without them.
5. Set a re-check date. Gartner describes AI-ready data as an ongoing practice, so schedule a repeat of the sample after go-live and assign someone to own it.

If you want a first estimate of where AI could cut cost in your own workflows, the free pre-audit is a short questionnaire that returns one.