<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Joel Lim on Medium]]></title>
        <description><![CDATA[Stories by Joel Lim on Medium]]></description>
        <link>https://medium.com/@joel-lim?source=rss-14976f2cf2ef------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*3F3qsBWgHm8be-MnNgfLyg.jpeg</url>
            <title>Stories by Joel Lim on Medium</title>
            <link>https://medium.com/@joel-lim?source=rss-14976f2cf2ef------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 22:09:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@joel-lim/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How to Get More Value Out of Microsoft Copilot in 2026]]></title>
            <link>https://joel-lim.medium.com/how-to-get-more-value-out-of-microsoft-copilot-in-2026-78c37d840da5?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/78c37d840da5</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[tokenomics]]></category>
            <category><![CDATA[microsoft-copilot]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 10:07:04 GMT</pubDate>
            <atom:updated>2026-09-02T11:49:21.556Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XWdnvF8tQCkbjenMj63z6A.png" /></figure><p>More companies are giving Copilot to their staff. 30% of companies have turned it on for at least half their employees, almost double last year’s number. 21% have given it to more than three-quarters of their workforce.</p><p>What are employees using it for?</p><p>Writing, editing, summarising, searching for information on SharePoint or OneDrive. These are what the majority, about half, of Copilot users are using it for, and these are fairly basic tasks. Just 1 in 4 use Copilot for data analysis. And only 1 in 5 are using it to carry out tasks, agentically.</p><p>According to the Gartner report, <em>Microsoft 365 Copilot and Agents: Assessing Impact and Value in 2026, </em>tech leaders believe only a quarter of their employees know how to and regularly use Copilot’s advanced features such as selecting which AI model to use, remembering past conversations and personalising its responses.</p><p>A major reason people aren’t getting more value out of Copilot is that they simply don’t know how to use its more advanced features. And new features are being released faster than employees can learn them.</p><p>For companies that have subscriptions beyond the free <em>Copilot Chat</em>, this is a problem. Potential value derived from using the AI assistant is being left on the table. And companies are paying for services their staff aren’t using.</p><p>The solution is quite simple: prioritise the licence for people who’re actually using it and fix the knowledge gap.</p><p><strong>1. Upgrade Gradually</strong></p><ul><li>According to the report, only about 40% of Copilot licences are actually being used on a daily basis. 60% are unused.</li><li>Employees who aren’t using <em>free</em> probably won’t use a paid version. Encourage the use of the basic tier first, then once someone is using it regularly, upgrade them to a full licence.</li><li>Pair the upgrade with training, e.g. using Copilot in Office apps and building their own agents/assistants.</li><li>Revoke paid versions for those who aren’t using the AI chatbot’s advanced features. For some users, <em>free</em> is good enough to get their jobs done.</li></ul><p><strong>2. Enable Continually</strong></p><ul><li>Training should not be a one-time programme. Build an ongoing enablement programme because the tools keep changing.</li><li>Appoint <em>Copilot Champions</em> in each department to give their colleagues hands-on help on how to make Copilot applicable for their tasks.</li><li>Create regular idea-sharing sessions for employees to share tips, prompts and success stories with each other.</li></ul><p>Proving the payoff Copilot is bringing to your organisation can be difficult. 51% of tech leaders reveal that proving Copilot is worth the money is one of their top three challenges.</p><p>In my analysis, value harvested from Copilot-assisted basic tasks is hard to convert into financial return. This is because time saved and efficiency gains from basic tasks aren’t being channelled into trackable revenue generation or cost savings. Employees are not necessarily using all the time saved to deliver more for customers, for example.</p><p><strong>Teaching staff how to make use of more advanced features, building division-specific (vertical) agents that allow for metrics tracking, and offering licences commensurate with actual usage are tactics companies can implement immediately to get more out of Copilot.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78c37d840da5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When a Human-in-the-loop Might be Worse than No Human At All]]></title>
            <link>https://joel-lim.medium.com/when-a-human-in-the-loop-might-be-worse-than-no-human-at-all-99fcda75355c?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/99fcda75355c</guid>
            <category><![CDATA[human-in-the-loop]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Fri, 07 Aug 2026 11:43:36 GMT</pubDate>
            <atom:updated>2026-08-07T12:04:43.249Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DZcZj4x5_ZXHdNHj7WsL-Q.jpeg" /></figure><p>A study called “When Combinations of Humans and AI are Useful” (Vaccaro, Almaatouq and Malone, 2024) breaks conventional wisdom that a task executed by human+AI collaboration will always deliver better outcomes than the best human or AI can on their own.</p><p>Vaccaro et al. studied 106 experiments and concluded that the best human or AI outperforms AI+humans in tasks involving decision-making from a set of options. In open-ended creation tasks, however, specifically content creation, AI+human workflows produced better outcomes than the best human or AI working alone.</p><p>The paper also revealed, counter to popular thinking, that adding model explainability — describing how an AI arrived at its decision — or providing model confidence scores, did not significantly alter the quality of AI-human team output.</p><p>Why this matters to AI in business, five principles for better solution design:</p><ol><li>The research evidences that human-in-the-loop (HITL) alone cannot deliver better outcomes in AI adoption. HITL is not a silver bullet.</li><li>A poorly-designed AI-human process cannot outperform the best human or AI.</li><li>Good AI-human process (interaction) design must take into account the type of task and relative strength of each actor.</li><li>Avoid over-investing in AI performance analytics as metrics alone do not drive output quality; but do invest in process design, as this is most likely to produce significant outcomes for the enterprise.</li><li>Test POCs using best human working alone, AI alone, and AI+human team workflows to determine if the solution needs to be pivoted to get it right.</li></ol><p>Reference: Vaccaro, M., Almaatouq, A. and Malone, T. (2024) ‘When combinations of humans and AI are useful: A systematic review and meta-analysis’, <em>Nature Human Behaviour</em>, 8, pp. 2293–2303. Available at:<a href="https://doi.org/10.1038/s41562-024-02024-1"> https://doi.org/10.1038/s41562-024-02024-1</a> (Accessed: 3 August 2026).</p><p>#ai #enterpriseai #productdevelopment #humanintheloop #airesearch #aiforbusiness</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99fcda75355c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Plain English Summary of the AI Omnibus that Simplifies Compliance with and Sharpens Definition of…]]></title>
            <link>https://joel-lim.medium.com/plain-english-summary-of-the-ai-omnibus-that-simplifies-compliance-with-and-sharpens-definition-of-06bd6af7b56a?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/06bd6af7b56a</guid>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[eu]]></category>
            <category><![CDATA[regtech]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 10:05:47 GMT</pubDate>
            <atom:updated>2026-07-31T10:05:47.287Z</atom:updated>
