<?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 Kishor Kukreja on Medium]]></title>
        <description><![CDATA[Stories by Kishor Kukreja on Medium]]></description>
        <link>https://medium.com/@kishorkukreja?source=rss-98105f25f905------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Kishor Kukreja on Medium</title>
            <link>https://medium.com/@kishorkukreja?source=rss-98105f25f905------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 21:03:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kishorkukreja/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[A Fertilizer Cargo Can Arrive Too Late to Matter]]></title>
            <link>https://medium.com/@kishorkukreja/a-fertilizer-cargo-can-arrive-too-late-to-matter-0d8d44c374a8?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d8d44c374a8</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 15:48:02 GMT</pubDate>
            <atom:updated>2026-10-18T15:48:02.725Z</atom:updated>
            <content:encoded><![CDATA[<p>A proposed fertilizer passage through Hormuz can become operational and still fail the farm it is meant to serve. The missing test is whether a specified nutrient reaches a specified crop before its latest useful application date.</p><p>A UN official said the proposed passage mechanism remained inactive and estimated that activation could take 7–10 days after a US–Iran agreement. Sailing, unloading, customs clearance and inland delivery would follow. IFAD’s president separately reported that some farmers were planting more cautiously and using less fertilizer after existing stocks initially absorbed the disruption. Neither interview supplies a crop-level application series or a measured yield loss.[2][3]</p><p>That evidence calls for a narrow workflow. It should turn mixed maritime, inventory and agronomy records into one dated approval request. It should not forecast a harvest.</p><h3>The operating transformation</h3><p><strong>The workflow converts a shipment status into a latest-useful-receipt decision for one crop and one nutrient.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LIvMuwnRinuXF_D19LQ4JQ.png" /></figure><p>Start with the <strong>farm decision, not the vessel</strong>. For e<strong>very exposed order, record the crop, fertilizer grade, locally approved application interval, latest useful date, usable inland stock, quantity already committed, backed projected field receipt, status of each date, source document and responsible owner.</strong></p><p>The distinction between nutrient grades is essential. IFPRI’s September review found urea prices easing by late August while phosphate and sulphur remained under pressure. Those historical observations do not establish October stock at a named depot, but they show why a general fertilizer index cannot approve a substitution.[8]</p><h3>How the workflow runs</h3><h4>1. Extract facts without upgrading their status</h4><p>A language model can read carrier notices, bills of lading, port updates, customs releases, supplier emails and depot reports. It extracts the product grade, quantity, location, expected date, last verified time and source passage.</p><p>Every extracted date keeps one of three labels: observed, estimated or conditional. The proposed passage plan stays conditional until an official activation notice exists. A port arrival remains an observed port event, not a field receipt. Conflicting estimates remain separate until a named logistics owner resolves them.</p><h4>2. Calculate the receipt margin deterministically</h4><p>A rules service compares usable inland stock and the backed projected field-receipt date with the agronomist-maintained application interval. It includes the local haulage and handling allowance approved for that route.</p><p>The output is a receipt margin, not a yield forecast. A positive margin says the current evidence supports delivery before the local cutoff. A missing or negative margin opens a review. Software must not invent an ETA to complete the calculation.</p><h4>3. Prepare the decision case</h4><p>When the margin narrows, the system assembles the evidence. Procurement confirms grade-specific stock or an alternative origin. Logistics validates carrier acceptance, insurance, customs status and inland capacity. An agronomist assesses whether another product, rate or application date is safe. Finance approves incremental cost, compliance checks trade restrictions, and the grower approves crop and field work.</p><p>The system may draft questions and show the consequences of each supported option. It cannot place an order, change agronomic advice, reroute cargo or promise a yield.</p><h4>4. Close the evidence loop at the field</h4><p>The workflow records depot receipt, grower delivery and, where access and consent allow, actual application. A later application record can confirm or overturn the earlier risk judgment.</p><p>Farm records require tight access controls. Delivery receipts and field data should stay in approved systems, with only the minimum necessary fields sent to a model. Each recommendation needs a retained source trail and the names of the human approvers.</p><h3>Scenario stress test</h3><h4>1. Base case: passage starts after a relevant application window</h4><p>The workflow increases the evidence threshold for claiming that the order can serve the crop. Procurement must show usable inland stock or a backed alternative receipt. The agronomist decides whether a safe substitute or changed timing exists. If neither is supported, the approval request records that the planned application cannot be promised. No crop effect is assigned without field and weather evidence.[2][3]</p><h4>2. Upside case: agreement and passage arrive early enough</h4><p>An activation notice changes the plan’s status, but it does not close the case. The workflow asks for a fertilizer-specific loaded vessel, destination receipt, inland availability and actual farmer purchase. Human owners can keep the original plan only when the crop-and-nutrient receipt remains inside the approved interval.