Short notes on making technology deliver
Practical views from our team, and from the best public research on AI.
All articles
AI agents are moving from reading to acting
What changes when AI can act, not just answer.
4 min readArticleAsk, don't tell: getting honest answers from AI
Three habits that stop AI agreeing with you.
3 min readArticleAI saves time on some tasks and none on others. Measure which
What a 500-person trial says about real gains.
3 min readArticleA practical way into the NIST AI Risk Management Framework
Four functions, four leadership questions.
3 min readArticleISO/IEC 42001 or the NIST AI RMF: which do you need?
Guidance versus certification, and how to choose.
3 min readArticleFive generative AI risks every leadership team should name
Where generative AI risk sits for users.
3 min readLeadershipBefore you deploy AI, make the case in writing
A one-page habit borrowed from safety-critical industries.
3 min readArticleWhy AI pilots stall before they reach operations
Three patterns, and the fix.
4 min readArticleAI governance starts with an inventory, not a policy
Find the roads before you set the rules.
3 min readLeadershipMeasure open innovation by adoption, not by challenges launched
What separates results from events.
4 min readForesightAI agents are moving from reading to acting
By Haytham AliRead
Most organizations still think of AI as something that answers questions. That picture is already out of date. AI agents now connect to tools that let them change things: edit files, update records, send messages and even move money.
The UK AI Security Institute studied more than 177,000 tools built for AI agents between late 2024 and early 2026. It sorted them into tools that read data, tools that analyze it, and tools that act on the world outside the system. Over sixteen months, the share of usage going to action tools rose from 27% to 65%. Most support medium-stakes tasks such as editing files, but some already handle financial transactions.
The question is no longer what your AI knows. It is what your AI is allowed to do.
Security has not kept pace. In a large public red-teaming competition reported by the Institute, participants launched 1.8 million attacks against 22 AI agents in 44 realistic scenarios. More than 60,000 attacks caused policy violations such as unauthorized data access or illicit financial actions. Nearly all agents broke policy for most tested behaviors within 10 to 100 attempts, and larger, more capable models were not reliably safer.
Even professional testers have been caught out. In August 2026 the Institute disclosed an incident in which AI agents, during a cyber evaluation with internet access switched on, took actions on the live internet that no one had sanctioned, including an attempt to slip malicious code into a public software project. A human reviewer blocked it and no real-world harm was found. The Institute's response is worth copying: internet access must now be actively justified rather than granted by default, and monitoring will flag or block out-of-scope actions as they happen.
What to do now
First, inventory agent tools the way you inventory user permissions, and label each one: read, analyze or act. Second, make access a decision, not a default. Any action that touches money, personal data, external parties or official records should need explicit approval. Third, monitor in real time and keep logs you can actually investigate. The Institute warns that our ability to audit, monitor and investigate AI systems is likely to erode without deliberate effort, so build that ability in from the start.
Agents will bring real productivity gains. The organizations that benefit most will be those that decided in advance what their agents may do, and can show it.
Sources
- How are AI agents used? Evidence from 177,000 MCP tools · UK AI Security Institute, March 2026
- Security challenges in AI agent deployment: insights from a large-scale public competition · UK AI Security Institute, July 2025
- Incident report: unsanctioned agent behaviour during cyber testing · UK AI Security Institute, August 2026
- Loss of Oversight: how AI systems may become harder to audit, monitor and investigate · UK AI Security Institute, May 2026
ArticleAsk, don't tell: getting honest answers from AI
By Haytham AliRead
AI assistants have a well-known weakness: they tend to agree with you. Researchers call it sycophancy. For leaders who use AI to test a strategy or review a decision, it is a real risk, because the tool can hand your own view back to you with added confidence.
Recent research from the UK AI Security Institute looked at what triggers this behavior. Three findings stand out. Models were far more sycophantic when users made statements than when they asked questions. Sycophancy rose steadily with how certain the user sounded. And framing a point from your own perspective, "I think…", made it worse.
Tell an AI what you believe, and it will often tell you that you are right.
The remedy is simple. The researchers found that having the model restate a user's statement as a question before answering significantly reduced sycophancy, and worked better than simply instructing it not to be sycophantic.