            <content:encoded><![CDATA[<h3>Plain English Summary of the AI Omnibus that Simplifies Compliance with and Sharpens Definition of the EU AI Act</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pb8AvPEC0dYuN5lsEh1o2A.jpeg" /></figure><h3>Simplifying Compliance with the EU AI Act.</h3><p>In November 2025, the EU released an AI Omnibus as part of a larger digital omnibus package in an effort to reduce complexity and burdens imposed by the EU AI Act. On <strong>27 July 2026</strong>, amendments to the AI Act came into force. Here is a plain English summary by Claude of <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:L_202601744">Regulation (EU) 2026/1744</a>, the AI Omnibus:</p><h4>Background and rationale (the recitals)</h4><p>The AI Act (Regulation (EU) 2024/1689) came into force in August 2024, with different rules taking effect on a staggered timetable up to August 2027. Since then, the EU has found that delays in preparing technical standards for high-risk AI providers and delays in setting up governance and conformity assessment systems at national level have created a heavier compliance burden than expected. Feedback from stakeholders showed a need for extra measures to make implementation clearer, without lowering protection for health, safety and fundamental rights. This Omnibus is the EU’s response: a set of targeted fixes rather than a rewrite.</p><p>Some of the driving themes running through the recitals (recitals provide additional information, in this case to previous regulations):</p><ul><li><strong>Supporting smaller businesses</strong>: the EU wants consistent treatment for SMEs (small and medium enterprises) and a newly defined category, “SMCs” (small mid-cap enterprises, companies that have just grown out of SME status). SMCs face similar burdens to SMEs, so several of the simplifications for SMEs are extended to them too.</li><li><strong>Narrowing what counts as a “safety component”</strong>: the current definition risked classifying too many AI systems as high-risk. <em>The Omnibus clarifies that only systems whose actual intended purpose is to prevent or mitigate risks to health, safety or property count as safety components.</em> Systems intended purely for user assistance, performance optimisation, convenience or similar are excluded, even if built into a regulated product.</li><li><strong>AI literacy</strong>: providers and deployers will now only need to “take measures to support” AI literacy in their staff, rather than face a stricter, harder-to-define obligation. Governments and the Commission are expected to help by providing training resources and good practice.</li><li><strong>Bias detection using sensitive data</strong>: because bias in AI (leading to discrimination) is a serious public interest concern, the Omnibus extends the legal basis <em>allowing limited processing of special category data (e.g. data on ethnicity, health) specifically for detecting and correcting bias.</em> Previously this was only available to providers of high-risk systems; it is now extended to deployers of high-risk systems and to providers/deployers of other AI systems and models, all subject to the same strict safeguards.</li><li><strong>New prohibited practices (Article 5)</strong>: two new bans are added. First, AI systems that generate or manipulate realistic sexual or intimate images, video or audio of an identifiable person without that person’s clear consent (so-called “<em>nudification</em>” tools) are prohibited where this is either the system’s intended purpose, or a reasonably foreseeable outcome that the provider has failed to guard against. Second, AI systems generating or manipulating child sexual abuse material are banned, except where a “without right” legal defence applies (such as law enforcement using such material for investigations). The recitals go into considerable detail on how these bans are meant to work in practice: they target realistic, identifiable depictions, <em>they do not stop legitimate tools (medical imaging, virtual try-on apps, artistic non-realistic nudes)</em>, and providers are expected to build in reasonable safeguards (filtering, consent verification, abuse detection) rather than face an absolute duty to prevent all misuse.</li><li><strong>Reducing duplication with other product laws</strong>: where AI is embedded in products already regulated by other EU laws (medical devices, machinery, aviation, cybersecurity products), the Omnibus tries to avoid forcing companies to comply twice for the same risk. It allows the Commission to identify, by delegated act, cases where another law’s requirements already deliver an equivalent level of protection, so the AI Act’s own requirements can be scaled back accordingly.</li><li><strong>Machinery is moved to a “sectoral” approach</strong>: AI embedded in machinery will now be governed mainly through the Machinery Regulation (2023/1230) rather than directly through the AI Act’s high-risk regime, with the Commission required to write AI-specific health and safety requirements into the Machinery Regulation by way of delegated acts (applying from 2 August 2028).</li><li><strong>Conformity assessment and notified bodies</strong>: to cut duplication, companies and testing bodies that already go through conformity assessment under another EU product law (e.g. medical devices) will be able to use a single, unified application and assessment process to also be recognised under the AI Act, rather than applying twice.</li><li><strong>A new coding/classification system (Annex XIV)</strong>: a structured list of codes is introduced so that “<em>notified bodies” (organisations authorised to carry out conformity assessments) can be clearly designated </em>for the specific categories and technologies of AI systems they are competent to assess, for example biometric systems, generative AI, or systems using symbolic/expert-system approaches.</li><li><strong>Registration simplified</strong>: the information that must be entered into the EU’s public database for certain <em>lower-risk AI systems</em> is streamlined, cutting paperwork while still requiring providers to keep a <em>documented internal assessment</em> that regulators can request.</li><li><strong>Regulatory sandboxes and real-world testing expanded</strong>: EU countries must have <em>at least one national “AI regulatory sandbox” </em>(a controlled testing environment) up and running by 2 August 2027. The EU’s own AI Office may now also run a Union-level sandbox, with priority access for SMEs and SMCs. <em>Real-world testing (testing outside a sandbox, under safeguards) is extended to cover more categories of high-risk AI systems</em>, including those tied to other EU product laws.</li><li><strong>Quality management system simplification</strong>: the lighter, <em>simplified compliance</em> option previously available only to “microenterprises” is now extended to <em>all SMEs, including start-ups.</em></li><li><strong>Governance and enforcement powers of the AI Office</strong>: this is one of the most substantial parts of the Omnibus. <em>The AI Office (the EU’s central AI regulator) gets clarified and strengthened powers to supervise and enforce the AI Act directly, but only for a defined set of systems: AI systems built on general-purpose AI models by the same provider (or same corporate group), </em>and AI systems that are, or are embedded in, very large online platforms or search engines already regulated under the Digital Services Act. For these, the AI Office can request information, conduct inspections (including with judicial authorisation where needed), accept binding commitments from companies, find non-compliance, and impose fines and periodic penalty payments (capped, for the latter, at 5% of average daily turnover per day). Companies retain a right of defence, access to the file (subject to protecting business secrets), and a right to have decisions reviewed by the Court of Justice of the EU. <em>National market surveillance authorities remain responsible for everything else</em>, and there are cooperation and information-sharing duties between the AI Office and national bodies, as well as with authorities that protect fundamental rights (like equality bodies).</li><li><strong>Fees for the scientific panel</strong>: if a Member State calls on the AI Act’s expert scientific panel for advice, the fee structure is simplified to match what the Commission itself pays in similar cases.