[2]</p><h4>3. Downside case: no agreement soon</h4><p>The system increases monitoring for exposed crops and separates each nutrient rather than applying a global shortage assumption. IFAD’s reported cautious planting supports targeted review; IFPRI’s evidence of alternative exporters and different nutrient price paths blocks a uniform forecast. Agronomists and growers still approve any change, and weather and fuel evidence remain separate from shipment delay when causes are assessed.[2][3][8]</p><h3>Failure modes worth testing</h3><ul><li>The 7–10-day estimate becomes an active transit time instead of a post-agreement activation estimate.</li><li>A carrier update names a vessel but not its cargo grade.</li><li>A port arrival is presented as depot stock or farm delivery.</li><li>An oil-flow figure or container-vessel incident is treated as fertilizer throughput.</li><li>A lower urea price is used to justify a phosphate substitution.</li><li>Conflicting ETAs are averaged into false precision.</li><li>A crop calendar is stale or copied across regions.</li><li>Private farm records reach an unapproved model.</li><li>Reduced fertilizer use is assigned entirely to shipping despite IFAD also naming weather, fuel and Black Sea grain pressures.[3]</li></ul><p>Each failure has a practical control: preserve status labels, require product-specific documents, keep metric families separate, date the agronomy record and retain human approval.</p><h3>Bottom line</h3><p>The useful output is a dated decision for one crop and one nutrient: keep the plan, seek a verified alternative, or tell the grower that timely application cannot yet be supported.</p><blockquote>The workflow prevents a proposed maritime passage from becoming a fictional field delivery.</blockquote><h3>Sources</h3><p>[2] <a href="https://www.al-monitor.com/originals/2026/10/failure-activate-hormuz-fertilizer-plan-could-trigger-food-crisis-un-official">https://www.al-monitor.com/originals/2026/10/failure-activate-hormuz-fertilizer-plan-could-trigger-food-crisis-un-official</a></p><p>[3] <a href="https://www.arabnews.pk/middle-east/hormuz-shipping-disruption-is-starting-to-hit-harvests-un-agency-says-3004306">https://www.arabnews.pk/middle-east/hormuz-shipping-disruption-is-starting-to-hit-harvests-un-agency-says-3004306</a></p><p>[8] <a href="https://www.ifpri.org/blog/how-are-fertilizer-markets-coping-with-the-continued-closure-of-the-strait-of-hormuz">https://www.ifpri.org/blog/how-are-fertilizer-markets-coping-with-the-continued-closure-of-the-strait-of-hormuz</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d8d44c374a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do Not Turn a Dealer Reserve Into an Oil-Cost Assumption]]></title>
            <link>https://medium.com/@kishorkukreja/do-not-turn-a-dealer-reserve-into-an-oil-cost-assumption-2c23e316f497?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c23e316f497</guid>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[oil]]></category>
            <category><![CDATA[reserve-bank-of-india]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 17:52:03 GMT</pubDate>
            <atom:updated>2026-10-11T17:52:03.569Z</atom:updated>
            <content:encoded><![CDATA[<p>India has changed two financial routes around oil imports. One supplies dollars to three named state oil companies. The other requires banks to hold cash against a defined class of new import hedges. A planning system that merges those rules can turn a bank obligation into a false procurement cost.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aL9CJMfOSxERsWNw8f2K7g.png" /><figcaption>Generated with Codex</figcaption></figure><ol><li><a href="https://rbi.org.in/scripts/BS_PressReleaseDisplay.aspx?prid=63784">RBI’s oil-company facility</a> starts on 12 October for Indian Oil, Hindustan Petroleum and Bharat Petroleum through designated banks.</li><li><a href="https://rbidocs.rbi.org.in/rdocs/notification/PDFs/292NT63009643AFA847A48C9A73F31E26175F.PDF">Circular 26</a> requires an authorised dealer to hold cash equal to 20% of INR-equivalent notional at RBI when a qualifying new current-account import hedge exceeds USD 2 million. The cash stays there until the contract ends.</li></ol><p>The rule publishes no importer fee. No public record yet shows what a buyer will pay for comparable cover.</p><h3>The operating transformation</h3><p>The workflow turns a policy notice and a bank quote into a cited treasury decision record. It keeps three questions separate:</p><ol><li>Is the legal entity eligible for the oil-company dollar facility?</li><li>Does the proposed hedge meet the reserve rule’s contract-date, purpose, currency-direction and notional tests?</li><li>Has a like-for-like dealer quote changed enough to justify a different hedge decision?</li></ol><p><strong>An LLM can extract candidate facts. Deterministic checks apply the thresholds. Treasury approves the commercial decision. Legal or the bank resolves disputed applicability.</strong></p><h3>What the workflow needs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gYFeDG2zuMBx8WiAdQsCPA.png" /><figcaption>Generated with Codex</figcaption></figure><p>Keep the source documents, invoice and quote together. The minimum record is:</p><ul><li>RBI document version, effective date and cited passage;</li><li>importer’s legal entity and underlying invoice;</li><li>payment currency, purpose and due date;</li><li>hedge creation date, currency direction, notional and tenor;</li><li>dealer, quote timestamp, forward points, spread, fees and conditions;</li><li>a comparable prior quote, if one exists;</li><li>unresolved fields, reviewer and approval time.</li></ul><p>Do not reuse the latest bank email across several imports. A quote belongs to one request and one market moment.</p><h3>The workflow</h3><h4>1. Version the rules</h4><p>Store the oil-company facility, circular 26 and <a href="https://rbidocs.rbi.org.in/rdocs/notification/PDFs/291NT8DE5483CA3F84609A0CA56720CD64103.PDF">circular 25</a> as separate rule records. Circular 25 changes cancellation and rebooking and reduces two no-underlying thresholds to USD 5 million. Those limits are not the USD 2 million reserve trigger.