Industrial Research Unit — research note, Straits Institute for Applied AI
Applied AI: the systems are running. Almost none have been measured.
The Institute has verified 5,322 applied AI systems at work in 249 countries. Roughly one in nine of them carries a measured result. The rest is an account of AI applied and switched on, with nothing published about what happened next.
By [ researcher name ], Head of the Industrial Research Unit, Straits Institute for Applied AI. Drawn from the applied AI statistics the Institute maintains and sweeps weekly.
Last revised 19 September 2026
5,322 verified systems
Press pack, data and charts
Applied AI means artificial intelligence put to work inside a job somebody already holds — in an organisation that can be named, on a date that can be fixed, with a published source. It is the opposite end of the field from AI research: research asks what a model can do, applied AI asks what changes when that model is handed work a customs officer, a radiologist or a procurement officer is currently paid to do.
Applied AI systems are defined by their placement, not their architecture. A language model drafting a council’s minutes, a vision model reading cargo scans and a scoring model ranking tax returns are all applied AI. The same models sitting in a laboratory are not.
Findings
- 1
Of 5,322 verified applied AI systems worldwide, 594 — 11% — carry a measured outcome. 66% are a system going live with no published result of any kind.
Figure 1Of 5,322 verified applied AI systems across 249 countries, 594 (11%) carry a measured outcome; 66% are a system going live with no published result. Source: Straits Institute for Applied AI, State of Applied AI, https://www.straitsai.institute/state-of-applied-ai - 2
Applied generative AI is a minority of what actually runs. Generative and language models account for 46% of verified applied AI systems; the remaining 54% is computer vision, prediction and scoring.
Figure 2Applied generative AI accounts for 46% of verified applied AI systems worldwide; the remaining 54% is computer vision, prediction and scoring. Source: Straits Institute for Applied AI, State of Applied AI, https://www.straitsai.institute/state-of-applied-ai - 3
The region buying the most AI writes the least of its own. In-house building runs at 41% of entries in Asia-Pacific and 19% in North America.
Figure 3In-house building of public-sector AI runs at 41% of entries in Asia-Pacific and 19% in North America: the region buying the most AI writes the least of its own. Source: Straits Institute for Applied AI, State of Applied AI, https://www.straitsai.institute/state-of-applied-ai - 4
21% of all verified applied AI systems point at a person who did not choose to interact with them — facial recognition, risk scoring, behaviour flagging, biometric checks on benefit recipients.
Figure 421% of verified applied AI systems worldwide are pointed at a person who did not choose to interact with them – facial recognition, risk scoring, behaviour flagging and biometric verification. Source: Straits Institute for Applied AI, State of Applied AI, https://www.straitsai.institute/state-of-applied-ai - 5
The rule arrives after the system. Where deployment came first, it preceded the first governing rule by a median of 22 months.
Figure 5Where public-sector AI deployment preceded the first governing rule, it did so by a median of 22 months. Source: Straits Institute for Applied AI, State of Applied AI, https://www.straitsai.institute/state-of-applied-ai
What applied AI means — and what we refused to count
Applied AI has no agreed definition in common use, which is why almost nothing written about it can be checked. A survey reporting that four firms in five “use AI” counts answers to a question. A vendor case study counts a sale. Neither can be traced to a single system in a single building on a single date.
The Institute therefore works from a definition written as an admission test. An entry enters the record only if four conditions hold at once, and this is what applied AI means in every figure below.
- 1
A named organisationNot a sector, not “a European bank”. If the body cannot be named, nobody can check the claim or ask them how it went.
- 2
A real jobSomeone currently holds the role the system touches. AI applied to work nobody was doing is a product, not an application.
- 3
A date“Will deploy”, “is piloting” and “has deployed” are three different claims. A report without a date collapses all three into one.
- 4
A published sourceSomething a third party can open: a ministry release, a regulator’s filing, a procurement notice, a local news report.
What we refused to count
Naming the exclusions is the more useful half of the definition, because most reporting on this subject consists of them.
- Vendor case studies. The organisation is named by the party that sold the system. The claim and the interest are held by the same person.
- Adoption surveys. “78% of firms use AI” counts answers, not deployments. It cannot be traced to a single system anywhere.
- Strategy and roadmap announcements. An intention to deploy is evidence about a budget cycle, not about a workplace.
- Model capability results. A benchmark score describes a laboratory. Nothing follows about what happens when the model meets an institution.
- Funding rounds and partnerships. Capital and memoranda precede applied AI by years, and frequently precede nothing at all.
- Unattributed pilots. “A major hospital group is trialling” fails the first condition, whatever the trial is worth.