</li><li><strong>Timelines pushed back</strong>: this is one of the headline changes. The high-risk AI system rules in Chapter III, Sections 1 to 3 (which cover most of the detailed obligations for high-risk systems, such as risk management, data governance, technical documentation, human oversight) will now apply from <strong><em>2 December 2027</em></strong><em> for systems classified as high-risk under Article 6(2)/Annex III (the list-based high-risk categories, such as recruitment or credit-scoring tools),</em> and from <strong><em>2 August 2028</em></strong><em> for systems classified as high-risk under Article 6(1)/Annex I (AI embedded as a safety component in already-regulated products)</em>. This is a significant delay from the original 2 August 2026 date, justified by the <em>slow progress on technical standards and national enforcement infrastructure.</em></li><li><strong>A four-month transition for existing generative AI content-marking rules</strong>: providers of systems that generate synthetic audio, image, video or text and were already on the market before 2 August 2026 get until <strong><em>2 December 2026 </em></strong><em>to comply with the labelling/marking obligations</em> (Article 50(2)).</li><li><strong>Removing some implementing-act requirements</strong>: the Commission is relieved of the need to formally approve codes of practice on marking of AI-generated content and on GPAI models by implementing act, since these codes carry limited legal weight anyway. Similarly, a planned harmonised <em>template for post-market monitoring plans becomes voluntary guidance rather than a binding template</em>, giving companies more flexibility, though the Commission must still publish guidance (with an optional template) by 2 September 2027.</li><li><strong>Right of entry into force</strong>: given the approaching application dates, the Regulation itself enters into force very quickly, three days after publication in the Official Journal.</li></ul><h4>What the operative Articles actually do</h4><p><strong>Article 1</strong> contains the bulk of the changes, amending the AI Act itself. It:</p><ul><li>Adds SME and SMC definitions, tied to existing EU recommendations on company size.</li><li>Rewrites the “safety component” definition (as above).</li><li>Replaces the AI literacy obligation with a softer duty to “take measures to support” literacy, and gives the EU’s AI Board a role recommending common objectives.</li><li>Inserts new Article 4a, setting out the detailed conditions under which special category personal data may be processed for bias detection and correction (necessity, security safeguards, no data sharing with third parties, deletion once no longer needed, and documented justification).</li><li>Amends Article 5 to add the two new prohibitions described above, with detailed carve-outs for consent, lawful defences, and technical/non-realistic content.</li><li>Clarifies in Article 6 that non-safety-related convenience or efficiency features do not make a component a “safety component”, while confirming that anything whose failure would endanger health and safety still does.</li><li>Makes various technical fixes to Articles 10, 11, 17, 25, 27, 28, 29, 30, 40, 42, 43, 50, 56 to 60, 63, 64, 69, 70, 72, 75 to 77, 95 to 99, 111 and 113, largely implementing the simplifications, timeline changes and enforcement powers described in the recitals above.</li><li>Moves the reference to the Machinery Directive/Regulation from Section A to Section B of Annex I (the sectoral-approach change for machinery), and removes it from the list of laws requiring third-party conformity assessment duplication.</li><li>Adds the new Annex XIV, the classification code system for notified bodies described above.</li></ul><p><strong>Article 2</strong> makes small, consistent amendments to the Aviation Safety Regulation (2018/1139), so that whenever the Commission adopts technical rules for the aviation sector concerning AI safety components, it must take into account the equivalent requirements set out in the AI Act’s Chapter III, Section 2, ensuring consistency between the two regimes without creating duplicate obligations.</p><p><strong>Article 3</strong> amends the Machinery Regulation (2023/1230) to reflect the sectoral approach for AI in machinery: the Commission must, by delegated act, insert AI-specific health and safety requirements into the Machinery Regulation’s Annex III, mirroring the AI Act’s own requirements (applying from 2 August 2028). In the meantime, manufacturers can rely on AI Act harmonised standards to demonstrate conformity with the Machinery Regulation’s own requirements for AI-related risks.</p><p><strong>Article 4</strong> confirms that the Regulation enters into force three days after publication in the Official Journal and applies directly across all Member States without need for separate national implementation.</p><h4>In short</h4><p>The AI Omnibus does not weaken the AI Act’s core protections. Instead, it buys businesses more time (pushing back the main high-risk compliance deadlines by roughly one to two years), narrows some definitions that had been drawn too broadly (safety components, AI literacy), gives smaller companies specific relief (SMEs and SMCs), reduces duplicated paperwork with other EU product laws (medical devices, machinery, aviation, cybersecurity), builds a clearer and more powerful enforcement structure for the AI Office over the largest AI players, and closes two clear safety gaps by explicitly banning “nudification” AI tools and AI-generated child sexual abuse material.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=06bd6af7b56a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Large Numbers to the Rescue]]></title>
            <link>https://joel-lim.medium.com/large-numbers-to-the-rescue-7dcc490a53e4?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/7dcc490a53e4</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 14:40:01 GMT</pubDate>
            <atom:updated>2026-08-07T14:38:29.409Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W3OtN_GuX05nH7kexutqoA.jpeg" /></figure><p>Despite how “smart” we think artificial intelligence is, many complex tasks require the human touch, or human thought.</p><p>For example, can AI (LLMs to be precise) be good enough to do something as simple (to humans) as binary classification really well?</p><p>In one particular use case I worked on, the answer was No.</p><p>We gave an analyst and an LLM the same one data point and we found that the human was better than AI at classifying the quality of an employee’s work as “good” or “poor”.</p><p>Did we chuck out AI? Also No.</p><p>In cases like the above, there is a really helpful mathematical principle we can use: The Law of Large Numbers.</p><p>According to Investopedia, “The law of large numbers, in statistics, states that the results of a test on a sample get closer to the average of the whole population as the sample size gets bigger. That is, it becomes more representative of the population as a whole.”</p><p>So even though the accuracy of AI was not on par with an analyst’s within a few classification cycles, after classifying a sufficiently large volume of data points (remember that AI classification was still not as good as a human’s) the results surprised us.</p><p>The AI’s “good” and “poor” distribution was close enough to the human’s to be useful. What happened was that once the AI’s errors were spread across enough data points, its overall “good” versus “poor” split started to look a lot like the human’s. (Note that the LLM’s judgement was not skewed towards “good” or “poor” by hidden factors — it was given data points together with some guidance, and it made decisions.)</p><blockquote><strong>Numerous inferences can increase the overall reliability of results where AI performance is lower than a human’s, thanks to the law of large numbers.</strong></blockquote><p>// <a href="https://www.investopedia.com/terms/l/lawoflargenumbers.asp">https://www.investopedia.com/terms/l/lawoflargenumbers.asp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7dcc490a53e4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Costs are Hiding in Your Custom-built Enterprise AI Apps?]]></title>
            <link>https://joel-lim.medium.com/are-you-aware-of-the-hidden-costs-in-custom-built-enterprise-ai-apps-0a6a053546f5?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a6a053546f5</guid>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[tokenomics]]></category>
            <category><![CDATA[finops]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 13:01:07 GMT</pubDate>