</p><p>A person approves each rule version before it reaches the checker.</p><h4>2. Extract only cited fields</h4><p>The model extracts entity, date, purpose, direction, notional and tenor from the source record. Every field must point to a document span. Missing evidence stays missing. The model cannot infer a current-account purpose from confident prose or infer facility use from the announcement.</p><h4>3. Apply deterministic tests</h4><p>The checker compares the legal entity with the three named oil companies. For the reserve rule, it tests whether the contract was undertaken after issuance, whether the user buys foreign currency against INR for a current-account transaction and whether the notional exceeds USD 2 million.</p><p>Exact wording matters. A contract at exactly USD 2 million does not satisfy a threshold written as exceeding USD 2 million. A larger contract may still fall outside the rule if another condition fails.</p><h4>4. Compare quotes on the same basis</h4><p>Match importer, currency pair, notional, contract date, tenor and purpose. Record differences in credit terms, market timing and dealer appetite. A wider spread after the rule takes effect is evidence of a changed quote, not proof that the reserve caused it. Ask the bank to explain a new term and record that answer as the bank’s position.</p><p>If no fair comparison exists, preserve both quotes and mark the price inference inconclusive.</p><h4>5. Route a decision to a person</h4><p>Treasury approves changes to hedge duration or coverage. Legal reviews uncertain application. The CFO approves changes that affect cash limits or customer economics. Procurement receives the signed instruction, not raw model output.</p><p>The workflow must never execute a trade, instruct a dealer, split contracts to avoid a threshold, change a customer price or state a compliance position.</p><h3>Approval boundaries</h3><p>Automation can fetch approved sources, detect changed documents, extract fields, run explicit tests and assemble a comparison. It cannot decide whether two financial contracts are legally equivalent or whether a wider quote was caused by the reserve. It cannot turn an analyst expectation into a budget assumption.</p><p>The useful output may be unresolved: no comparable quote. That prevents a Saturday policy announcement from becoming a Monday landed-cost input.</p><h3>Failure modes</h3><ul><li>The model treats dealer-held cash as a 20% customer charge.</li><li>It marks an importer outside the three named firms as eligible for the dollar facility.</li><li>OCR drops after issuance or changes exceeding to at least.</li><li>The checker mixes the USD 2 million reserve trigger with the USD 5 million documentation thresholds.</li><li>A cancelled contract is rebooked despite circular 25.</li><li>Quotes with different tenors or market timestamps are compared as though they were equivalent.</li><li>A document changes without a new rule version and human review.</li></ul><p>Source spans, explicit unknowns and signed approvals address these failures better than a longer prompt.</p><h3>Scenario stress test</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lq-_BnzCu9NdP_dLGXM2Pg.png" /><figcaption>Generated with Codex</figcaption></figure><h4>Base case: dealer terms change</h4><p>If a dealer widens a spread, shortens available tenor or adds conditions, the workflow needs a new quote plus a like-for-like comparator. Treasury can change coverage only after reviewing the terms and alternative dealers. The cause remains unproved unless the bank attributes the change and the comparison controls for other differences.</p><h4>Upside case: dealers absorb the cash requirement</h4><p>If comparable qualifying quotes barely move, the legal reserve still applies but the commercial-cost hypothesis weakens. The workflow should preserve the stable comparison and leave procurement assumptions unchanged. A human still approves the hedge.</p><h4>Downside case: the dollar facility operates while physical costs remain high</h4><p>Facility use would answer where three named companies obtained settlement dollars. It would not remove charter costs, create distillate or prove a lower delivered oil price. The workflow must keep shipping, product availability and currency cover as separate evidence. Treasury may approve cover while procurement keeps the physical cost assumption unchanged.</p><h3>Bottom line</h3><p>The first test is reconstructability. <strong>A reviewer should be able to see the invoice, quote, RBI passage, extracted fields, rule version, unresolved facts and human approval.</strong></p><blockquote>Any claimed importer cost change still needs an actual comparable quote.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c23e316f497" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When a Fuel Release Moves Futures Before Your Contract]]></title>
            <link>https://medium.com/@kishorkukreja/when-a-fuel-release-moves-futures-before-your-contract-7f9ea4c7ccd5?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/7f9ea4c7ccd5</guid>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[procurement]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 17:36:10 GMT</pubDate>
            <atom:updated>2026-10-04T17:36:10.712Z</atom:updated>