Applying that test to a week of AI coverage removes most of it. What survives is smaller, duller and considerably more informative — a municipality of 2,500 people in Liguria running 43 checking systems over its council paperwork; a finance-ministry engineer in Nepal assembling a flood-response portal overnight; an armed forces IT unit in Montenegro writing its own document assistant rather than send files abroad.
Counts by industry, region, role and country are published and updated separately in the applied AI statistics. This note deals only with what the record shows.
Only 11% of the world’s AI deployments have published a result
Every entry is filed as one of four kinds, and the four form a ladder. A deployment is a system in service. Evidence of adoption attaches a usage figure — seat counts, query volumes, share of cases handled. Evidence of impact attaches a measured change in time, cost, error rate or throughput. Regulation is a rule now governing the practice.
What the record contains, by kind of evidence
Figure 1
Each entry is filed as exactly one kind. Where a source reports both a deployment and an impact figure, the entry is filed as evidence of impact and the deployment is logged separately.
Roughly one in nine verified applied AI systems carries a measured outcome. Anyone quoting a return figure for AI in general is extrapolating from that tenth — and that tenth is itself only what has been disclosed.
Straits Institute for Applied AI
The scarcity is not evenly spread. Impact evidence clusters in three places — medical imaging, tax administration and contact centres — and the reason is the same in all three: the prior job was already measured. There was a baseline read rate, a collection figure, a call-handling time. Where the old work was never measured, the AI applied to it is not measured either, and there is no prospect of it being measured later.
Almost every figure that does exist is self-reported by the deploying body. Nebraska’s statewide assistant is credited with answering 88,000 questions and cutting call volume by a fifth; a New Orleans emergency-dispatch system with triaging 3,500 events and cutting duplicate calls by 30%; Indian Railways with lifting waitlist prediction accuracy from 53% to 94%. The record carries all three, marked as unverified, because marking them is more useful than either dropping them or repeating them as fact.
Credits 30 projects across 13 departments with cutting some waiting times by up to 90%. No independent evaluation exists.
Deployment · 17 Aug 2026
Selects 100% of verification cases by automated risk assessment; 1.9m taxpayers auto-assessed, R8bn refunded within 72 hours.
Evidence of impact · 1 Jul 2026
A Croydon facial-recognition trial scanned 470,000 faces over six months, yielding 173 arrests and one false match.
Evidence of impact · 23 Jun 2026
The AI actually running is mostly not generative
Public discussion of applied AI has narrowed to one model family. The record does not support that narrowing. Sorting every verified system by the technique doing the work produces a field in which applied generative AI is the largest single category and nowhere near the whole of it.
Applied AI systems by the technique doing the work
Figure 2
Each entry is filed by the technique performing the task described in the source. Systems combining techniques are filed by the one the source names as primary — a document pipeline that reads scans and then summarises them is filed as computer vision if the extraction is the point, generative if the summary is.
Applied generative AI — generative and language models put into a real job — accounts for 46% of the record, concentrated where the work is text: drafting, summarising, translating, answering. Austria rolling GovGPT out to 180,000 federal employees. Japan’s Digital Agency giving 180,000 officials a generative tool for translation and minutes. Brazil’s courts adopting a judge-built system that has summarised 1.55 million documents across 106,000 cases.
What gets missed is the other 54%. Computer vision is the second-largest category and in several countries the largest: fire-spotting cameras across 143 Croatian sites, orbital flood detection over the Pamir Mountains, tattoo matching that produced a first forensic identification in Guatemala, AI X-ray reading on Dominican cargo. Prediction and scoring is third: risk-ranked tax audits in Greece, Sri Lanka and South Africa, pedestrian-risk scores on 17,000 Calgary intersections, waitlist forecasting on Indian Railways.
Applied gen AI is the loudest category, not the largest field. Just over half of the applied AI systems now running use no generative model at all — and that half holds most of the measured results, because vision and scoring systems arrive with a baseline to be measured against.
Straits Institute for Applied AI
The distinction matters for anyone buying or teaching this. Applied generative AI fails in ways that need review, disclosure and an audit trail. Applied vision and scoring systems fail in ways that need thresholds, error-rate monitoring and appeal routes. A single course called “AI for professionals” that does not separate the two is teaching neither.
The countries buying the most AI write the least of their own
The most consequential thing about a deployment is not what it does but who made it, and it is the dimension almost nobody tracks. Three postures recur.
- Built in-house. Montenegro’s armed forces built Digital Adjutant internally to keep sensitive files off foreign systems. Korea’s Public Procurement Service shipped seven tools from a 26-person team, specified conversationally by staff who do not write code. A Nepali finance-ministry engineer built a flood portal in a night.
- Local vendor or national model. Poland’s central offices testing an assistant on PLLuM; Serbia’s welfare chatbot on the Kragujevac supercomputer; Thailand’s five-agency pilot on Pathumma, a Thai-built model.