            <atom:updated>2026-06-26T15:18:23.372Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j71oROnDVR0y8tv_HBagRQ.jpeg" /></figure><p>In traditional software, costs are driven by licences and seats. In AI enterprise software, costs are driven by GenAI model used, number of queries, prompt and response length, and task complexity.</p><p><strong>TOKEN USAGE-BASED PRICING</strong>, not licences or seats, is affecting the speed of GenAI adoption in enterprises. What’s more, costs increase dramatically with Agentic AI workflows that can cost 50x to 1000x more than a simpler chatbot use case*. This emphasises the importance of FinOps/TokenOps in AI development to track usage.</p><blockquote>Crystal ball time: Will an increase in foundation model supplier competition drive token costs down? Will a highly competitive market encourage pricing models such as per-user or tiered? Will new pricing models incentivise optimisation of AI systems with regards to token use?</blockquote><p>The key to price reduction is an increase in data centres. If data centres continue to be built at the current pace**, if they are able to obtain the electricity needed to power their systems (infrastructure resources), if chip manufacturers keep up with growing demand (hardware resources), if enough engineering talent can be sourced (human resources), if compute demand is price sensitive (market behaviour), and if countries welcome and support data centre construction (political accommodation), then my prognosis is that run costs for LLM-dependent apps will drop commensurate with the increase in data centres.</p><p>In a nutshell, I hold that enterprise AI run costs will decrease when two factors gather momentum:</p><p>(1) the build rate of data centres worldwide and</p><p>(2) the addition of new foundation model suppliers (1 and 2 correlate)</p><p>As these factors accelerate, in my opinion, the good news is that token spend now will probably be the highest you will ever pay. But regardless of the rise or fall of token cost, what is critical to do today is to <strong>optimise agentic workloads to minimise token use where possible.</strong></p><p>*Why AI agents consume more tokens: The autonomy, tools and procedures given to agents mean that to fulfil a user request it might invoke &gt;10 GenAI model API calls; a chatbot querying a dataset might only do one. Token-wise, complicated, multistep, multi-tool, looping multi-agent tasks may use up 50 to 1000 times more tokens than a straightforward non-agentic use case. I advise you to speak to your engineers to confirm GenAI invocation and token count per typical interaction, per use case. Tips on how to make your agentic AI workflows more efficient are a topic for another day.</p><p>**(Refer to chart) Data centre construction accounted for 0.15% of US nominal GDP in Q1 2026, which is an 87.5% increase from Q4 2023 (0.08%). Zooming out for a bigger picture, in the first quarter of 2026, AI-related infrastructure accounted for 0.83% of US nominal GDP and was worth more than half of total computing infrastructure (1.49%). The rate of growth has been phenomenal: In Q1 2025 total AI -related infrastructure accounted for 0.3% of nominal GDP; by Q1 2026 it had grown by 0.53 percentage points, or ~177%, largely driven by compute hardware (chips).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lJUWTsbuX6rGHVvjCurb6Q.png" /></figure><p>#GenerativeAI #AgenticAI #LLM #EnterpriseAI #AIStrategy #AICosts #FinOps #TokenOps #UsageBasedPricing #AIInfrastructure #DataCentres #AIAdoption</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0a6a053546f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Investment Opportunity Framework]]></title>
            <link>https://joel-lim.medium.com/the-ai-investment-opportunity-framework-50812b7381e7?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/50812b7381e7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-product-management]]></category>
            <category><![CDATA[generative-ai-strategy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Tue, 26 May 2026 13:13:13 GMT</pubDate>
            <atom:updated>2026-06-15T15:18:56.209Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qzN_nccPTQjrz8l2zc4mWg.jpeg" /></figure><h4>A quantitative approach to evaluating use cases with insurance sector examples</h4><h4>SETTING THE CONTEXT</h4><p>I am an AI strategy and product consultant currently working for a UK-based insurer, where I assist with the setup of an AI Centre of Excellence. My work involves engaging with stakeholders across divisions including Underwriting, IT, HR, Agency, and Customer Service. I evaluate and refine use cases, create high-level product specifications for solution development, and support the design of AI and data-for-AI governance.</p><p>The business is committed to adopting AI, but the central challenge is determining which use cases warrant investment. This has driven the development of a structured assessment framework to bring rigour and consistency to that decision-making process.</p><p>The framework builds on the four categories of AI usage within an enterprise context defined in a “4-level AI framework” proposed by Oxford Professor Kejia Hu (2026). One of the ways I have sought to improve on the existing framework is to surface and document the risks associated with each AI implementation — a particularly critical consideration in a heavily regulated sector like insurance.</p><h4>SURVEYING THE AI LANDSCAPE IN INSURANCE</h4><p>AI is already embedded into the way insurance companies work. A survey focussed on the insurance sector by AI benchmarking firm Evident AI (Haynes <em>et al.</em>, 2025, p. 9) indicates that the most common use cases are in claims (26.4%), customer service (15.1%), knowledge democratisation (13.2%), and underwriting and pricing (13.2%). With reference to a BCG report (Apotheker <em>et al.</em>, 2025, p. 9), 50% of surveyed insurers use AI in fraud detection and claims, while 39% use it in underwriting. These statistics inform the following examples of deployed AI solutions to give us an idea of the capabilities of AI:</p><p><strong>Customer Service &amp; Knowledge Democratisation</strong></p><p>A report by BCG (De Bellefonds <em>et al.,</em> 2024, p. 10) revealed that a life insurer in Asia identified two use cases, one for customer service and a second for sales and marketing, with the former enjoying a reduction of effort of up to 30%, and the latter saving 30–40% time producing marcomm assets. The same report also indicates that the biggest use of Generative AI is in summarisation and the provision of answers to queries.</p><p><strong>Fraud Probability Detection</strong></p><p>A US-based insurer’s rules-based fraud algorithm was replaced by a machine learning algorithm developed by Zensar (2026) to reduce the number of false positives. The Precision Score of the new ML system rose from 12% to 57% when tested and validated as part of blind validation with Risk Analysts, while False Positives fell by 90%.</p><p><strong>Underwriting &amp; Pricing</strong></p><p>According to Evident AI (Haynes<em> et al.</em>, 2025, p. 10), the North American, UK and Ireland insurer Travelers expedited the underwriting and pricing of small business insurance with an AI model that consumed a business’ location and public data (e.g., social media), offered real-time business risk classification guidance to human agents, and cut down the number of questions asked to potential customers from 40 to nine.</p><p><strong>Damage Assessment</strong></p><p>A Department for Science, Innovation &amp; Technology UK report on the AI sector (2025, p. 12) states that computer vision (CV) could be used in claims processing. They gave an example of a machine learning and CV solution by Tractable that analyses images of natural disasters and road accidents, thereby speeding up a manual process.</p><p><strong>No one-size-fits-all</strong></p><p>These reports clear any doubt that AI is capable of executing tasks with acceptable degrees of accuracy or usable output. One might draw the conclusion that any of these use cases could be implemented at scale by any insurance company and deliver similar workflow efficiency or revenue increase.</p><p>I am of the opinion that while it is good to be aware of how insurers are using AI, it is a mistake to assume that we can copy-paste solutions into different contexts and expect the same outcomes. No two businesses are alike. There are deep differences in AI understanding, workflows, leadership alignment, risk tolerance, data readiness, staff know-how, governance policy, IT infrastructure, internal bureaucracy, and so on, that can make or break AI operations.