            <content:encoded><![CDATA[<p>The G7 committed to release 100 million barrels of crude oil and petroleum products over four months, with substantial diesel frontloaded within 20 days.</p><p>Diesel futures fell after the announcement. The G7 also agreed to coordinate refinery maintenance and asked the IEA to report on implementation and replenishment within 20 days.</p><p>A fuel buyer now has three different calls to make.</p><ol><li>Should the price forecast change?</li><li>Can the business rely on more product reaching an operating site?</li><li>Should it pass a saving or a delivery promise to a customer?</li></ol><p>Each call needs different evidence. Treating the futures move as the answer to all three is how a temporary intervention turns into an unsupported commitment.</p><p>The agentic workflow I would build is a recurring check on those three decisions and on the two resources the intervention draws down: <strong>emergency stocks and flexibility in the refinery maintenance calendar.</strong></p><h3>Start with the decision, then ask for evidence</h3><ol><li>For a price forecast, start with the fuel contract. Identify its product index, location basis, pricing period and next reset date. If the contract follows a diesel benchmark, a market move may affect the next invoice before any released barrel arrives locally. The formula and reset date tell finance when to revise the assumption. <strong>A Brent price or a national diesel average may have no direct bearing on that contract.</strong></li><li>For availability, ask a different question. <strong>Which product has been allocated, where will it land, and has the supplier confirmed the delivery window?</strong> The G7 announcement does not specify country shares, product split or terminal destinations. A buyer can mark availability as possible, but cannot treat the full 100 million barrels as deliverable diesel.</li><li><strong>For a customer promise, test both sides. </strong>Commercial leadership needs the approved price basis and credible delivery evidence before changing a rate or service commitment. The workflow can show what is missing; it cannot approve the promise.</li></ol><h3>How the check runs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5qA6KKuPdjwQB0WKXEFlrQ.png" /></figure><ol><li>Collect official G7, IEA and ministry updates, supplier notices, contract terms, regional product prices and physical delivery information. Save the publication date and source with each claim.</li><li>Extract the product, volume, location, timing and status. Keep <em>announced</em>, <em>allocated</em>, <em>dispatched</em> and <em>received</em> separate. Leave fields blank when the source does not provide them.</li><li>Compare each new claim with the buyer’s actual contract and demand window. A diesel futures move matters differently to an indexed fleet contract than to a fixed price agreement. A barrel at another terminal may not help either.</li><li>Show the human owner only the changed assumption and its evidence. Procurement handles supplier and contract questions. Finance signs off on forecast changes. Operations checks delivery plans. Commercial leadership approves customer terms.</li><li>Add the later obligations to the same review. A stock release needs a replenishment plan. Any refinery schedule change needs engineering and safety approval and a completion window. Recheck both when the IEA reports or operators disclose a change.</li></ol><p>This is one evidence check, run often enough to catch a changed allocation or contract reset. It does not require a swarm of agents. <strong>An LLM can read notices and highlight differences; a rules layer can match products, dates and contract terms; people make the calls.</strong></p><h3>What must be recorded</h3><p><strong>Every claim needs a source and time, product and unit, geography, evidence status and caveat. </strong>Every exposed purchase needs its contract index, pricing period, reset date and delivery point. Any proposed forecast change needs a named approver and the evidence that would reverse it.</p><p>Keep crude, diesel, gasoline and jet fuel separate. Reuters reported U.S. diesel futures down 3.25% and European diesel futures down about 5.75% after the G7 decision, while Brent settled nearly flat. <strong>Those prices measure different things. None discloses where released product will arrive.</strong></p><h3>Scenario stress test</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GO8cc3YAOxswPopZbXhkkA.png" /></figure><ol><li><strong>Base case: relief reaches buyers and the restoration work is spread out.</strong> Allocations turn into deliveries, contracts reset according to their terms, stocks are replenished gradually and refinery maintenance is completed within approved limits. Finance can update the relevant price forecast when the contract supports it. Procurement and operations still verify supply for each exposed site. Engineering and safety approve any maintenance change.</li><li><strong>Upside case: commercial supply recovers as well.</strong> If Chinese product exports resume and Gulf product flows improve, local supply may become less dependent on emergency stocks. The workflow asks for physical flow and regional price evidence before reducing contingency. Procurement and finance approve that change; a source headline alone does not.</li><li><strong>Downside case: prices move, but product does not reach the exposed site.</strong> A contract may pass through a lower benchmark while a terminal remains tight. Later, government replenishment could overlap with refinery work that still has to be done. Finance keeps a later tightening case, procurement checks the local supplier, and engineering retains control of maintenance. The overlap is a risk to test, not an announced schedule.</li></ol><h3>Where the workflow fails</h3><p>An official commitment gets mislabeled as a shipment. A crude release gets counted as delivered diesel. A lower future changes a fixed price contract in the forecast.<strong> A supplier’s unconfirmed delivery becomes a customer promise. Or a request to coordinate maintenance becomes pressure to delay required work.</strong></p><p>Each error has a direct control: preserve the evidence status, match the product and contract, request delivery confirmation, and send maintenance questions to qualified engineering and safety staff.<strong> If a source does not answer a question, the review should say so plainly.</strong></p><p>The practical next step is to <strong>pick one fuel exposure and answer three questions: when can its price change under the contract, what proves supply at its delivery point, and who must approve the resulting commercial decision.</strong></p><p><strong>Keep replenishment and safe maintenance completion in the next forecast review.</strong></p><p>Sources: <a href="https://www.gov.uk/government/news/g7-leaders-statement-on-global-energy-security-and-market-stability">G7 leaders’ statement</a>; <a href="https://www.reuters.com/business/energy/europe-discussing-plan-diesel-stock-releases-after-us-pressure-source-says-2026-10-02/">Reuters on the diesel response</a>; <a href="https://www.reuters.com/business/energy/oil-rises-slightly-market-weighs-mixed-supply-signals-2026-10-02/">Reuters on the crude settlement</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7f9ea4c7ccd5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Three Days at Origin Can Break a Month-End Inventory Plan]]></title>