- Global platform. Copilot, Gemini, Claude, ChatGPT, Agentforce, deployed largely as sold. Iceland across its state institutions; the Philippines to 50,000 public servants, rising to 200,000; the US Department of Defense to a workforce of three million.
Who built the deployed system, by region
Figure 3
Share of each region’s entries by build posture. Where an organisation self-hosts or fine-tunes a global model, the entry counts as in-house only if the application layer was written by the organisation itself.
A country’s real AI capability is better read from who writes its public-sector software than from how much compute it has bought. In-house building concentrates in mid-sized states with strong civil services and unresolved sovereignty questions — not in the largest AI economies, which import platforms and configure them.
Straits Institute for Applied AI
The pattern is not a ranking, and reading it as one gets it backwards. Buying a platform is often the correct decision, and building in-house is frequently a response to a constraint rather than a display of strength — no budget, no vendor willing to localise, or a rule against sending files offshore. Guatemala’s planning secretariat runs its assistant entirely offline, which removes token costs and lets officials handle sensitive files directly. Japan’s Soka City went on-premise for the same reason.
What the split does predict is what a workforce needs to learn. A country that configures purchased systems needs evaluation, oversight design and change management. A country that builds needs those and engineering as well.
One in five systems is pointed at someone who didn’t opt in
Each entry is also read for direction. Citizen-facing systems are ones the public chooses to use: municipal assistants, benefits chatbots, appointment bookers. Internal administration is staff-only: minute-taking, drafting, records, rostering. Enforcement and monitoring systems have a subject rather than a user — the person they act on did not opt in.
Direction of use, by region
Figure 4
Share of each region’s entries by direction, assigned from the occupational role and the function described in the source. Mixed-use entries are filed by the function named first.
The enforcement column is where the record has grown fastest. Live facial recognition on the London Underground and in Bradford city centre; 130,000 faces scanned in a Western Australian trial; facial verification made compulsory for 877,000 pensioners in Tajikistan; 487 AI cameras switched on across Phnom Penh; a Guatemala City installation bundled with crowd-counting tools the vendor pitches for identifying protests, in a country with no data-protection statute.
Direction predicts public reaction better than industry or technology does. Two systems running the same model — one drafting a council’s minutes, one scoring a citizen’s risk — belong to different arguments. Filing both as “AI in government” is how most coverage of this subject loses the thread.
Straits Institute for Applied AI
The rule arrives a median of 22 months after the system
Because the record dates both deployments and regulation, it can settle a question usually argued from intuition. For each country the Institute measures the interval between its first logged public-sector deployment and the first logged rule governing that use.
Months between first deployment and first governing rule
Figure 5
+38
+31
+27
+24
+19
+17
+14
+11
-3
-5
-8
-11
Bars to the right: the deployment came first, by that many months. Bars to the left: a rule was already in force. Public-sector entries only, where both dates are reliably documented.
The countries where rules came first are not the countries deploying least. They are the ones whose existing administrative-law tradition already covered automated decisions before anyone called the technology AI — which means the governing rule for most systems now being built already exists somewhere in general administrative or data-protection law, and will be applied retrospectively the first time somebody complains.
The Dutch immigration service is the shape of that problem. Two assistants, Robin and ChatIND, have been helping roughly 230 caseworkers summarise documents, translate and analyse interviews in asylum processing since early 2026. Neither appears in the national algorithm register the service’s own transparency rules require.
Where public-sector AI deployment came first, it preceded the governing rule by a median of 22 months. The exceptions are not the slow adopters. They are the states whose administrative law already covered automated decisions.
Straits Institute for Applied AI
What this looks like in your country
The Institute maintains a country cut of the record for every country it tracks. Each is read on four axes, all computed from the entries rather than asserted, which is what lets two countries with near-identical deployment counts produce opposite profiles.
How far a country’s entries spread across the industries. Everything in one sector describes a state applying AI to its own administration. A spread profile describes an economy doing it in several places at once.
Computed: share held by the largest industry, and how many industries hold more than 5%
The in-house, local-vendor and global-platform mix from section 4. The strongest available signal of whether a country is developing capability or purchasing capacity.
Computed: share of entries by build posture
The citizen, internal and enforcement mix from section 5. Two countries with identical counts can be running opposite programmes, and this is where that shows.
Computed: share of entries by direction of use
Whether rules preceded deployments, and by how long. A governance reading taken from dates rather than from policy documents, which describe intent rather than practice.
Computed: interval between first regulation entry and first deployment entry, per industry
Regional cuts and the full filterable record sit with the applied AI statistics.