</p><p>We must ensure a good fit between AI adoption strategies and organisational realities. To that end, the conceptual tool I propose is designed to help organisations navigate this new era of AI disruption and opportunity, particularly when it comes to quantifying use case value.</p><h4>AI INVESTMENT OPPORTUNITY (AIO) FRAMEWORK</h4><p>The AIO framework consists of three main parts:</p><ol><li>Use case categorisation</li><li>Return on investment (ROI)</li><li>Risk estimation</li></ol><p>The framework can improve the chances of full deployment of AI applications that deliver measurable results.</p><p>Use cases do not automatically go from proof-of-concept (POC) to production. According to MIT NANDA research with 153 business leaders (Challapally <em>et al.</em>, 2025, pp. 2–3) only 20% of custom- or pre-built AI systems are selected for pilots, and only 5% make it to deployment. That’s one in 20 use cases. A more strategically-sound AI adoption framework can lift businesses out of resource-draining POC/pilot quagmires.</p><h4>UNPACKING THE FRAMEWORK</h4><p>The AIO framework assumes that a round of use case generation has been carried out by stakeholder-proposers — business function or division leads with vested interest in successful AI implementation.</p><p>The framework also assumes that at a minimum, a core team comprising business analysts/strategists and technical AI experts are available to assist with the implementation of methods in the two stages outlined below.</p><h4>AIO FRAMEWORK STAGE ONE: USE CASE CATEGORISATION &amp; ROI PROJECTION</h4><p>My AI Investment Opportunity (AIO) framework adds a fifth factor, “Augmentation”, to the existing four factors within the 4-level AI Framework proposed by Professor Hu (2026). The core team tags each potential use case with one of the following categories:<strong> Automation, Augmentation, Personalisation, Operational Innovation, or Business Innovation</strong>.</p><p>Here are definitions for each category and example use cases for the purposes of my AIO framework:</p><p>1) <strong>Automation</strong> use cases are characterised by AI systems completing tasks with minimal user involvement. For example, after customer interactions over the phone with human customer service agents, call summaries, analyses of customer sentiment and AI-recommended next steps are automatically generated and stored in a relevant CRM database.</p><p>2) <strong>Augmentation</strong> use cases include, for instance, AI agents that assist underwriters with data collection, analysis, pricing and documentation. AI augmentation of tasks enables employees with poor or average skills to do better work, and helps star personnel reach higher levels of performance, according to Ethan Mollick, referencing a BCG study in his book “Co-Intelligence” (2024, p. 156).</p><p>3) In terms of <strong>Personalisation</strong> use cases, these are where customers receive customised interactions or tailored output. For example, chatbot exchanges or product recommendations grounded on customer data.</p><p>4)<strong> Operational Innovation</strong><em> </em>use cases reimagine internal business processes and workflows to reduce cost or increase the speed and quality of services. They have significant payoff potential. Examples: Employing predictive AI for human resource allocation or using generative AI to produce marketing collateral and expedite approval processes.</p><p>5) Lastly, <strong>Business Innovation</strong> use cases are extremely transformative; they leverage AI to materially alter the way a business serves their customers, i.e. their business model. They have the potential to deliver transformational payoffs and open up substantial new revenue streams. Of course, they also carry significant risks of failure. An example of this kind of use case might be a new service by an insurer that helps customers reduce their risk profile based on years of historical predictive analysis data covering their entire customer base (Agrawal, Gans and Goldfarb, 2022, p. 236–237). For a new life insurance product, the insurer could offer bespoke, risk-lowering lifestyle recommendations to policyholders based on data they have on them. While this may reduce premiums over time, the proposition is market-disrupting enough to attract a significant volume of new customers.</p><p>Now that we have a clear definition of each of the five use case categories, the next thing to do is to calculate roughly how much <em>value</em> a use case will deliver. There are different ways to do this and various metrics we could employ, but for the the purpose of this essay we will estimate the return on investment using the ROI formula below:</p><p>ROI = [ (Estimated Value — Cost of Investment ) / Cost of Investment ] x 100</p><p>For instance, if the estimated cost reduction from deploying Use Case X is £55,000 per year, the total cost to build the app is £30,000 in year one, and it costs £10,000 to maintain it every year:</p><ul><li>Year One ROI = [ (55,000–40,000) / 40000 ] x 100 = 37.5%</li><li>Year Two ROI = [ (110,000–50,000) / 50,000 ] x 100 = 120%</li></ul><p>The core team shall calculate ROI for five years. Additionally, they shall present the estimated P&amp;L, or financial value, the use case will offer the business in year five.</p><p>Why do we need to measure ROI at the onset of the AIO framework in the first place? I hold that we need to be able to establish quantitative financial benefits a use case can deliver, e.g. revenue gains or cost reductions, and not merely qualitative benefits, e.g. the pain point it might overcome. <em>Qualitative must be converted to quantitative</em>; for example, we must attach financial value to happier customers, more tasks taken on by staff, averted risk, etc, in order to justify investment.</p><p>Plainly speaking, establishing estimated ROI up front saves businesses from pursuing AI solutions that bring no or very low return. A 2025 McKinsey report states that eight in ten enterprises adopting new AI technology show minor revenue or cost savings (Sawaya <em>et al.</em>, p. 1). This aligns with the UK’s Department for Science, Innovation and Technology’s research with 3,500 businesses indicating that 77% do not observe topline growth from AI initiatives (Department for Science, Innovation and Technology, 2026, pp. 10 &amp; 43) despite about half of these companies acknowledging a boost in employee output (Department for Science, Innovation and Technology, 2026, p. 44).</p><p>Another important metric to include in this stage is the payback period or breakeven point for each use case. In Use Case X’s example above, breakeven happens by the end of Year One with 37.5% ROI. It is worth noting that it can take<em> “two to four years” for ROI to be achieve</em>d, according to a Deloitte survey (Horton <em>et al</em>., 2025, p. 5). The same survey goes on to reveal that only 13% among top performing use cases see ROI inside of a year (Horton <em>et al.</em>, 2025, p. 5). My perspective is that the more transformative an AI use case is, the longer it might take to realise its value. Estimating how long use cases will take to break even will provide leaders with another important datapoint to consider when deciding which use cases to move ahead with.</p><p>Once the ROI, P&amp;L and payback period have been worked out for each potential use case, we plot each one on a table as per the example in Table 1. To illustrate this, we will make use of use case examples from Stage One.</p><figure><img alt="Use Case Category Use Case Description ROl in 5 years [%] Breakeven Point [Year(s)] P&amp;L for Year 5" src="https://cdn-images-1.medium.com/max/1024/1*Lmi1B6X3DhzyEKzLqW-qNQ.png" /></figure><p>Table 1 Notes:</p><ul><li>“Breakeven Point” is the year when financial value gained from the solution overtakes investment cost inside a five-year rollout window</li><li>“P&amp;L for Year 5” is the estimated return at the end of five years</li><li>If breakeven is not realisable over a five-year period, consider removing the use case from the consideration set</li></ul><h4>AIO FRAMEWORK STAGE TWO: RISK IDENTIFICATION &amp; MITIGATION</h4><p>After ROI and payback duration have been worked out for each potential use case, two more steps remain before we can select use cases to go ahead with. We need to enrich each use case with a prediction and mitigation of key risks associated with each use case.