.svg){kind=icon}
Learn
Data readiness levels are a graded scale that rates how prepared a dataset is to be used in an AI or machine-learning system, from raw, ungoverned data at the lowest level up to cleaned, labeled, documented, and mission-usable data at the highest.
Last updated: October 18, 2026
Why it matters to your mission. A model that fails in the field usually did not fail because the algorithm was wrong; it failed because the data underneath it was never ready. Data readiness levels give a program a shared, graded way to say exactly how prepared a dataset is before anyone trains or fields a model on it. The outcome is that a commander or program lead can see where the data actually stands, fix the gap that matters, and stop paying for AI that was doomed at the data layer.
Data readiness levels are a graded scale that rates how prepared a dataset is to be used in an AI or machine-learning system, from raw, ungoverned data at the lowest level up to cleaned, labeled, documented, and mission-usable data at the highest. The idea borrows from technology readiness levels, which the Department of War has long used to grade how mature a technology is before fielding it, and applies the same discipline to data. A data readiness level is an honest label on a dataset: this is how far along it is, and this is what still has to happen before a model can safely rest on it.
The reason the scale exists is that "we have the data" and "the data is ready for AI" are two different claims, and programs routinely confuse them. A readiness level forces the second question. It is the difference between a hard drive full of reports and a corpus a model can learn from without inheriting its errors.
Key takeaways
Data readiness levels describe a dataset's maturity in stages, the way a thermometer reads a single number off a lot of underlying conditions. At the bottom, data exists but is raw: it is scattered across systems, undocumented, inconsistent, and nobody can vouch for it. In the middle, it has been gathered, cleaned, and structured, and someone understands what the fields mean. At the top, it is labeled where labels are needed, documented with its provenance and limitations, governed, and fit for the specific mission use it is meant to support.
The exact number of levels and their names vary by framework, because no single government-wide standard fixes them. What is consistent across frameworks is the direction of travel: from "data we happen to have" to "data a model can learn from and an evaluator can trust." A readiness level is always relative to a purpose. The same dataset can be high-readiness for one mission question and low-readiness for another, because readiness is about fitness for a use, not cleanliness in the abstract.
Three things move a dataset up the scale:
Because in defense the data, not the model, is usually the thing that is not ready. The U.S. Government Accountability Office examined the state of AI for weapon systems and found that alongside familiar acquisition problems, the Department faced AI-specific obstacles, chief among them "having usable data available to train the AI." The report's own example is a submarine-recognition model that depends on "gathering many images of various submarines and labeling them so the AI can learn to identify one on its own" (see the U.S. Government Accountability Office, Artificial Intelligence: Status of Developing and Acquiring Capabilities for Weapon Systems, GAO-22-104765, 2022). That is a data readiness problem described in plain language: the data has to be collected, cleaned, and correctly labeled before the model is worth building.
Without a shared readiness scale, that gap stays invisible until late. A program buys a model, integrates it, and only then discovers the training data was never representative, never labeled correctly, or never cleared for the classification it has to run at. Every dollar after that is spent fixing the data a readiness assessment would have flagged at the start. Naming the level up front turns a late, expensive surprise into an early, scoped task.
It also fixes accountability. "The AI does not work" is a dead end in a program review. "The data is at a low readiness level on labeling and provenance, and here is the plan to raise it" is something a program lead can fund, schedule, and track. The scale gives the whole team one language for a problem that otherwise gets blamed on the model.
A data readiness assessment scores a dataset against the mission it is meant to serve, not against an abstract ideal. In practice it walks the same dimensions that move data up the scale and asks, for each, how far along the data is and what the gap costs the mission:
The output is not a single grade but a profile: a dataset can be high on access and low on labeling, or clean but undocumented. That profile is what tells a program which gap to close first. A readiness assessment is most useful when it is repeated, because readiness changes as data ages, as sources are added, and as the mission question shifts.
Technology readiness levels grade how mature a technology or system is, from basic research up to a proven system in operation; data readiness levels apply the same graded thinking to the dataset the technology depends on. They are complementary, not interchangeable. A program can hold a high technology readiness level for its AI system and still fail, because the model is mature but the data feeding it is not.
