An index of published deployment evidence answers one question: what has been said in public, by whom, and when. It cannot answer six questions readers will reasonably want answered, and the honest thing is to list them rather than let a grade imply more than it carries. These are the defence capability assessment limits of the Israeli Defence-AI Deployment Index, stated with the entries that show where each one bites.
It cannot tell you whether a system works
No grade in the index contains performance data. There are no detection ranges, no false-positive rates, no availability figures, no test results.
SMARTSHOOTER is graded L3 because Breaking Defense reported in June 2025 that the US Army was expanding its use of the fire-control system. Expansion of use is a procurement fact reported by a journalist. It is not a measurement, and the index does not convert it into one. The same holds for every other entry in the targeting and fire control category, which contains exactly one qualifying company.
A reader who wants to know whether a system performs needs test data, and test data for defence AI is very rarely published by anyone.
It cannot tell you who the customer is
Roughly half the index sits at L2, and L2 announcements routinely withhold the buyer. Odysight.ai's entry rests on a release announcing a purchase order from "an international defense contractor." Cyabra's rests on a release describing a multi-year contract with an unnamed APAC intelligence agency. Sentrycs' rests on a release announcing multimillion-dollar deals to protect military bases in Europe.
Each of those is a real, dated, checkable statement. None identifies a country, a service, a quantity or a configuration. A grade of L2 therefore establishes that a transaction was announced and leaves the buyer unidentified, which is a limit of what defence contracting publishes rather than a research failure.
It cannot tell you whether an L0 company holds a classified contract
Nine companies are graded L0. The grade records the absence of an independent public source corroborating a defence application. It cannot distinguish a company with no defence customer from a company whose defence customer does not announce purchases.
That second possibility is structural. The European Union's defence and security procurement framework provides that member states may exempt contracts from the ordinary procurement rules where strictly necessary to protect essential security interests, and applies special handling to contracts involving classified information. Comparable carve-outs exist elsewhere. Where those mechanisms are used, no announcement reaches the public record, and a supplier remains at L0 regardless of what it has sold.
It cannot tell you how big anything is
The index has no revenue field, no headcount field and no market-share field. It also excludes the largest Israeli defence organisations by design, because its inclusion test requires that a company's core product be AI or autonomy applied to defence. The reasoning, and the cost of it, is set out in why the Israeli primes are excluded.
The consequence is that the index cannot be read as a market map. Fifty companies is a working sample of a sector that Calcalist reported as having expanded to 312 firms on Startup Nation Central's landscape map. A reader wanting to know where Israeli defence AI money sits will not find it here.
It cannot tell you how autonomous a system is
The index records categories, not autonomy levels. It does not distinguish a system where a human authorises each action from one where a human supervises a process, and it does not attempt to, because the public sources do not use the terms consistently enough to support a coding scheme.
Entries in the autonomous air, autonomous ground and maritime, and C2 categories describe systems whose degree of human involvement varies and is often not specified in the source at all. The vocabulary problem is examined in in the loop, on the loop, out of the loop.
It cannot tell you what a system was used for
The index does not link any product to a specific incident, operation or casualty event, and will not, even where press reporting has done so. That rule and its reasoning, including the argument against it, are set out in why this index does not link products to named strikes.
The practical effect is that an L3 grade tells a reader a military used a system in operations, and stops. Readers looking for consequences are looking at the wrong document.
The precedent for stating limits this plainly
This is not an unusual position for a public-sources dataset. SIPRI, which maintains the longest-running open database of the world's largest arms-producing companies, states in its own sources and methods that publicly available financial and employment information on the arms industry is limited, that the scope and geographical coverage of the data are largely determined by the availability of information, and that the comparability of company revenue figures is limited. Where companies do not report arms revenue, SIPRI estimates it and says so.
A dataset built from disclosure ends where disclosure ends. The alternative is to fill the gap with inference, and inference is what makes an index unfalsifiable.
Limits of this reading
Even this list of limits is incomplete in a specific direction. The index is weighted toward companies with English-language coverage, which biases the sample against firms publishing only in Hebrew and therefore probably understates the contract and operational bands rather than overstating them. That skew is examined in the English-language bias in Israeli defence-tech research.
Category assignment, the choice of a single strongest source per entry, and the reading of ambiguous language such as "deployed" are all editorial judgements. A different researcher applying the same four grades to the same 50 companies would produce a similar but not identical result. The distribution the index does report — 13 companies at L3, 22 at L2, six at L1 and nine at L0 — is set out with its sourcing in the deployment evidence gap, and every entry is listed in the full index with the grading rules beside it.
Frequently asked questions
Does a high grade in the index mean a system is effective?
No. The index contains no performance data of any kind. An L3 grade means a named public source reported a military using the system in operations. It records that a report exists, not that the system performed to any standard.
Can the index identify who bought a system?
Often not. Many L2 entries rest on announcements that withhold the customer, describing an "international defense contractor" or "military bases in Europe." The index records the announcement and the fact that the buyer was not named.
Could an L0 company have a classified defence contract?
Yes. Procurement frameworks allow contracts to be exempted or withheld from publication on security grounds, and companies also sign agreements not to announce customers. An L0 grade records that no public source corroborates a defence application, not that none exists.
Sources
- idf.ai, The Israeli Defense-AI Deployment Index v1.0, 24 August 2026
- Breaking Defense, Israel's SMARTSHOOTER says US Army expanding use of high-tech fire control system, June 2025
- Nasdaq, Odysight.ai announces purchase order from an international defense contractor
- PR Newswire, Sentrycs secures multimillion dollar deals to protect military bases in Europe
- EUR-Lex, Public procurement in the fields of defence and security
- SIPRI, Arms Industry Database: sources and methods
- Calcalist, Israel's defense tech map expands to 312 firms amid wartime innovation boom
Independent publication of idf.ai. Not affiliated with, endorsed by, or connected to the Israel Defense Forces, the Israeli Ministry of Defense, or any government body. Compiled entirely from publicly published sources. No classified, restricted or non-public information. Listed companies may dispute any entry: send the published source that contradicts it and the entry will be amended or removed.