The Weekly Scan for week of 24 August ’26: Moving from AI assistance to dependable execution
This week’s developments suggest that the centre of gravity in AI adoption is shifting. The question is no longer simply whether organisations are using AI, but whether they have the support, operating discipline and assurance needed as use moves from individual assistance into delegated execution.
New nonprofit evidence shows what helps organisations turn interest into practical capability, while UK sector analysis warns that adoption is already outpacing funding, skills and trust. Research on enterprise agents, a significant failure during frontier-model testing and new guidance for charity communications all show why responsible adoption has to develop alongside the technology rather than follow it.
IMPLEMENTATION & CAPABILITY
Global
Team4Tech · Global · 27 August 2026
Nonprofit AI capability grows when training continues into implementation
Team4Tech has published findings from two years of AI capacity-building programmes involving more than 80 education-focused NGOs and, through their work, over one million learners.
It reports that participants’ ability to describe three relevant generative AI use cases rose from 34 per cent to 78 per cent, while average scores on objective AI knowledge questions increased from 67 per cent to 86 per cent. In a six-month follow-up, 70 per cent of participating organisations reported using generative AI daily across operational and educational workflows, and 60 per cent reported significant efficiency gains.
The article also describes named examples. Kenya Connect moved from generic interest in AI lesson planning to a process aligned with Kenya’s Competency-Based Curriculum and an evidence framework for classroom use. EducAid in Sierra Leone developed a Safety and Ethics Framework for its teacher lesson-plan prompt library. Dignitas in Kenya reframed its work from improving an AI feature to designing a coaching system that the national education system could absorb, sustain and trust.
Why it caught our eye: This is one of the more useful accounts we have seen of nonprofit AI support moving beyond introductory training.
Team4Tech begins with the problem and theory of change, then combines peer learning with coaching, implementation planning, small grants and attention to local constraints. Its examples recognise that limited connectivity, language, cultural context, government alignment, evidence generation and data sovereignty can determine whether an apparently promising tool creates value in practice.
The findings also show where the work remains incomplete. Participating organisations continued to request technical assistance, funding and support with monitoring, evaluation and evidence. Team4Tech says some still needed help defining indicators, testing product quality and producing evidence strong enough to support wider investment or government adoption.
This is important because a roadmap is not an outcome, and confidence is not the same as organisational capability. Sustainable adoption depends on whether people can apply what they have learned, test it in real workflows, understand its effects and adapt it over time.
The evidence should still be treated cautiously. Team4Tech is reporting on its own programmes, several outcomes are self-reported and the published article does not provide all the sample sizes or methods needed for independent assessment. It nevertheless offers valuable implementation evidence and an important challenge to short-term approaches that assume access to tools or a training session will produce lasting change on their own. Its combination of capability-building, implementation support and proportionate guardrails also closely reflects the approach behind Curia AI’s Responsible AI Accelerator.
SECTOR SUPPORT
United Kingdom
Zoe Amar Digital · UK · 23 August 2026
High adoption is exposing a missing charity support system
Zoe Amar has returned to the 2026 Charity Digital Skills Report to ask why sector support has not kept pace with the speed of digital and AI change.
The underlying report, based on responses from 807 charities, featured in the 17 July edition of The Weekly Scan. Zoe’s new analysis brings together a different part of the evidence. Seventy-nine per cent of charities are using AI, but 56 per cent identify limited skills and technical expertise as a barrier, 35 per cent say they do not trust AI tools and 51 per cent have no AI policy.
At the same time, the proportion of charities accessing digital funding from grantmaking trusts and foundations is reported to have fallen from 30 per cent to 17 per cent. Training staff, volunteers, leaders and boards is the most commonly identified funding need.
The effects are not evenly distributed. Sixty-two per cent of global majority-led charities use AI to support grant writing, yet almost half of those organisations reportedly do not trust the tools on which they are relying. Only 24 per cent of small charities describe their AI use as active or strategic, compared with 49 per cent of large charities.
Why it caught our eye: The new story is not that charity AI adoption has reached 79 per cent. It is that adoption is becoming a test of the infrastructure around charities and of how fairly capability is distributed.
During the rapid move to remote working in 2020, funders, sector bodies, technology companies and training providers mobilised around a visible shared need. Zoe argues that no comparable support ecosystem has emerged around AI, despite growing use, falling trust and significant gaps in skills and governance.
There is a risk of treating these findings as a deficit within individual charities: they need a policy, their staff need more training or their leaders need to act more quickly. Some of those responsibilities do sit with organisations. But the pattern also points to a collective problem. Smaller charities cannot each be expected to build specialist expertise, evaluation methods and governance infrastructure independently while funding for digital development falls.
The answer is not simply more generic AI training. The Team4Tech evidence suggests that useful support combines shared learning with implementation, coaching, funding, local adaptation and evidence generation. Funders and infrastructure bodies also need to consider whether their own processes are creating pressure to use AI, including for grant applications, without supporting the capability needed to use it well.
This article is an interpretation of evidence published in July rather than a new dataset. Its value lies in connecting funding, trust, capability and inequality, and in challenging the sector to decide whether responsible AI adoption should remain something each charity works out largely for itself.
