Thirty-one of 36 OECD countries now use AI for internal government processes, but only 13 use it to support policymaking. That roughly 50-point gap between back-office adoption and policy work is the clearest signal in public sector AI data: governments have embraced the technology where consequences are contained and hesitated where decisions affect citizens. Meanwhile OECD reports public servants adopting generative AI faster than oversight mechanisms can keep pace. The figures below come from the OECD Digital Government Outlook and related public governance research.
TL;DR
- 31 of 36 OECD countries use AI for internal processes, about 86% (OECD)
- That is up from 23 of 33, roughly 70%, two years earlier (OECD)
- 27 of 36 use AI in public services, about 75% (OECD)
- Up from 22 of 33, roughly 67% (OECD)
- Only 13 of 36 use AI to support policymaking, about 36% (OECD)
- Up from 11 of 33, roughly 33% (OECD)
- Internal adoption rose about 16 percentage points (derived)
- Public service adoption rose about 8 points (derived)
- Policymaking adoption rose about 3 points (derived)
- The internal-to-policymaking gap is about 50 points (derived)
- Internal skills gaps are the most cited barrier (OECD)
- Public servants adopt generative AI often without formal oversight (OECD)
- The country base grew from 33 to 36 between rounds (OECD)
1. Internal Adoption Is Nearly Universal
The back-office numbers show a technology that has effectively won its first argument. 31 of 36 OECD countries, about 86%, now use AI for internal processes, up from 23 of 33 or roughly 70% in the earlier measurement round.
A 16-point rise in two years puts internal AI use past the point where it can be described as experimental. Only five measured countries report not using it at all. The pattern makes sense given where the risk sits: document processing, translation, scheduling, and internal search are domains where an error is caught by the person who requested the output, and no citizen is affected by a bad result. Governments have adopted AI most readily in precisely the settings where the accountability question does not arise, which is a rational sequencing rather than a timid one.
| Metric | Value | Source |
|---|---|---|
| Countries using AI for internal processes | 31 of 36 | OECD |
| Share | approx. 86% | OECD |
| Earlier round | 23 of 33 | OECD |
| Earlier share | approx. 70% | OECD |
| Change | up approx. 16 points | Derived from OECD figures |
| Countries not using AI internally | 5 of 36 | Derived from OECD figures |
| Share not using | approx. 14% | Derived from OECD figures |
| Status of internal use | past the experimental stage | Derived |
Source: OECD Digital Government Outlook 2026, adopting and governing AI in government.
2. Public Services Trail Internal Use
Citizen-facing deployment sits a clear step behind. 27 of 36 countries, about 75%, use AI in public services, up from 22 of 33 or roughly 67%.
The 8-point rise is half the pace of internal adoption, and the resulting 11-point gap between the two categories has widened rather than closed. That divergence is informative. Public service deployment brings in legal exposure, equality obligations, appeal rights, and the practical difficulty of explaining an automated decision to someone it went against. Nine countries still report no use here at all. The gap between 86% and 75% is not large in absolute terms, but the fact that it grew while both categories rose suggests the constraints on citizen-facing use are structural rather than a matter of governments simply catching up.
| Metric | Value | Source |
|---|---|---|
| Countries using AI in public services | 27 of 36 | OECD |
| Share | approx. 75% | OECD |
| Earlier round | 22 of 33 | OECD |
| Earlier share | approx. 67% | OECD |
| Change | up approx. 8 points | Derived from OECD figures |
| Countries not using AI in public services | 9 of 36 | Derived from OECD figures |
| Gap versus internal use | approx. 11 points | Derived from OECD figures |
| Direction of that gap | widening | Derived from OECD figures |
Source: OECD.AI resources on AI in government.
3. Policymaking Has Barely Moved
The third category is where adoption stalls almost completely. Only 13 of 36 countries, about 36%, use AI to support policymaking, up from 11 of 33 or roughly 33% across the same period.
A 3-point increase over two years is close to flat, and against a growing denominator it is arguably no movement at all. Twenty-three countries report no such use. OECD also notes that adoption remains limited in accountability activities specifically. The reluctance is defensible on its own terms, since policy analysis involves contested value judgements, causal claims that resist verification, and outputs that must survive legislative and judicial scrutiny, none of which suit a system whose reasoning cannot be fully inspected. But the practical result is that AI is transforming how governments administer while leaving how they decide almost untouched.
| Metric | Value | Source |
|---|---|---|
| Countries using AI in policymaking | 13 of 36 | OECD |
| Share | approx. 36% | OECD |
| Earlier round | 11 of 33 | OECD |
| Earlier share | approx. 33% | OECD |
| Change | up approx. 3 points | Derived from OECD figures |
| Countries not using AI in policymaking | 23 of 36 | Derived from OECD figures |
| Gap versus internal use | approx. 50 points | Derived from OECD figures |
| Adoption in accountability activities | also limited | OECD |
Labour market context sits in our AI impact on jobs statistics. Source: OECD Digital Government Outlook 2026.
4. The Governance Gap
The most consequential finding is not a percentage but a mismatch. OECD reports that public servants are increasingly adopting generative AI often without formal oversight, describing a widening gap between rapid decentralised uptake and lagging public governance mechanisms.