The table separates the two and adds the related ideas programs often conflate with them.
| Term | What it grades | Relationship to data readiness levels |
|---|---|---|
| Technology readiness level (TRL) | The maturity of a technology or system, from research to fielded operation | The parent concept data readiness levels adapt; a mature system can still sit on immature data |
| Data readiness level (DRL) | How prepared a specific dataset is to train or run an AI system, for a specific purpose | The subject of this page; graded low to high across access, quality, labeling, documentation, and governance |
| Data quality | Whether data is accurate, complete, consistent, and timely | One dimension inside a readiness level, not the whole of it; quality can be high while governance or labeling is low |
| AI-ready data | Data that is prepared, governed, and fit to be used by an AI system | The end state a high data readiness level describes; readiness levels are the scale, AI-ready is the top of it |
The takeaway is that readiness is about the data's fitness for a use, assessed on its own terms. Grading the system without grading the data is how a program ends up with a mature model trained on data that was never ready.
Torch.AI builds government-owned reasoning infrastructure that raises data readiness by connecting to mission data where it already lives, rather than demanding a long, manual cleanup project before any value appears. The reasoning layer complements existing systems of record instead of replacing them: it reaches fragmented, multi-source data, preserves each source's meaning and provenance, and turns raw and unstructured records into a representation a mission can reason over. That is readiness work done in place, on the authoritative data, not a copy that loses its lineage.
Where the public products fit: NEXUS makes data meaningful, extracting entities and structure from unstructured documents and reporting and turning them into a semantic representation that a model and an analyst can both use, which is the move from low-readiness raw text to higher-readiness, usable data. HALO makes the relationships understandable, fusing resolved records into a graph so the data is not only clean but connected. Because the infrastructure can be deployed government-owned, the customer keeps control of the data, its provenance, and the readiness logic rather than renting them inside a vendor's proprietary environment; ownership varies by deployment and purchase, and the government-owned model is the one to require when control of the data matters to the mission. You can see how this is packaged as a capability on the Torch.AI software page.
This is the systems of record versus system of reason distinction at the center of Torch.AI's approach: existing systems keep recording the authoritative data, and the reasoning layer on top raises it to a readiness level a model can safely rest on, with every element traceable back to its source so the result can be governed and defended.
Scoping a data readiness assessment? Bring these seven requirements to a technical walkthrough and we will run them against a sample of your mission data and show where it stands: request a technical walkthrough.
Data readiness levels are a graded scale, low to high, that rates how prepared a dataset is to be used in an AI or machine-learning system. The lowest levels describe raw, scattered, ungoverned data; the highest describe cleaned, labeled, documented, governed data that is fit for a specific mission use.
Because in defense the data, not the model, is usually the thing that is not ready. GAO found that having usable data available to train the AI, including collecting and correctly labeling it, is a central obstacle to fielding AI in weapon systems (GAO-22-104765, 2022). A readiness level surfaces that gap early, before a program funds a model its data cannot support.
You score a dataset against the mission it is meant to serve across several dimensions: whether the data is accessible, consolidated without losing meaning, clean and consistent, correctly labeled where labels are needed, documented with provenance and classification, and representative of the conditions the model will face. The result is a profile that shows which gap to close first.
Technology readiness levels grade how mature a technology or system is; data readiness levels grade how prepared the dataset feeding it is, for a specific purpose. They are complementary: a program can hold a high technology readiness level and still fail because the data underneath the model was never ready.
There is no single government-wide standard that fixes the exact number of levels or their names; frameworks vary. What is consistent is the direction, from raw and ungoverned to cleaned, labeled, and mission-usable, and the underlying DoD expectation, set in the DoD Data Strategy's VAULTIS goals, that data be visible, accessible, understandable, linked, trustworthy, interoperable, and secure before it can be trusted for decisions.
Describe the problem.
Define the constraints. We will structure the session.
No demo.
No slide deck.
Real work.
Describe the problem.
Define the constraints. We will structure the session.
No demo.
No slide deck.
Real work.