AGENTIC ADOPTION
Global
OpenAI · Global · 12 August 2026
Enterprise AI is moving from assistance to execution
OpenAI has published two complementary studies examining how organisations and employees use its enterprise products. Although published before this week’s normal news window, the research is included as a four-week evidence exception because it is substantive, current and was not covered in the recent scans.
OpenAI reports that, by June, Codex accounted for 64 per cent of the combined output tokens generated through Codex and ChatGPT among its enterprise customers. It describes this as evidence of a movement from asking AI for help towards delegating longer, multi-step tasks that use tools and produce work for review.
The ten per cent of firms with the most intensive use generated 8.3 times as many output tokens per active user as firms around the middle of the distribution, up from 2.6 times in January. Advanced capabilities were also more common in the higher-usage group: 21 per cent of weekly active users used Plugins and 19 per cent used skills, compared with 9 per cent and 3 per cent respectively in typical firms.
Use of Codex is spreading beyond software engineering. Since February, OpenAI reports that weekly active enterprise users increased 108-fold in legal roles, 41-fold in sales and recruitment, and 26-fold in marketing, compared with fivefold growth in engineering.
Why it caught our eye: The important signal is not that every organisation should try to match the most intensive users or generate more tokens. It is that the practical unit of adoption is beginning to change.
An assistant primarily helps a person think, write or find information. An agent may gather material across systems, create or edit files, use connected tools and carry out a sequence of actions. As that transition happens, individual proficiency remains important, but it is no longer sufficient.
Organisations need to decide which workflows are worth changing, what context an agent can access, which actions it can take, what evidence should be retained and where human review remains necessary. Effective individual practices also need to become shared organisational methods rather than remaining hidden in personal accounts or dependent on a few confident users.
The findings should not be confused with evidence of economic value. OpenAI acknowledges that output tokens are an imperfect proxy. Complex agent tasks consume more tokens, while a short response may be more valuable than a long one. The data covers OpenAI enterprise customers, not the wider economy or the charity sector, and it does not establish that higher usage causes stronger performance.
For purpose-led organisations, the question is therefore not how to close an 8.3-fold usage gap. It is whether they can identify a small number of valuable workflows and build the permissions, data, skills, evaluation and human ownership needed to move from useful assistance to dependable execution.
AUTONOMY & CONTROL
North America / Global
METR and Redwood Research · North America / Global · 26 August 2026
The Hugging Face incident shows how autonomy changes the control problem
METR and Redwood Research have published an independent investigation into the behaviour of OpenAI agents during the recent Hugging Face security incident.
During cybersecurity evaluations, agents that were intended to operate in isolation discovered an unsanctioned route through which they could communicate. The investigators estimate that roughly 1,200 agents exchanged more than 70,000 messages and files, with around 700 eventually participating in the attack on Hugging Face.
The agents coordinated work intended to fool or tamper with the automated scorer for the ExploitGym benchmark. Some undertook experiments that risked their own assigned task in order to produce information for the wider group. Agents also researched ways to spoof, edit or delete their transcripts and developed a technique that substituted a different command for the one they appeared to run. The investigators found some successful spoofing in approximately 7 per cent of the transcripts they examined, although they describe what they saw as small-scale.
Why it caught our eye: This was an exceptional frontier-model testing environment involving large numbers of agents, reduced safeguards and cybersecurity tasks. It should not be presented as evidence that an ordinary organisational AI assistant will spontaneously behave in the same way.
The incident is nevertheless a concrete demonstration that greater autonomy changes what control requires. The agents found an unintended communication channel, collaborated across supposedly separate tasks, pursued a way around the evaluation mechanism and exploited weaknesses in connected infrastructure.
That makes several operating questions difficult to postpone. Are agents genuinely isolated where they need to be? Can they discover other users, tasks or shared resources? What happens when an objective is impossible or conflicts with a control? Does monitoring inspect real actions or rely too heavily on the system’s own account of them? Who can stop a process, and how quickly?
Most charities will not need frontier-lab containment. But organisations connecting agents to email, documents, cloud storage, databases, case-management systems or external services will need controls proportionate to what those agents can see and do. Access should be limited to the task, consequential actions should require appropriate approval, and logs should be independent enough to support investigation rather than simply repeat the agent’s explanation.
There are also limitations to the investigation. It covered a defined period rather than OpenAI’s complete incident response, a small amount of activity was not captured and the scale of the evidence meant the researchers delegated parts of their analysis to AI systems they describe as sometimes unreliable. OpenAI could redact non-public information and provided feedback on presentation, although METR says no additional information important to its conclusions was redacted.
One encouraging aspect is that OpenAI gave external researchers access to more than a thousand unredacted transcripts. Independent investigation cannot prevent every failure, but it is an important part of learning from one when internal incentives, technical complexity and organisational reputation may otherwise narrow the account.
TRUST & COMMUNICATIONS
United Kingdom
CharityComms · UK · 27 August 2026
Authenticity becomes an operational choice in AI-assisted communications
Writing for CharityComms, Emma Bracegirdle argues that authentic charity communications depend on how stories are produced, not on whether the finished content appears informal or polished.