</p><p>There is no lack of studies or literature to help us list typical risks that can arise from enterprise AI transformation programmes. According to a report by the Department for Science, Innovation and Technology (Department for Science, Innovation and Technology, 2026, p. 8), data governance, reliability of AI output and ethics (potential job losses) were cited as main concerns by those surveyed.</p><p>Research by the IBM Institute for Business Value states that “trust, bias, and explainability” (Boinodiris, Handa and Simons, 2025, p. 3) are top factors blocking the progress of AI, according to over half of business leaders surveyed.</p><p>Mo Gawdat offered a pessimistic view in his book “Scary Smart” (p. 114) in 2021 that job loss attributed to AI will be a reality and only those who know how to use AI will retain their roles. Now in 2026, the media is awash with news of an impending “jobpocalypse”. The fear and risk persists.</p><p>Informed by the above, we cluster risk into three broad categories:</p><ul><li><em>Reliability</em> — Technical robustness, infrastructure scalability, AI model accuracy and explainability</li><li><em>Governance</em> — data privacy, AI model bias, regulatory compliance</li><li><em>Societal</em> — ethics, job displacement (job ceases to exist), job replacement (AI takes over the job)</li></ul><p>The next step is to describe the risks for each use case, provide a 1 to 3 risk impact score, as well as actions a business can take to mitigate them. To illustrate this, we will make use of three use case examples described earlier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*87e-KUy_xDt1KUAw_MSf2A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*isJgCZrHhQ-PgawE9BTL4Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6FkGR3dV-AoyoUrla9Qizw.png" /></figure><p>Now that Stage One and Stage Two are complete, all that’s left to do is to tag each potential use case with their Total Risk Impact Score. This will give us a high-level view of each use case’s category, description, ROI, breakeven point, P&amp;L and level of risk. For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5iNKc1oPDccInTeXR3dXew.png" /></figure><p>The final output of the AI Investment Opportunity framework enables leadership assessment of potential use case desirability along financial return and risk factors. The most desirable, generally speaking, would have high returns potential, reach breakeven faster, and low total risk impact scores — relative to other use cases in the consideration set. That said, a use case with a very high risk score should not be immediately taken off the table; it might bring business-transforming returns should it be successful, therefore worth the risk.</p><h4>CONCLUSION</h4><p>The output of the framework is not prescriptive. It helps leaders reflect and make an informed decision on which use cases to pilot, deploy, adopt and exploit — a decision grounded on quantitative data. The AI Investment Opportunity framework can be used as-is or adapted to fit your organisation.</p><p>AI use cases can be better evaluated through clearer categorisation, weighing up their risks, understanding what kind of ROI to expect and how long it typically takes to see returns. The biggest takeaway is that by putting concrete numbers on the value of AI use cases we are able to cut out human bias in deciding which ones to pursue.</p><p>///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////</p><h4><em>REFERENCE LIST</em></h4><p>Agrawal, A., Gans, J., and Goldfarb, A. (2022) <em>Power and Prediction: The Disruptive Economics of Artificial Intelligence</em>. Harvard Business Review Press.</p><p>Apotheker, J., Beauchene, V., De Bellefonds, N., Forth, P., Franke, M., Grebe, M., Kataeva, N., Kirvelä, S., Kleine, D., De Laubier, R., Lukic, V., Luther, A., Martin, M., Walters, J. and Schweizer, C. (2025) <em>The Widening AI Value Gap. </em>Boston Consulting Group. Available at: <a href="https://media-publications.bcg.com/The-Widening-AI-Value-Gap-Sept-2025.pdf">https://media-publications.bcg.com/The-Widening-AI-Value-Gap-Sept-2025.pdf</a> (Accessed: 20 April 2026)</p><p>Boinodiris, P., Handa, R., and Simons, D. (2025) <em>The AI ethics trust engine Converting capital into capabilities. </em>IBM Institute for Business Value. Available at: <a href="https://www.ibm.com/downloads/documents/us-en/1443d5cb56c02cb4">https://www.ibm.com/downloads/documents/us-en/1443d5cb56c02cb4</a> (Accessed: 22 April 2026)</p><p>Challapally, A., Pease, C., Raskar, R. and Chari, P. (2025)<em> The GenAI Divide STATE OF AI IN BUSINESS 2025.</em> MIT NANDA. Available at: <a href="https://www.artificialintelligence-news.com/wp-content/uploads/2025/08/ai_report_2025.pdf">https://www.artificialintelligence-news.com/wp-content/uploads/2025/08/ai_report_2025.pdf</a> (Accessed: 21 April 2026)</p><p>De Bellefonds, N., Charanya, T., Franke, M., Apotheker, J., Forth, P., Grebe, M., Luther, A., De Laubier, R., Lukic, V., Martin, M., Nopp, C. and Sassine, J. (2024) <em>Where’s the Value in AI?</em> Boston Consulting Group. Available at: <a href="https://web-assets.bcg.com/a5/37/be4ddf26420e95aa7107a35aae8d/bcg-wheres-the-value-in-ai.pdf">https://web-assets.bcg.com/a5/37/be4ddf26420e95aa7107a35aae8d/bcg-wheres-the-value-in-ai.pdf</a> (Accessed: 20 April 2026)</p><p>Department for Science, Innovation &amp; Technology (2025) <em>Artificial Intelligence sector study 2024</em>. Department for Science, Innovation &amp; Technology. Available at: <a href="https://www.gov.uk/government/publications/artificial-intelligence-sector-study-2024/artificial-intelligence-sector-study-2024">https://www.gov.uk/government/publications/artificial-intelligence-sector-study-2024/artificial-intelligence-sector-study-2024</a> (Accessed: 20 April 2026)</p><p>Department for Science, Innovation and Technology (2026) <em>AI Adoption Research. </em>Department for Science, Innovation and Technology. Available at: <a href="https://www.gov.uk/government/publications/ai-adoption-research/ai-adoption-research">https://www.gov.uk/government/publications/ai-adoption-research/ai-adoption-research</a> (Accessed: 21 April 2026)</p><p>Gawdat, M. (2021) <em>Scary Smart: The Future of Artificial Intelligence and How You Can Save Our World. </em>Bluebird.</p><p>Haynes, A., Gilbert, C., Fynn, C., Jaén, G., Manji, M and Meeson, S. (2025).<em> Evident AI Index Insurance.</em> Evident AI. Available at: <a href="https://evidentinsights.com/api/payload/file-download/EAI-KFR-Insurance-2506-11.pdf">https://evidentinsights.com/api/payload/file-download/EAI-KFR-Insurance-2506-11.pdf</a> (Accessed: 20 April 2026)</p><p>Horton, R., Michalski, J., Winters, S., Gunn, D. and Holland, J. (2025). <em>AI ROI: The paradox of rising investment and elusive returns.</em> Deloitte. Available at: <a href="https://www.deloitte.com/uk/en/issues/generative-ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html">https://www.deloitte.com/uk/en/issues/generative-ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html</a> (Accessed: 21 April 2026)</p><p>Hu, K. (2026) ‘AI Transformation in Business’ [Lecture]. <em>DipAI 2026 S2 M1: The Landscape of Technological Disruption</em>. University of Oxford. 12 March 2026.</p><p>Mollick, E. (2024) <em>Co-Intelligence: The Definitive, Bestselling Guide to Living and Working with AI.</em> Ebury.</p><p>Sawaya, A., Wang, K.W., Bu, L. and Chang, M. (2025) <em>Beyond the Hype: Unlocking Value from the AI Revolution</em>. McKinsey. Available at: <a href="https://www.mckinsey.com/cn/~/media/mckinsey/locations/asia/greater%20china/our%20insights/beyond%20the%20hype%20%20unlocking%20value%20from%20the%20ai%20revolution/mckinsey-beyond-the-hype-unlocking-value-ai-revolution.pdf">https://www.mckinsey.com/cn/~/media/mckinsey/locations/asia/greater%20china/our%20insights/beyond%20the%20hype%20%20unlocking%20value%20from%20the%20ai%20revolution/mckinsey-beyond-the-hype-unlocking-value-ai-revolution.pdf</a> (Accessed: 21 April 2026)</p><p>Zensar.com (2026). <em>Designed and developed an AI Powered fraud analytics solution to improve operational efficiency and business outcome for a leading US based Insurer. </em>Available at: <a href="https://www.zensar.com/insights/case-study/insurance/data-engineering-and-analytics/designed-and-developed-an-ai-powered-fraud-analytics-solution-to-improve-operational-efficiency-and-business-outcome-for-a-leading-us-based-insurer">https://www.zensar.com/insights/case-study/insurance/data-engineering-and-analytics/designed-and-developed-an-ai-powered-fraud-analytics-solution-to-improve-operational-efficiency-and-business-outcome-for-a-leading-us-based-insurer</a> (Accessed: 20 April 2026)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50812b7381e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Amara’s Law: Are You Moving Too Fast or Too Slow?]]></title>