Three habits for leaders
Ask, don't assert. "What are the risks of entering the Saudi market next year?" will get you a better answer than "We should enter the Saudi market next year."
Hold back your preference. Let the tool assess every option before you say which one you favor.
Ask for the opposing case. Request the strongest argument against your plan, then weigh both sides yourself.
These habits cost nothing. They matter more each month as AI finds its way into board papers, investment memos and policy drafts. A tool that only agrees with you is no advisor.
Source
- Ask don't tell: reducing sycophancy in large language models · UK AI Security Institute, April 2026
ArticleAI saves time on some tasks and none on others. Measure which
By Haytham AliRead
Most AI business cases assume a single productivity number applied across the board. The evidence says gains are real but uneven, and the unevenness is where good decisions are made.
In a pilot study with the UK government's Future of Work Unit, the UK AI Security Institute gave 500 participants four common workplace tasks and randomly split them into a group with AI tools and a group without. Across all four tasks, the AI group produced work of 19% higher quality in 25% less time. Task by task, the picture changed. Drafting technical specifications was 34% faster and 23% better. Interpreting information for others was 42% faster, with no gain in quality. Organizing, planning and prioritizing work showed no significant gain at all.
AI is strongest on structured, well-defined work and weakest on open-ended planning. Your business case should say which is which.
The researchers describe this as the "jagged" capability of AI, and they are careful to call the results preliminary: the sample was small and the uncertainty high. That caution is itself a lesson. If a national research institute treats its own numbers as provisional, a vendor's single headline figure deserves the same care.
What this means for your AI investment
Measure by task, not by tool. Break a role into its main tasks and test AI on each one separately.
Use a comparison group. Even a small pilot with and without AI tells you more than a satisfaction survey.
Track quality and time separately. Faster is not always better, and better is not always faster.
Organizations that do this will put AI where it pays and stop forcing it where it does not. That alone can decide whether an AI program earns its budget.
Source
- AI and the future of work: measuring AI-driven productivity gains for workplace tasks · UK AI Security Institute, February 2026
ArticleA practical way into the NIST AI Risk Management Framework
By Haytham AliRead
Boards and regulators in the Gulf and the United States are asking the same question: how are you managing AI risk? The NIST AI Risk Management Framework is one of the most useful places to start. It is voluntary and free, and it was built with input from more than 240 organizations across industry, academia, civil society and government.
Its core is four functions. Govern sets the culture, roles and policies, and runs through everything else. Map establishes the context of each AI system so its risks can be framed. Measure uses quantitative and qualitative methods to assess and track those risks. Manage puts resources against the risks that matter most, on a regular cycle.
Govern, map, measure, manage. Four words that turn "we should be careful with AI" into work someone can own.
Four questions for your next leadership meeting
Govern: Who is accountable for AI risk, and where are our decisions recorded?
Map: For each AI system, what is it for, who does it affect, and what could go wrong?
Measure: How do we know it works as intended, and what evidence do we keep?
Manage: Which risks are we treating first, and who reviews them, how often?
Three practical points
It is being revised. NIST states that a revised version of AI RMF 1.0 is in progress. That is no reason to wait. The four functions are stable ideas, and work done now will carry over.
One effort can serve several frameworks. NIST publishes crosswalks that map the framework to other standards, including ISO/IEC 42001 for AI management systems and ISO/IEC 23894 for AI risk management. For organizations working across the US and the GCC, that means less duplicated work.
Start with a gap analysis. NIST's collection of real-world examples includes a city government's playbook and a gap-analysis template shared by Google DeepMind. Comparing what you already do against the framework is a faster first step than writing policy from scratch.
Frameworks do not reduce risk by themselves; people with clear roles do. What the AI RMF offers those people is a shared structure and a common language.
Sources
- AI Risk Management Framework · NIST AI Resource Center
- Crosswalk documents · NIST AI Resource Center
- Examples of use cases · NIST AI Resource Center
ArticleISO/IEC 42001 or the NIST AI RMF: which do you need?
By Haytham AliRead
Two names come up in almost every AI governance conversation: the NIST AI Risk Management Framework and ISO/IEC 42001, the international standard for AI management systems published in December 2023. Leaders often ask which one to adopt. The honest answer is that they do different jobs, and most organizations will benefit from both.
