Gartner expects organisations that build unified semantics for their data to raise AI agent accuracy by as much as 80% and cut agentic AI running costs by as much as 60% by 2027. Those are unusually large numbers for an initiative that produces no user-facing feature, and they point at something teams deploying agents keep rediscovering the hard way.
The agent is rarely the hard part now. Good models are available on tap, the tooling to connect them to systems has matured, and a competent team can have an agent calling internal APIs inside a fortnight. What determines whether it works is underneath: whether the data it reaches is findable, interpretable and trustworthy.
The headroom here is genuinely large. A Cloudera and Harvard Business Review Analytic Services survey of 1,574 enterprise IT leaders found that 7% describe their data as completely ready for AI adoption. IDC found 94% of IT leaders naming data quality as the top factor in AI project success. Gartner puts 63% of organisations in the category of lacking, or being unsure whether they have, the data practices AI requires.
Read together, those figures say the constraint is widely shared and widely unaddressed. Which means it is available as an advantage. A Singapore mid-market company that does this work properly is not catching up to its competitors on data foundations. It is getting ahead of them.
What “AI-ready” means in practice
The phrase gets used loosely, so it helps to be concrete. An agent needs four things from your data, and a warehouse migration supplies roughly one of them.
It needs to find the relevant record, which is a retrieval and indexing question. It needs to interpret what it found without guessing, which is a semantics question. It needs to know whether the information is current and where it came from, which is a provenance question. And it needs to be prevented from reading what the requesting user should not see, which is an access question that most data platforms were designed before anyone needed.
The second of those is where the Gartner number comes from, and it is the one most often skipped because it looks like paperwork rather than engineering. Take a question as ordinary as how many active customers a business has. Sales counts anyone with an open opportunity. Finance counts anyone who has been invoiced in the period. Support counts anyone with a live entitlement. All three are correct inside their own function, and all three are stored in different systems with the same column name.
Ask an agent that question and it will answer confidently from whichever table it happened to query. Ask it twice through different paths and you get two different numbers, both defensible, neither wrong. That is the failure that erodes trust in an agentic system faster than any hallucination, because it is not obviously an error. It is just an answer that does not match the one someone else got.
The four pieces of work, in order
Agree the definitions for your core entities. Pick the ten or fifteen nouns your business actually runs on. Customer, order, active, churned, qualified, closed. Write down what each one means, which system is authoritative for it, and who owns that definition. This is a series of meetings rather than a build, it frequently surfaces disagreements that predate the AI project by years, and it is the single highest-return piece of work on this list.
Classify the unstructured material. More than 80% of a typical enterprise footprint is documents, email, contracts, call recordings and images, and most of it has never been catalogued. This is where the context lives that makes an agent genuinely useful rather than merely fluent, because the reason a deal stalled is in an email thread and not in the CRM stage field. Classification no longer requires a manual taxonomy project, since models handle the extraction and labelling themselves, but it does require deciding what categories matter and where the output is stored.
Attach provenance and freshness. Every record an agent can reach should carry where it came from and when it was last true. This lets an agent qualify its own answers, which is the behaviour that makes it trustworthy, and it lets you debug a wrong answer by tracing the input rather than interrogating the model. Teams that skip this find themselves unable to distinguish a reasoning failure from a stale-data failure, and those have completely different fixes.
Make access controls queryable at runtime. An agent acting for a user must resolve that user’s permissions at the moment of the query, not inherit a service account’s broader reach. This is the piece most likely to require real engineering, because many internal data stores assume access is enforced by the application sitting in front of them rather than by the store itself.
The order matters. Definitions before classification, because classification without agreed categories produces labels nobody trusts. Provenance before access, because you cannot write a sensible rule about data whose origin you cannot establish.
What you can skip
Two things commonly get bundled into data readiness programmes and neither is a prerequisite for useful agents.
A full warehouse consolidation is the first. Moving every source into one platform is a multi-year commitment that solves findability and leaves semantics, provenance and access exactly where they were. Agents read through APIs and retrieval layers, not through a single schema, so the consolidation can proceed on its own merits and timeline rather than blocking the AI work.
The second is an enterprise-wide data catalogue populated before anything uses it. Catalogues built speculatively go stale, because nobody maintains metadata for a consumer that does not exist yet. Catalogue the sources your first two agents actually touch, keep that current, and let demand pull the rest in. A small catalogue that is true beats a large one that was true in March.
There is a related trap in governance sequencing. Committees sometimes want the full data governance framework signed off before a single agent reads a single record, which inverts the useful order. Governance designed against a real workload is specific and enforceable. Governance designed in the abstract produces a policy document and no change in behaviour.
A sensible first 90 days
Nobody should fund a two-year data programme to find out whether agents help. The sequence that works runs the other way around.
Choose one workflow where a single system already holds the truth. Invoice matching, support ticket triage, lead enrichment. One source, one clear outcome, a number you can measure before and after. Build the agent against that, properly, with logging and a defined owner. It will work, because the data question is contained.
Then use what you learn to scope the rest. The second workflow you attempt is the one that will hit the semantics problem, and it will hit it with a specific, fundable shape: these two systems disagree about this entity, and here is what that costs us weekly. That is a far easier case to make than an abstract argument about data maturity, and it gets funded.
Meanwhile start the definitions work in parallel, because it needs calendar time rather than engineering time. Fifteen entities, one owner each, written down. A quarter is enough if someone chases it.
Webpuppies builds data foundations and agentic systems together for clients across Singapore and the region, because in our experience the two cannot be sequenced cleanly. If you have agents working well in one place and stalling everywhere else, the cause is usually three rows down in this article. Get in touch and we will find out which one.
Sources
- Only 7% of Enterprises Have AI-Ready Data
- The 2026 Enterprise Guide to AI-Ready Data: Definition, Requirements, and How to Get There
- AI Agent Readiness: A 2026 Data Checklist, EnFuse Solutions
- Data Quality for AI: How Enterprises Improve Accuracy, Reduce Bias and Scale AI in 2026, Techment
Frequently Asked Questions
What does AI-ready data mean?
Data an agent can find, interpret without guessing, and trust. In practice that means agreed definitions for core business entities, classified unstructured content, provenance and freshness metadata, and access controls an agent can respect at query time.
How many organisations have AI-ready data?
A Cloudera and Harvard Business Review Analytic Services survey of 1,574 enterprise IT leaders found 7% say their data is completely ready for AI adoption. Gartner reports 63% of organisations either lack or are unsure whether they have the data practices AI requires.
Why does unstructured data matter for AI agents?
Unstructured content such as documents, email, call recordings and images makes up more than 80% of a typical enterprise data footprint, and most of it has never been classified. It holds the context agents need, but requires extraction or classification before an agent can use it reliably.
What is a semantic layer and why do agents need one?
An agreed set of definitions for core business entities, such as what counts as an active customer or a closed deal. Without one, an agent picks a plausible interpretation from whichever table it queried, and two correct-looking answers disagree.
Can we start using agents before the data work is done?
Yes, on narrow tasks against a single well-understood source. Pick a workflow where one system holds the truth, prove the pattern, and let that prove which data work is worth funding next.