            <link>https://joel-lim.medium.com/-23e5dd9d0376?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/23e5dd9d0376</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Wed, 11 Feb 2026 13:04:14 GMT</pubDate>
            <atom:updated>2026-02-11T13:05:16.321Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/624/1*bYi_CAMXN8soNb1dnVMavA.jpeg" /></figure><p>Roy Amara, late researcher and co-founder of the Institute for the Future, said: “𝘞𝘦 𝘵𝘦𝘯𝘥 𝘵𝘰 𝘰𝘷𝘦𝘳𝘦𝘴𝘵𝘪𝘮𝘢𝘵𝘦 𝘵𝘩𝘦 𝘦𝘧𝘧𝘦𝘤𝘵 𝘰𝘧 𝘢 𝘵𝘦𝘤𝘩𝘯𝘰𝘭𝘰𝘨𝘺 𝘪𝘯 𝘵𝘩𝘦 𝘴𝘩𝘰𝘳𝘵 𝘳𝘶𝘯 𝘢𝘯𝘥 𝘶𝘯𝘥𝘦𝘳𝘦𝘴𝘵𝘪𝘮𝘢𝘵𝘦 𝘵𝘩𝘦 𝘦𝘧𝘧𝘦𝘤𝘵 𝘪𝘯 𝘵𝘩𝘦 𝘭𝘰𝘯𝘨 𝘳𝘶𝘯.”</p><p>What are some implications of his statement for the world of work, in light of advances in generative artificial intelligence?</p><p>𝟭. 𝗗𝗶𝘀𝗿𝘂𝗽𝘁𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘁 𝗮𝘀 𝗿𝗮𝗽𝗶𝗱 𝗮𝘀 𝘆𝗼𝘂 𝗺𝗶𝗴𝗵𝘁 𝘁𝗵𝗶𝗻𝗸</p><p>There is often an expectation that AI will completely take over, or heavily augment, jobs as soon as a new agentic or LLM-powered application comes online. In reality, change takes longer. People need time to learn and adapt to new tools, and new applications do not integrate perfectly with existing workflows from day one.</p><p>Implication: Avoid panic-driven decisions, such as rushed job cuts or dramatic reorganisations which can lead to unforeseen problems very quickly.</p><p>Real-world example: Despite the success of a prototype, and the 4.2 months it could save each year on a specific, repeatable process, our client decided not to progress the app to production right away but to consult with teams to simplify the process further and prepare them for change before committing to deployment.</p><p>𝟮. 𝗙𝗮𝗶𝗹𝗶𝗻𝗴 𝘁𝗼 𝗽𝗿𝗲𝗽𝗮𝗿𝗲 𝗶𝘀 𝗽𝗿𝗲𝗽𝗮𝗿𝗶𝗻𝗴 𝘁𝗼 𝗳𝗮𝗶𝗹</p><p>AI investment may take a back seat due to data security concerns, unclear ROI, or a lack of internal skills. However, technical and literacy debt builds over time, and organisations may find themselves too slow to respond to major shifts in their sector, or unable to compete with AI-savvy firms.</p><p>Implication: Ensure your organisation is making sufficient investment in AI planning and innovation initiatives. How much? Around 27% of companies commit roughly a quarter of their IT budget to AI, according to EY (December 2025, n=500).</p><p>Real-world example: We helped a bank with Copilot Enterprise licences make full use of it by proposing a learning programme that empowered staff to build their own agents. In addition, we outlined a six-month plan to define and build agents for more complex use cases.</p><p><strong>∴ If you are thinking about AI’s impact on your organisation, slow down the panic and speed up the preparation.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23e5dd9d0376" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It’s 2026, Why Haven’t You Github Copilot-ed Yet?]]></title>
            <link>https://joel-lim.medium.com/its-2026-why-haven-t-you-github-copilot-ed-yet-524deb1edb9b?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/524deb1edb9b</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ide]]></category>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Tue, 10 Feb 2026 12:30:32 GMT</pubDate>
            <atom:updated>2026-02-11T12:52:44.469Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VzUW0QHXwTDWZrTy5fGvSw.png" /></figure><p>AI‑assisted coding is rapidly gaining ground and it’s here to stay. If you haven’t already enabled GitHub Copilot or a similar code assistant for your developers, here are reasons why you should.</p><p><strong>Faster task completion</strong> — Developers complete tasks up to 55% faster<a href="https://resources.github.com/learn/pathways/copilot/essentials/measuring-the-impact-of-github-copilot/">[1]</a> in controlled studies (n = &gt; 2,000 developers), enabling the same team to deliver significantly more every week.</p><p><strong>Improved code quality </strong>— Copilot supports higher-quality output across multiple criteria. For example, it can automatically generate unit tests, helping developers strengthen their test coverage for app resilience.</p><p><strong>Mature, widely adopted IDE plugin</strong> — First launched in VS Code in March 2022<a href="https://en.wikipedia.org/wiki/GitHub_Copilot#:~:text=GitHub%20announced%20Copilot%27s%20availability%20for%20the%20Visual%20Studio%202022%20IDE%20on%20March%2029%2C%202022">[2]</a>, GitHub Copilot now has an estimated 20 million cumulative users and 1.3 million paid subscribers (as of July 2025)<a href="https://www.quantumrun.com/consulting/github-copilot-statistics/">[3]</a>.</p><p><strong>Low cost, high ROI</strong>— At just $39 / £28.55 per user per month<a href="https://github.com/features/copilot/plans">[4]</a>, it’s a great return on investment: Increase throughput with the existing team, or maintain the same throughput with a smaller team<a href="https://joel-lim.medium.com/how-fixed-or-unlimited-ai-demand-affects-headcount-39f6d5370c51">[5]</a>.</p><p><strong>Polyglot coding support</strong> — Copilot assists across multiple programming languages<a href="https://docs.github.com/en/get-started/learning-about-github/github-language-support#core-languages-supported-by-github-features">[6]</a><a href="https://github.com/orgs/community/discussions/132097">[7] </a>and can help developers work in languages they may not be fluent in (e.g., a team member unfamiliar with C#). It also accelerates both front‑end and back‑end development, with support for JavaScript, Python, Java, Ruby, Go, PHP, C++, C#, and Swift, as well as frameworks like React, Angular, Vue.js, Node.js, Django, Flask, and Ruby on Rails.</p><p>//</p><p><strong>Sources:</strong></p><p>Github Copilot impact on development: <a href="https://resources.github.com/learn/pathways/copilot/essentials/measuring-the-impact-of-github-copilot/">https://resources.github.com/learn/pathways/copilot/essentials/measuring-the-impact-of-github-copilot/</a><br> <br>IDE plugin launch: <a href="https://en.wikipedia.org/wiki/GitHub_Copilot#:~:text=GitHub%20announced%20Copilot%27s%20availability%20for%20the%20Visual%20Studio%202022%20IDE%20on%20March%2029%2C%202022">https://en.wikipedia.org/wiki/GitHub_Copilot#:~:text=GitHub%20announced%20Copilot%27s%20availability%20for%20the%20Visual%20Studio%202022%20IDE%20on%20March%2029%2C%202022</a><br> <br>Pricing: <a href="https://github.com/features/copilot/plans">https://github.com/features/copilot/plans</a><br> <br>Code languages supported: <a href="https://docs.github.com/en/get-started/learning-about-github/github-language-support#core-languages-supported-by-github-features">https://docs.github.com/en/get-started/learning-about-github/github-language-support#core-languages-supported-by-github-features</a> &amp; <a href="https://github.com/orgs/community/discussions/132097">https://github.com/orgs/community/discussions/132097</a><br> <br>Github Copilot statistics: <a href="https://www.quantumrun.com/consulting/github-copilot-statistics/">https://www.quantumrun.com/consulting/github-copilot-statistics/</a></p><p>How AI Demand Affects Headcount: <a href="https://joel-lim.medium.com/how-fixed-or-unlimited-ai-demand-affects-headcount-39f6d5370c51">https://joel-lim.medium.com/how-fixed-or-unlimited-ai-demand-affects-headcount-39f6d5370c51</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=524deb1edb9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How AI Demand Affects Headcount]]></title>
            <link>https://joel-lim.medium.com/how-fixed-or-unlimited-ai-demand-affects-headcount-39f6d5370c51?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/39f6d5370c51</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Mon, 26 Jan 2026 15:00:28 GMT</pubDate>