            <link>https://medium.com/@kishorkukreja/three-days-at-origin-can-break-a-month-end-inventory-plan-634bb89f8398?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/634bb89f8398</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[shipping]]></category>
            <category><![CDATA[port-congestion]]></category>
            <category><![CDATA[logistics]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 16:01:41 GMT</pubDate>
            <atom:updated>2026-09-27T16:01:41.485Z</atom:updated>
            <content:encoded><![CDATA[<h3>Port dwell needs its own planning clock, and the inventory model must see the risk before a shipment misses its vessel</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BSaDawFK4Zjzv1MagaGfEg.png" /><figcaption>Generated with Codex</figcaption></figure><blockquote><strong>A container can lose three days before its ocean voyage starts.</strong></blockquote><p><strong>The cargo is ready. The booking exists. The planning system still carries the expected departure and arrival dates. Yet the container is sitting at origin while its remaining schedule buffer disappears.</strong></p><p>In FY 2026 week 37, <strong>average vessel waiting time at Shanghai reached 78 hours. Ningbo was close behind at 77 hours. </strong>Both readings rose from the previous week, according to Drewry’s 17 September update.</p><p>Three days can look modest beside a 25-day ocean transit. <strong>Inside a monthly inventory cycle, those three days can move a receipt from September into October. If the delay causes the container to miss a weekly sailing, the scheduling penalty is no longer three days. It may become ten.</strong></p><p><strong>The inventory model needs to recognise that risk before the planned vessel departs without the shipment.</strong></p><h3>The delay begins before the voyage</h3><p>Most transport plans give the ocean leg most of the attention. A booking contains an estimated departure, an estimated arrival and a transit time. Origin handoff and port dwell often sit inside one lead-time assumption or appear only after the carrier changes the ETA.</p><p>That approach fails when congestion changes faster than the planning parameters.</p><blockquote><strong>Shanghai’s average vessel wait rose by 13 hours in one week. Ningbo’s rose by 11 hours. During the same period, Drewry’s Intra-Asia Container Index increased 6% to $1,402 per 40-foot container, its fourth consecutive record. Shanghai to Laem Chabang and Ho Chi Minh City both rose 15%.</strong></blockquote><p><strong>The rate increase is visible because someone must approve the higher freight cost</strong>. The timing loss is easier to miss. A planner can accept the quoted rate while MRP continues to carry an available date that the shipment is unlikely to meet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wv91P8v4Dy4WjN2dhYgd1Q.png" /><figcaption>Generated with Codex</figcaption></figure><p>These signals describe two related pressures, but they answer different questions. <strong>Freight rates show the price of capacity. Port dwell shows how much of the delivery window may disappear before the ocean leg begins.</strong></p><h3>One lead-time field cannot explain the risk</h3><p>A purchase order moves through at least five planning clocks:</p><ol><li>Supplier readiness</li><li>Origin handoff</li><li>Port dwell</li><li>Ocean transit</li><li>Destination handling</li></ol><p>Each clock responds to different events.</p><p><strong>A factory delay changes the ready date. A typhoon or berth queue changes port dwell. A blank sailing changes the departure interval. A route diversion changes transit time. Customs congestion or labour disruption changes destination handling.</strong></p><p><strong>Compressing these clocks into one lead-time field makes the planning model easier to maintain, but harder to diagnose</strong>. When a receipt moves, the team cannot see whether the cause is a supplier promise, a port constraint, a carrier schedule or destination clearance.</p><p>The usual response is blunt:<strong> add a general buffer or expedite the order.</strong></p><p>The planning error is not that dwell changed. The error is that the new dwell assumption reached the inventory model after the relevant decision had already passed.</p><p>There is a second failure mode. <strong>Inventory waiting at the terminal can disappear from the planner’s mental model. </strong>A stale ETA creates a projected shortage,<strong> so the team places another order or books airfreight.</strong> <strong>The original container then moves, leaving the business with two solutions to one requirement.</strong></p><h3>Keep dwell separate, then test the missed sailing</h3><p>The fix does not require a new global control tower. Start with a small date model for exposed lanes.</p><p><strong>Track recent port dwell by origin, terminal, carrier service and week. Use a median dwell time for the base plan and a 90th-percentile dwell time for the risk case. Keep both values separate from ocean transit.</strong></p><p>Then connect the result to the actual vessel calendar. The date calculation is straightforward:</p><pre>Available date =<br>ready date<br>+ origin handoff<br>+ port dwell<br>+ ocean transit<br>+ destination handling</pre><p>That calculation still misses one important event: <strong>the planned vessel may depart without the shipment.</strong></p><p><strong>Add a missed-sailing case using the service interval, often seven days on a weekly service, and compare the revised available date with the inventory requirement.