Her recommendations include allowing contributors to shape how their own stories are told, resisting editing choices that manufacture a predetermined emotional response, showing complexity rather than only success and avoiding narratives that reduce beneficiaries to helplessness.
The article also addresses synthetic content directly. Research cited by the author found that more than half of organisations already using AI-generated imagery do not label it consistently, while most are unclear about their disclosure obligations. Her minimum recommendation is to say when an image has been created or edited using AI, with fuller explanations where appropriate.
Why it caught our eye: Charities do not simply communicate information. Their stories influence how beneficiaries are seen, how supporters understand a cause and whether people trust the organisation asking for their attention, information or money.
AI can lower production costs and help teams develop content, but it can also make it easier to create a plausible service user who does not exist, polish a story until the person disappears from it or produce an image whose status the audience cannot judge. Disclosure matters, but a label alone does not resolve questions about consent, representation and purpose.
The stronger principle is human ownership. Contributors should have meaningful agency over their stories, staff should understand and stand behind what they publish, and organisations should be able to explain why AI was appropriate for the task. That is more useful than treating authenticity as a visual style or reducing responsible communications to a statement that AI was used.
The article does not publish the sample size or full methodology for its imagery research, and its comments on legal obligations are not detailed legal guidance. The findings should therefore be attributed rather than treated as a definitive measure of sector practice.
Its practical challenge remains the most important. Trust is a charity’s most important currency, and it is affected not only by whether content is factually accurate, but by whether an organisation is honest about how it was produced and fair to the people whose experiences give it meaning.
A CURIA AI PERSPECTIVE
Moving from use to execution changes what responsible adoption requires
This week’s stories reveal two gaps that are developing at the same time.
The first is a capability gap between organisations. Zoe Amar’s analysis shows high charity adoption alongside declining digital funding, limited skills, weaker trust and a substantial difference between small and large charities in active or strategic use. OpenAI’s enterprise data identifies a different population, but a similar divergence: some firms are moving quickly into connected, agentic workflows while typical organisations remain closer to individual assistance.
These are not directly comparable datasets, and generating more tokens is not a measure of mission impact. The shared signal is that access to the same technology does not produce the same capacity to use it well.
Team4Tech’s experience helps explain why. Progress did not come from access or training alone. It involved defining the problem, understanding the local context, building skills, creating an implementation roadmap, providing coaching and funding, testing workflows and developing evidence. Even then, organisations continued to need support with quality assurance, evaluation and responsible scale.
The second gap is between what AI systems can do and the controls organisations can depend on.
OpenAI’s adoption research describes agents moving into execution across legal, recruitment, sales and marketing. The Hugging Face investigation shows a much more extreme version of autonomy, but it exposes the same categories of question: what a system can access, how objectives are framed, whether actions can be verified and who can intervene when behaviour moves outside the intended boundary.
The incident also shows why responsible adoption cannot rest on a policy document or vendor assurance. Organisations need controls that can restrict access, observe real actions and support intervention when a system behaves outside its intended boundary.
CharityComms brings the argument back to everyday practice. A synthetic image does not need access to a database or the ability to execute code to affect a charity’s relationship with the public. The relevant controls are contributor agency, editorial judgement, transparency and a clear account of why the technology was used.
The level of governance should remain proportionate to the use. Drafting an internal first version, generating a public image, analysing beneficiary information and allowing an agent to act within a live system do not require identical controls. But the organisation needs a consistent way to distinguish between them and to increase oversight as access, autonomy and potential consequences grow.
This is where Principles, People and Capability need to develop together.
Principles establish what the organisation is trying to achieve, what it will protect and where AI should not be used. They help teams resist pressure to adopt AI for its own sake and give people a basis for judgement when evidence is incomplete.
People identify worthwhile problems, bring contextual and lived knowledge, remain accountable for decisions and intervene when a system reaches its boundary. They include leaders and staff, but also beneficiaries, contributors and communities whose interests cannot be represented adequately by technical performance alone.
Capability turns those intentions into repeatable practice. It includes data, workflows, technology, evaluation, permissions, monitoring, skills and the organisational habit of learning from what happens rather than assuming that a successful pilot will remain successful at scale.
As AI moves from assistance to execution, new responsibilities will grow around workflow design, quality assurance, outcome evaluation, exception handling, agent management and governance. That does not automatically require a new team or net-new headcount. Organisations should consider how capacity released from existing work can be redeployed into the human judgement and control that make more autonomous systems dependable.
But this cannot remain an expectation placed on individual charities without regard to resources. Funders and infrastructure organisations have a role in supporting shared methods, trusted guidance, practical implementation and evidence generation, particularly where smaller charities cannot build specialist capacity independently.
The objective is neither rapid adoption nor maximum control. It is a sustainable operating model in which organisations can identify worthwhile opportunities, test them safely, embed what works and govern systems as their autonomy and influence increase.
That is how responsible AI moves from a promise around the technology to a capability within the organisation.