The NIST AI RMF is guidance. It is voluntary and free, and organized around four functions: govern, map, measure and manage. It helps you think clearly about AI risk and build good practice. There is no certificate at the end of it.
ISO/IEC 42001 is a certifiable management system standard. An independent certification body can audit your organization and certify that your management system for developing, providing or using AI meets the standard. As the US accreditation body ANAB points out, the certificate covers the organization's management system, not any individual AI product.
NIST helps you build the practice. ISO/IEC 42001 lets you prove it to others.
The overlap is large. A crosswalk published on NIST's AI Resource Center maps every part of the NIST framework to clauses and controls in ISO/IEC 42001 (it was prepared against the standard's final draft). Work you do under one will carry much of the way to the other.
How to choose
Start with who is asking. If the push is internal (a board wanting assurance, a team wanting structure), the NIST framework is a fast, low-cost start. If customers, partners or tenders are asking for evidence, certification is likely where you are heading.
Scope it tightly. Decide which AI systems and business units are in scope before anything else. A narrow, well-run scope beats a broad one that exists on paper.
Reuse what you have. If you already run an ISO/IEC 27001 information security management system, much of the machinery, such as policies, internal audit and management review, will feel familiar.
Check the certifier's accreditation. Accreditation bodies such as ANAB publish directories of certification bodies accredited for ISO/IEC 42001. And be wary of anyone who offers to both prepare you and certify you; good practice keeps those roles separate.
A certificate is not the goal in itself. The goal is AI that your organization can explain, control and stand behind. Done well, both frameworks get you there, and the certificate becomes the evidence rather than the point.
Sources
- NIST AI RMF to ISO/IEC 42001 crosswalk · NIST AI Resource Center
- AI Risk Management Framework · NIST AI Resource Center
- ISO/IEC 42001 AI management systems · ANAB, ANSI National Accreditation Board
- ISO/IEC 42001:2023 · ISO (the standard itself)
ArticleFive generative AI risks every leadership team should name
By Haytham AliRead
Generative AI brings risks that older systems did not. NIST's Generative AI Profile, a companion to its AI Risk Management Framework, sets out the risks that are unique to generative AI or made worse by it. Some, such as information on chemical or biological weapons, mainly concern the companies that build models. Others land squarely on any organization that uses these tools. Five stand out.
Confabulation. NIST's term for confidently stated but false content, often called hallucination. Ask: where would a wrong answer cause harm, and who checks it?
Data privacy. Personal or sensitive data can leak through prompts, outputs or training. Ask: what data may staff put into which tools?
Information security. Generative AI widens the attack surface, from manipulated inputs to misuse of the tools themselves. Ask: has security reviewed every AI tool connected to our systems?
Intellectual property. Outputs may reproduce protected material, and inputs may expose your own. Ask: who owns what our tools produce, and what are we giving away?
Value chain and component integration. Most organizations build on third-party models and data they do not control. Ask: what do our vendors guarantee, and what happens when their model changes?
For most organizations, generative AI risk sits less in what they build than in what they buy.
One more deserves attention: what NIST calls human-AI configuration. People tend to over-rely on systems that sound confident and to dismiss ones they distrust. Clear review rules and training address this better than technology does.
Naming risks is the first step in managing them. A short list agreed by leadership and owned by named people is worth more than a long policy no one reads.
Sources
- NIST AI 600-1: Artificial Intelligence Risk Management Framework, Generative AI Profile · NIST, July 2024
- Technical reports · NIST AI Resource Center
LeadershipBefore you deploy AI, make the case in writing
By Haytham AliRead
In aviation and other safety-critical industries, new systems are not approved on enthusiasm. Their owners present a safety case: a structured argument, backed by evidence, that the system is safe enough for its intended use. The UK AI Security Institute has argued that safety cases can help show decision-makers that AI is safe, and has published example cases for frontier AI.
Most organizations do not need the frontier version. The discipline, however, transfers well to everyday AI approvals. Before a system goes live, someone should be able to write on one page what the system is claimed to do safely, what evidence supports each claim, and who owns the risk if a claim proves wrong.
If no one can write the case, no one should sign the approval.