</strong></p><p>Consider a shipment with these planning assumptions:</p><ul><li>Stock required: 16 October</li><li>Cargo ready: 8 September</li><li>Origin handoff: 2 days</li><li>Risk-case port dwell: 4 days</li><li>Ocean transit: 26 days</li><li>Destination handling: 3 days</li></ul><p>The chain uses 35 days. The expected available date is 13 October, leaving three days of buffer.</p><p>If the container misses a weekly departure, the available date moves to 20 October. The order is now four days late.</p><p>That result is specific enough to support a decision. <strong>The team can secure the earlier sailing, split the order, add four days of cover or approve a premium mode.</strong></p><p>The useful escalation rule is equally specific:</p><blockquote><em>Recalculate shortage exposure when port dwell exceeds the order’s remaining date buffer. Escalate the missed-sailing case, not every delayed shipment.</em></blockquote><h3>A cheaper rate can still carry a slower clock</h3><p><strong>The wider ocean market makes this separation more useful.</strong></p><p>On <strong>17 September, Drewry’s World Container Index rose only 1% to $4,500. Shanghai to New York increased 7% to $10,394, while Shanghai to Rotterdam fell 9% to $3,626.</strong></p><ol><li>A procurement dashboard might summarise that as expensive America and cheaper Europe. An inventory planner needs the timing layer as well.</li><li>Shanghai’s average waiting time was 78 hours during the same reporting period, while four Asia-Europe blank sailings had been announced for the following week.</li><li>The lower Europe rate may still be commercially attractive.<strong> It does not establish that the booked departure will protect the required inventory date.</strong></li></ol><p>Price, port dwell and sailing availability belong in separate planning fields because each supports a different decision:</p><ul><li>Price affects the transport budget.</li><li>Dwell affects the probability of making the planned departure.</li><li>Sailing frequency determines the penalty if that departure is missed.</li><li>Inventory cover determines whether the business can absorb the penalty.</li></ul><blockquote>A low freight rate cannot compensate for a missed month-end receipt if the planning model never tested the departure interval.</blockquote><h3>Automation should move evidence, not invent dates</h3><p><strong>A bounded automation can monitor carrier messages, port-wait data and schedule changes. It can match those events to open shipments and identify which ETAs need review.</strong></p><p>The date arithmetic should remain deterministic. Each proposed change should retain:</p><ul><li>the source event;</li><li>the previous date;</li><li>the proposed date;</li><li>the planning assumption that changed;</li><li>the affected inventory requirement;</li><li>the owner who must review the response.</li></ul><p>A planner should approve any expedite, supplier change, inventory reallocation or customer-promise adjustment.</p><p>That approval boundary matters because congestion data does not guarantee an outcome for a specific container. An average wait can identify exposure. It cannot prove that one shipment will miss its vessel.</p><p>When the evidence is incomplete, the workflow should return an unclear outcome and request review. <strong>It should not quietly convert a port average into a shipment-level fact.</strong></p><h3>Change the planning clock before adding stock</h3><p>Inventory buffers can be the right response to unstable dwell. They should follow measured variation, not replace visibility.</p><ol><li>First, separate supplier readiness, origin handoff, port dwell, ocean transit and destination handling.</li><li>Second, update the affected order dates and test the next available sailing.</li><li>Third, recalculate stock coverage, month-end receipts and customer commitments.</li></ol><p>Only then should the team decide whether it needs more inventory, an earlier release, a split shipment or a premium move.</p><p>Shanghai and Ningbo did more than report congestion. Their waiting times exposed a planning field that many inventory models still bury inside a fixed lead time.</p><p>Port dwell needs to reach the inventory model while the shipment can still make its vessel.</p><h3>Sources</h3><ul><li><a href="https://www.drewry.co.uk/supply-chain-advisors/supply-chain-expertise/intra-asia-container-index">Drewry Intra-Asia Container Index, 17 September 2026</a></li><li><a href="https://www.drewry.co.uk/supply-chain-advisors/supply-chain-expertise/world-container-index-assessed-by-drewry">Drewry World Container Index, 17 September 2026</a></li><li><a href="https://www.drewry.co.uk/supply-chain-advisors/supply-chain-expertise/cancelled-sailings-tracker">Drewry Cancelled Sailings Tracker, 18 September 2026</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=634bb89f8398" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Diesel Intervention Can Change the Whole Refinery Slate]]></title>
            <link>https://medium.com/@kishorkukreja/a-diesel-intervention-can-change-the-whole-refinery-slate-4df11bc5e268?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/4df11bc5e268</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[foreign-policy]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 15:00:15 GMT</pubDate>