Surveys indicate a significant proportion of public servants already use open-access generative AI tools to support their work. This is shadow adoption in the classic sense: staff solving immediate problems with whatever is available, ahead of any policy telling them whether they may. The specific hazard OECD identifies is data protection, since using external tools in ways that conflict with organisational policies can move government data outside controlled environments. What makes this harder than a typical shadow IT problem is that the tools are free, require no procurement, and leave little trace, so an institution can have substantial AI usage without any record that it is happening.
| Metric | Value | Source |
|---|---|---|
| Pattern identified | decentralised uptake outpacing governance | OECD |
| Formal oversight | often absent | OECD |
| Public servants using open-access generative AI | a significant proportion | OECD |
| Primary risk named | data protection | OECD |
| Mechanism of risk | conflict with organisational policies and regulations | OECD |
| Procurement required | none for open-access tools | Derived |
| Visibility to the institution | low | Derived |
| Direction of the gap | widening | OECD |
Source: OECD working paper on generative AI experimentation in government.
5. Skills Are the Binding Constraint
When governments explain what holds them back, one answer dominates. Internal skills gaps are consistently cited as the most significant barrier to AI adoption in the public sector, ahead of time pressure, risk concerns, and limited capacity for innovation.
That ordering is worth taking seriously because it is not the answer most external commentary assumes. Public discussion tends to frame the obstacle as regulatory caution or budget, but administrations themselves point at capability. OECD’s workforce analysis identifies three distinct groups within the public workforce, each requiring a fundamentally different approach to training, which challenges the one-size-fits-all programmes many institutions have deployed. The practical implication is that generic AI literacy training is unlikely to shift the constraint, since the person who needs to evaluate a procurement, the person who needs to use a tool safely, and the person who needs to build one require entirely different preparation.
| Metric | Value | Source |
|---|---|---|
| Most cited barrier | internal skills gaps | OECD |
| Other barriers cited | time pressure from competing tasks | OECD |
| Also cited | concerns about generative AI risks | OECD |
| Also cited | limited capacity for innovative approaches | OECD |
| Consequence | promising applications slow to be identified and scaled | OECD |
| Distinct workforce groups identified | 3 | OECD |
| What each group requires | a fundamentally different training approach | OECD |
| Approach being challenged | one-size-fits-all training | OECD |
Education-sector parallels sit in our AI in education statistics, and model-availability context in our open source AI statistics. Source: OECD, building an AI-ready public workforce.
6. What the Denominators Hide
A methodological detail changes how these trends should be read. The country base grew from 33 to 36 between measurement rounds, which means every year-on-year percentage comparison mixes genuine adoption growth with a changed denominator.
In country counts rather than shares, internal use rose by 8 countries, public services by 5, and policymaking by 2, against 3 countries newly added to the measurement. Depending on whether those three were already using AI when they joined, some portion of the apparent increase reflects a wider survey rather than new adoption. This does not overturn the headline story, since an 8-country rise cannot be explained by 3 new entrants, but it does mean the policymaking figure is the shakiest of the three. A 2-country increase against 3 new entrants is fully consistent with no underlying change at all among the originally measured group.
| Metric | Value | Source |
|---|---|---|
| Countries measured, earlier round | 33 | OECD |
| Countries measured, later round | 36 | OECD |
| Countries added | 3 | Derived from OECD figures |
| Increase in internal use, country count | 8 | Derived from OECD figures |
| Increase in public services, country count | 5 | Derived from OECD figures |
| Increase in policymaking, country count | 2 | Derived from OECD figures |
| Which trend is least robust | policymaking | Derived |
| Data collection method | self-reporting by national administrations | OECD |
Source: OECD.AI publications on AI in government.
Summary: AI in Government by the Numbers
| Metric | Value | Source |
|---|---|---|
| Internal processes, later round | 31 of 36 (approx. 86%) | OECD |
| Internal processes, earlier round | 23 of 33 (approx. 70%) | OECD |
| Change | up approx. 16 points | Derived |
| Public services, later round | 27 of 36 (approx. 75%) | OECD |
| Public services, earlier round | 22 of 33 (approx. 67%) | OECD |
| Change | up approx. 8 points | Derived |
| Policymaking, later round | 13 of 36 (approx. 36%) | OECD |
| Policymaking, earlier round | 11 of 33 (approx. 33%) | OECD |
| Change | up approx. 3 points | Derived |
| Internal-to-policymaking gap | approx. 50 points | Derived |
| Countries not using AI internally | 5 of 36 | Derived |
| Countries not using AI in public services | 9 of 36 | Derived |
| Countries not using AI in policymaking | 23 of 36 | Derived |
| Most cited barrier | internal skills gaps | OECD |
| Governance pattern | uptake outpacing oversight | OECD |
| Distinct workforce groups identified | 3 | OECD |
| Country base change | 33 to 36 | OECD |
Methodology and Sources
- Adoption counts and shares for internal processes, public services, and policymaking come from the OECD Digital Government Outlook chapter on adopting and governing AI in government, drawing on the Digital Government Index (OECD, OECD.AI resources, OECD.AI publications).
- Governance gap findings and generative AI experimentation patterns come from OECD Working Papers on Public Governance No. 93 (OECD working paper PDF, OECD publication page).
- Workforce and skills findings come from OECD’s public workforce analysis (OECD full report, OECD publication landing page).
- Data watch: all adoption figures are self-reported by national administrations responding to an OECD survey, so a country reporting AI use is asserting it rather than demonstrating it, and the threshold for counting as a user is not standardised across respondents. The measured country base grew from 33 to 36 between rounds, which means percentage comparisons across years are not strictly like-for-like and the smallest of the three trends, policymaking, is consistent with no real change among originally measured countries. Percentages here are calculated from the published country counts and rounded. Several supporting OECD publications sit behind access restrictions, so barrier rankings are reported qualitatively as OECD describes them rather than with response percentages. Rows marked as derived are arithmetic on published counts.
- Last updated: August 2, 2026. We update this roundup quarterly, and the next major refresh is expected when OECD publishes an updated Digital Government Index round.