How we counted, and what the record misses
The record is swept weekly across national and local press, government and regulator publications, procurement notices and parliamentary records, in the working languages of each region. Candidates are tested against the four admission conditions in section 1. Admitted entries are filed by kind, industry, country and role, and carry the date and the source link.
Three limits, stated because a dataset that does not publish its blind spots is a marketing asset rather than a research one.
- Disclosure bias. The record measures what is published. Public bodies publish; private firms mostly do not. Comparisons across industries are comparisons of disclosure as much as of deployment.
- Language and press density. Countries with a dense local digital press are over-represented. Where deployments are reported in print, in minority languages, or not at all, they are under-counted.
- Self-reported performance. Where an entry carries a result, that result is almost always the deploying body’s own. Inclusion implies no independent evaluation.
The full method, including sweep sources, language coverage and the correction log, is published separately at the record’s method page. Corrections are accepted and logged; if an entry misstates a deployment you are responsible for, write to the research unit and it will be amended with the correction noted on the entry.
For journalists
Everything here is free to use. The Institute asks for attribution and a link, and would rather be quoted accurately than quoted often.
Contact
[ researcher name ], Head of the Industrial Research Unit, Straits Institute for Applied AI — research@straitsai.institute. Enquiries from journalists are answered the same working day. Available for attributable comment on any figure in this note.
Attribution line
Quotes, cleared for use
Attributable to [ researcher name ], Head of the Industrial Research Unit. Use as written or ask for a variant on your angle.
“We can tell you what has been switched on. Almost nobody can tell you what happened next.”
On the measurement gap · Figure 1
“The problem is not that AI is being evaluated badly. It is that in nine cases out of ten, it is not being evaluated at all.”
On the measurement gap · Figure 1
“Generative AI is the loudest category in this field, not the largest. Most of the AI actually doing a job is reading a camera or scoring a form.”
On applied generative AI · Figure 2
“If you want to know a country’s real AI capability, do not look at the compute it has bought. Look at who writes its government software.”
On build posture · Figure 3
“There is a difference between a system that helps you and a system that watches you, and it is not a technical difference. One in five of the systems we have verified is the second kind.”
On direction of use · Figure 4
“In most countries the rule is written after the system is already running. The rule that governs it usually existed all along, in administrative law, and nobody had thought to apply it yet.”
On the regulation lag · Figure 5
Data and charts
Every verified entry with kind, industry, country, role, date and source URL. Download
Figures 1 to 5 as PNG and SVG, credit line set inside the image. Linked under each figure above.
Every figure has a button that copies its underlying numbers as tab-separated values, ready for a spreadsheet.
Counts by industry, region, role and country, updated on every sweep. Applied AI statistics
Cuts on request
Handled by the research unit, usually within two working days: a country cut, a profession cut, a date-bounded cut for anniversary pieces, or a named-organisation check to verify a claim before publication.
Embed a chart
Paste this where the chart should appear. The credit and link travel with it.
<iframe src="https://www.straitsai.institute/embed/fig-1" width="100%" height="380" style="border:0" loading="lazy" title="Applied AI deployments by kind of evidence"></iframe> <p>Source: <a href="https://www.straitsai.institute/state-of-applied-ai">Straits Institute for Applied AI</a></p>
Citing this note
Common questions
What does applied AI mean?
Applied AI means artificial intelligence put to work inside a job somebody already holds, in an organisation that can be named, on a date that can be fixed, with a published source. Research asks what a model can do; applied AI asks what changes when a model is handed work somebody is currently paid to perform.
What are applied AI systems?
An applied AI system is any system doing that work, whatever technique it uses. In this record they divide into generative and language models, computer vision, prediction and scoring, speech and translation, and other machine learning. Applied AI systems are defined by placement rather than architecture: the same model is applied AI in a tax office and is not applied AI in a laboratory.
What is applied generative AI?
Applied generative AI, or applied gen AI, is the subset of applied AI that uses generative and language models in a real job: drafting, summarising, translating and answering. It accounts for just under half of the systems in this record. The rest of the field — vision, prediction, scoring — is larger in aggregate and carries most of the measured results.
How is applied AI different from machine learning?
Machine learning is a set of methods. Applied AI is a use of them inside an organisation. Almost all applied AI involves machine learning; almost no machine learning research is applied AI, because it has no organisation, no role and no date attached.
What does an applied AI practitioner actually do?
Four things recur across the record regardless of sector: choosing which task to hand over and which to keep; fitting the system to the existing workflow rather than the reverse; establishing what “working” means before deployment and measuring it after; and designing the human oversight the eventual rule will require. Building the model is rarely among them, because in most entries the model was bought.
Where can I see the full numbers?
On the applied AI statistics page, which carries counts by industry, region, role and country, the full method, and the record itself.