            <atom:updated>2026-09-27T15:00:15.619Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A diesel export restriction may lower domestic diesel prices first and reduce diesel, gasoline, and jet-fuel output later.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4QyFEpUmDG6IYiKfiwSB2Q.png" /><figcaption>Generated with Codex</figcaption></figure><p>That sequence creates a difficult planning problem.</p><ol><li>Policy status can change within hours. The physical effect depends on weekly stocks, regional storage, refinery runs, product yields, freight, and prices across several fuels.</li><li>A diesel-only monitor will miss the shared constraint.</li></ol><p>The useful AI pattern is a <strong>read-only evidence loop that tests how a product-specific intervention could propagate through the refinery slate</strong>. It collects<strong> source-of-record evidence, keeps unlike metrics separate, updates approved scenarios, and sends a receipt-backed decision request to a named person</strong>. It cannot trade, change a surcharge, commit volume, or alter a schedule.</p><h3>The operating transformation</h3><blockquote>The workflow converts a policy or market signal into a cross-product decision case.</blockquote><p>It starts with a <strong>possible change to the diesel outlet</strong>. It then checks whether <strong>domestic demand and accessible storage can absorb redirected volume, whether refinery throughput has changed, and whether gasoline or jet-fuel exposure has moved</strong>. The output is an approval request tied to one decision, one owner, and the evidence that would invalidate the case.</p><p>This structure<strong> fits refinery coupling because the intervention names one product while the plant operates through a shared crude input. </strong>It would not have fit the previous mineral-supply issue unchanged; <strong>that issue depended on conversion, specification, qualification, and buyer acceptance rather than joint production and finite product outlets.</strong></p><h3>Decisions the loop can support</h3><p>Fuel procurement, logistics, finance, legal, and operations can use the loop to decide whether to:</p><ul><li>retain the baseline fuel assumption;</li><li>activate a temporary diesel-relief case;</li><li>add gasoline or jet-fuel exposure to the risk case;</li><li>review terminal-level indexes and contract locations;</li><li>ask legal to interpret a formal restriction;</li><li>ask treasury to review a hedge;</li><li>reconsider the timing of transport, maintenance, or production.</li></ul><p><strong>The loop does not make those decisions. It shows when approved evidence has crossed an agreed trigger and routes the case to the person with authority.</strong></p><h3>Why an ordinary alert fails</h3><p><strong>A news alert can say “diesel ban” while the White House denies a flat restriction and officials discuss voluntary measures.</strong></p><p>A national inventory series can show scarce diesel without revealing <strong>whether the relevant tanks, pipelines, or terminals can accept redirected barrels.</strong></p><p>A fall in diesel prices can coexist with rising gasoline or jet-fuel exposure if refiners later reduce crude runs.</p><p>The system must preserve these distinctions:</p><ul><li>statement, voluntary request, proposed rule, and effective restriction;</li><li>national inventory and accessible regional storage;</li><li>diesel price and gasoline or jet-fuel price;</li><li>refinery utilization and product output;</li><li>observed data, official forecast, and consultant scenario;</li><li>crude spread, product price, and freight cost.</li></ul><blockquote><strong>Combining them into one risk score would hide the mechanism the workflow is meant to track.</strong></blockquote><h3>The minimum credible workflow</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O-foVvRW9Jo4XkM6hPF2hw.png" /><figcaption>Generated with Codex</figcaption></figure><h4>1. Define the decisions before collecting data</h4><p>Start with the decisions that could change: nominations, contract assumptions, hedges, customer surcharges, service promises, or operating schedules. Name the owner and approval threshold for each one.</p><blockquote>This prevents the collector from becoming a stream of interesting numbers with no accountable action.</blockquote><h4>2. Collect from named sources</h4><p>Use official policy pages for legal status, EIA tables for stocks and refinery data, approved market feeds for product prices and spreads, and contracted terminal or storage data where the company has access. Reuters or another approved wire can supply current reporting and attributed market context.</p><blockquote>Each receipt needs the publisher, URL, publication time, retrieval time, period, unit, geography, and extraction status. A partial page or search snippet cannot upgrade a scenario by itself.</blockquote><h4>3. Store each metric family separately</h4><p>An LLM can extract policy status, units, periods, and caveats into a fixed schema:</p><pre>metric_name<br>value<br>unit<br>geography<br>period<br>source_family<br>observed_or_forecast<br>policy_status<br>extraction_status<br>caveat<br>source_url</pre><p><strong>Deterministic checks should reject stale dates, incompatible units, missing comparison periods, and policy records without an official effective date.</strong> National stocks cannot stand in for regional tank capacity. A consultant model cannot be relabeled as an official forecast.</p><h4>4. Propagate the signal through the slate</h4><p>For each policy state, test the same sequence:</p><ol><li>How much diesel volume could remain in the domestic system?</li><li>Can domestic demand, tanks, pipelines, and terminals absorb it for the relevant period?</li><li>Has refinery throughput or product output changed?</li><li>Have gasoline or jet-fuel indicators crossed an approved threshold?</li><li>Which company decision changes under that evidence?</li></ol><blockquote>The arithmetic belongs in deterministic calculations. The LLM can compare documents, identify contradictions, and draft the decision note.</blockquote><h3>5. Route one evidence-backed request</h3><p>Each breached trigger should produce a compact record:</p><pre>Status: watch / review / act<br>Observed change: EIA distillate stocks moved from X to Y<br>Policy status: no flat ban; voluntary discussions reported<br>Scenario affected: short restraint<br>Storage evidence: unavailable for the relevant region<br>Cross-product check: gasoline stable; jet fuel above threshold<br>Owner: fuel procurement<br>Decision due: before the next nomination window<br>Receipts: official EIA URL, policy URL, approved market source<br>Falsifier: runs remain high and accessible storage remains available</pre><p><strong>The record should be short enough to review and specific enough to challenge.