Evidence has to reflect real use
Take customer-facing chatbots. Many organizations require them to disclose that they are AI when asked. The Institute's RealityTest study gathered more than 3,000 real identity questions from about 750 people in 49 countries and five languages. People asked in far more varied ways than machine-generated test questions did, and only 31% asked directly when the situation was ambiguous. Phrasing and context mattered more to whether a model disclosed than which model it was, and a single instruction could push disclosure below 30%, even in the best models.
For leaders, the lesson is to test with your real users, in their languages. In the Gulf that means Arabic and English at a minimum, across dialects, and in voice as well as text. A test set written by your vendor is a starting point. It is not your evidence.
A one-page case per system takes an afternoon. It is the quickest way to find out whether anyone really knows how the system will behave, before your customers find out for you.
Sources
- Safety cases: a scalable approach to frontier AI safety · UK AI Security Institute, February 2025
- An example safety case for safeguards against misuse · UK AI Security Institute, June 2025
- RealityTest: how people probe AI identity and whether models disclose it · UK AI Security Institute, June 2026
ArticleWhy AI pilots stall before they reach operations
By Haytham AliRead
Most organizations I speak with are not short of AI pilots. They are short of pilots that became part of how work gets done. The technology usually works. What fails is everything around it.
Three patterns come up again and again. First, the pilot was chosen because it was interesting, not because someone owned a business problem it would solve. When the demo ends, nobody is accountable for taking it further. Second, the pilot was tested on clean data in a protected environment. Real processes have exceptions, approvals and legacy systems, and nobody planned for them. Third, there was no decision point. A pilot without a clear "scale, change or stop" moment tends to drift until the budget runs out.
A pilot is a decision instrument. If it is not designed to support a decision, it is a demonstration.
The fix is not more pilots. It is fewer, better-chosen ones. Before starting, name the business owner, the measure of success and the date the decision will be made. Map the process the solution must fit into, including the exceptions. And budget for the work after the pilot: integration, training, support and governance often cost more than the pilot itself.
Organizations that do this well tend to run a short portfolio review first. Which opportunities matter most, what will each one take, and which three are worth doing now? That review is rarely exciting, but it is where value is decided.
ArticleAI governance starts with an inventory, not a policy
By Haytham AliRead
When leadership decides it is time to "govern AI", the first instinct is often to write a policy. A policy matters, but writing one first is like setting traffic rules before you know where the roads are.
Start by finding out where AI is already in use. In most organizations the answer is: in more places than anyone expected. It is inside productivity tools, embedded in vendor platforms, used by staff on personal accounts and built into small automations no one documented. Each of these touches data, decisions or customers in a different way.
A simple inventory answers five questions for each use: what it does, who owns it, what data it can access, what decisions it influences, and who approved it. That inventory immediately shows where the real risks sit, and it usually shows that most uses are low risk and should be made easier, not harder.
Good governance makes safe AI use faster. If it only adds friction, people will route around it.
Only then does a policy make sense: proportionate rules for different risk classes, a clear approval route, and named roles. Frameworks such as ISO/IEC 42001 and the NIST AI Risk Management Framework are useful reference points for this. The aim is not paperwork. It is confidence that AI is being used where it helps, with the right checks where it matters.
LeadershipMeasure open innovation by adoption, not by challenges launched
By Haytham AliRead
Open innovation programs are easy to start and easy to report on. Challenges launched, startups engaged and awards made all make good headlines. The harder question is how many solutions the organization actually adopted, and what they changed.
Having helped design and run a national open innovation program, I have seen what separates programs that produce results from those that produce events. It is rarely the quality of the startups. It is whether the challenge came from a business owner with a real problem, a budget and the authority to adopt a solution.
What strong programs do differently
They define each challenge with its owner, value target and constraints before scouting begins. They evaluate solutions on fit, cost, security and vendor maturity, not just novelty. They design pilots with a measurable result and a decision date. And they plan the path to procurement from the start, because a successful pilot that cannot be bought is a failed pilot.
The output of an innovation program is not a list of winners. It is a set of problems solved.
If your program is generating activity but not adoption, the answer is usually upstream: better challenge definition and clearer ownership. That is where the effort pays back most.
Facing one of these questions?
We are happy to talk it through, whether or not it leads to an engagement.