            <atom:updated>2026-01-26T20:58:51.917Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RIaEd5DYEkCEaTPJVCqLHA.jpeg" /></figure><h4><strong>Understanding the Link Between AI Consumption Potential and Human Labour</strong></h4><p>Picture a resource, a system that uses that resource, and a good produced by the system. <a href="https://en.wikipedia.org/wiki/Jevons_paradox">Jevon’s Paradox</a> shows us that, in some cases, system efficiency improvements will not result in stable or reduced resource usage; it might, in fact, lead to higher resource consumption.</p><p>This is because the more efficient a system becomes at extracting more out of a resource, the less it needs to produce a unit of output, thereby lowering production cost. This is a welcome outcome in contexts where goods and services are price-sensitive, or <em>price</em> <em>elastic:</em> a small price change has a big effect on demand. Examples of elastic goods include holiday travel, non-essential apparel, digital storage and compute.</p><p><em>Side note: The more efficient a resource-ingesting system becomes, the greater the demand for all elements in the value chain — infrastructure, raw materials, logistics, manufacturing, marketing, distribution, etc. Internal combustion engines are a useful example. As they delivered more power from the same amount of fuel, demand for fuel or cars did not plateau. Quite the opposite. We built more factories, invested in warehousing, sold more vehicles and consumed more fuel.</em></p><p><strong>This leads us to Point 1: The better a system becomes in converting a resource into a price-elastic good, the lower the price of that good, which in turn significantly increases demand.</strong></p><p>Applying Point 1 to AI, I suggest that the better an AI system becomes — better primarily in terms of costing less to use, and secondarily in requiring less human effort and time to get a desired output — the higher the demand for it, <em>provided there are few limits on consumption.</em></p><p>To help explain this point are three example use cases:</p><p>1. Customer Service Queries: If chatbots make answering questions nearly free, customers ask more questions.</p><p>2. Internal Reporting: Instant AI-generated reports mean business users request many more of them.</p><p>3. Document Summaries: With summarisation tools, users request them frequently because the effort barrier drops.</p><p>In each use case, lower cost, less time, less effort, combined with virtually unlimited consumption potential, drive increased AI usage.</p><p>On the other hand, <em>where there are hard limits on consumption</em>, demand will not rise beyond a certain point, even as AI systems improve. For instance:</p><p>1. Regulatory Submissions: AI assists with a fixed number of mandatory reports to be submitted every year.</p><p>2. Claims Settlements: AI speeds up processing of filed claims that only need to be resolved once.</p><p>3. KYC Checks: AI performs checks, but their frequency is set by regulation, not internal capacity.</p><p>In these use cases, AI reduces cost, effort and time, but limited consumption causes limited usage.</p><p><strong>This is summarised in Point 2: In AI use cases where there are almost no limits on consumption potential, demand for AI output will be virtually unlimited as systems require less money, effort and time to use. In contrast, in use cases where there are limits on consumption potential, demand for AI output remains fairly constant even as AI systems advance and costs drop.</strong></p><p>As the title of this thought piece suggests, this essay was written to work out the impact of AI on human labour, a key worldwide concern. According to the World Economic Forum’s <a href="https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf">Future of Jobs Report 2025</a>, “On average, workers can expect that two-fifths (39%) of their existing skill sets will be transformed or become outdated over the 2025–2030 period.” <a href="https://arxiv.org/pdf/2510.25137">MIT’s Project Iceberg</a> reports that AI adoption will encompass a range of jobs, beyond tech roles. It will extend to “cognitive and administrative tasks” found in 11.7% of jobs today.</p><p>I propose that Point 2 will be able to help us predict AI’s effect on labour as businesses ramp up AI adoption to boost productivity, reduce operating costs, make better decisions faster and turn the promise of AI into valuable outcomes. I offer two labour-impact scenarios based on Point 2:</p><p>Scenario 1: In enterprise AI use cases where there are <em>almost</em> <em>no limits on AI consumption potential</em>, the business will be able to deliver more products or services within the same timeframe with the same number of employees. The business may decide to use AI to increase revenue without incurring significantly higher labour costs by choosing to retain all staff and training them to use AI tools effectively.</p><p>Scenario 2: In use cases where <em>there are limits on AI consumption potential</em>, the business will be able to deliver the same volume of products or services within the same time frame with a smaller team. As a result, the business may decide to reduce the workforce and train remaining team members to use AI tools effectively. It’s also possible that the business would opt to retain everyone, enable them to use AI tools and increase their workload. Either path will result in lower labour costs to deliver fixed outcomes.</p><p>Table 1 describes both scenarios in a table and adds more detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/848/1*SjJTGhq9yJnZRZQL9JP1lg.png" /></figure><p><em>(Table 1. Possible headcount outcomes, inferred from AI use case consumption potential and ensuing business decision. Although not indicated in the table, it is assumed that regardless of consumption potential, the probability of headcount reduction is high where a use case is fully automatable.)</em></p><p><strong>Finally, Point 3: AI’s effect on labour is influenced by the level of AI output a use case can consume.</strong></p><p>AI is rewiring work. It’s redefining jobs and how it’s done. I suggest that the sooner business leaders grapple with the implications of AI transformation on their workforce, the clearer their decisions will be on the matter and better prepared they will be to navigate the difficult choices that lie ahead.</p><p>Towards that end, I propose that leaders:</p><ul><li>Recognise that AI consumption potential is a driver of labour outcomes, more so than improvements in AI, which can largely be assumed.</li><li>Use the framework above to anticipate the headcount implications of AI adoption per use case and make informed choices aligned with strategic, commercial and human resource objectives.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39f6d5370c51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How AI Automation Impacts Wages and Headcount]]></title>
            <link>https://joel-lim.medium.com/wages-and-headcount-how-ai-automation-impacts-work-78ccf3be23b2?source=rss-14976f2cf2ef------2</link>
            <guid isPermaLink="false">https://medium.com/p/78ccf3be23b2</guid>
            <category><![CDATA[jobs]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[work]]></category>
            <category><![CDATA[talent-development]]></category>
            <dc:creator><![CDATA[Joel Lim]]></dc:creator>
            <pubDate>Thu, 08 Jan 2026 15:20:06 GMT</pubDate>
            <atom:updated>2026-01-08T15:20:41.003Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HAdLUuqnxdXMpYIhu5sSQg.jpeg" /></figure><p>IF a job’s low complexity tasks are automated by AI and high complexity tasks are not,<br>THEN the number of jobs will decrease<br>AND pay will increase,<br>Due to <strong>higher complexity tasks that require human expertise</strong> and the <strong>know-how to benefit from the augmented parts of the role.</strong><br>(Example: bookkeeping, sales)</p><p>IF a job’s high complexity tasks are automated by AI and low complexity tasks are not,<br>THEN the number of jobs will increase <br>AND pay will be reduced,<br>Due to a l<strong>ower barrier to entry</strong> for people to assume these jobs. <br>(Example: Taxi driving, stock management)</p><p>In a nutshell:</p><h4><em>If low-level work is automated, fewer jobs remain but they are higher paid.</em></h4><h4><em>If high-level work is automated, more jobs are created but they are lower paid.</em></h4><p>The question for leaders: As you automate tasks with AI, how will you prepare for the consequent scenarios that will impact your workforce?</p><p>(Informed by MIT Management Research, 2025)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78ccf3be23b2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>