</strong></p><h3>Data requirements</h3><p>The public layer <strong>needs current policy documents, EIA distillate stocks, refinery utilization, gasoline and jet-fuel production, and approved market prices.</strong></p><p>The company layer <strong>may add terminal access, contracted storage, fuel consumption, nominations, supplier terms, local indexes, and customer commitments.</strong></p><p>Company records <strong>should remain confidential and outside public drafts.</strong> When regional storage or local basis data are missing, <strong>the workflow should state the gap. It should not estimate local capacity from a national inventory number.</strong></p><h3>Scenario stress test</h3><p>The scenarios change what evidence the loop requires and which owner receives the case.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GfTkvcao2Nx09IXi-DaiJA.png" /><figcaption>Generated with Codex</figcaption></figure><h4>1. Base case: no flat ban</h4><p>If the White House denial holds, voluntary measures remain limited, and refinery runs stay high, the loop should keep policy status under watch without claiming that storage congestion has begun. Procurement retains its baseline unless local delivered prices or contract conditions cross an approved trigger. Legal confirms any new official language. High runs, rebuilding stocks, and available storage weaken the intervention case.</p><h4>2. Upside case: short restraint with available storage</h4><p>If some exports shift temporarily and accessible tanks absorb the volume, domestic diesel may soften while gasoline and jet-fuel output remains broadly intact. The loop needs regional placement evidence, the restraint’s duration and scope, and a clear end condition. Fuel procurement can review short-term nominations, but people still approve every volume, hedge, and customer-term change. Storage tightening or the end of the restraint closes the case.</p><h4>3. Downside case: full restriction and run cuts</h4><p>If redirected diesel exceeds domestic demand and placement capacity, storage can tighten and crude runs can fall. The workflow then adds gasoline and jet fuel to the decision case, routes refinery and product-output evidence to logistics and finance, and asks legal to verify the source-of-record restriction. Wood Mackenzie’s figures remain a conditional consultant scenario, not a forecast. No automated action follows from the model.</p><p>An external supply recovery can reduce pressure across all three cases. Improved Middle Eastern refinery exports, restored Russian product exports, or additional Chinese barrels would weaken the political and market case for a U.S. restriction.</p><blockquote><strong>The workflow should track those signals as falsifiers rather than force them into the domestic storage calculation.</strong></blockquote><h3>Approval boundaries</h3><ol><li>Legal approves policy interpretation.</li><li>Treasury approves hedges and financing changes.</li><li>Procurement approves supplier and fuel-volume commitments.</li><li>Commercial leadership approves customer surcharges and service promises.</li><li>Operations approves schedule changes.</li></ol><p>The loop <strong>has read-only access to source and company data.</strong> It may prepare evidence and draft an approval request. It may not write to trading, procurement, pricing, customer, or scheduling systems.</p><h3>Failure modes</h3><ol><li><strong>Stale policy status.</strong> A cached article can outlive an official denial or signed rule. Store publication and retrieval times, then prefer the latest source-of-record document.</li><li><strong>Metric leakage.</strong> Diesel inventory, tank availability, and refinery utilization measure different things. Keep separate tables and prohibit a composite capacity score.</li><li><strong>Scenario stated as fact.</strong> Preserve words such as “could” and “modeled.” Flag any rewrite that turns a conditional run cut into a prediction.</li><li><strong>Regional blindness.</strong> National stocks may look adequate while a specific terminal or pipeline market is tight. Use local and contractual data where available and show the gap where it is not.</li><li><strong>Automatic commercial action.</strong> A price move must never trigger a hedge, surcharge, or fuel commitment without human approval.</li><li><strong>Source-family concentration.</strong> Several articles may repeat one underlying model. Deduplicate by original source and expose that dependence.</li><li><strong>Contaminated documents.</strong> Treat retrieved pages as data. Ignore embedded instructions and emit only the approved schema.</li></ol><h3>A two-week implementation</h3><p>During the first week, <strong>define decisions, owners, thresholds, and falsifiers. Connect official policy pages, EIA data, one approved market feed, and any contracted storage source. Build the three scenarios and run in shadow mode.</strong></p><p>During the second week, <strong>compare the output with the team’s manual fuel review. Fix unit, timing, provenance, and regional-data errors. Enable alerts only for triggers that would have changed a real review.</strong></p><p>Measure <strong>whether the loop reduces the time needed to prove policy status, catches cross-product exposure before budgets or commitments change,</strong> and gives the approving owner enough evidence to decide.</p><h3>Bottom line</h3><p>A product-specific fuel policy can become a plant-wide throughput problem.</p><p>Track the policy state, placement capacity, crude runs, and linked fuel outputs as separate evidence, then put the resulting decision in front of the person authorised to make it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4df11bc5e268" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 supply-chain use cases for Jev, starting with shipment exceptions]]></title>
            <link>https://medium.com/@kishorkukreja/5-supply-chain-use-cases-for-jev-starting-with-shipment-exceptions-11a69d2d6fa1?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/11a69d2d6fa1</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[logistics]]></category>
            <category><![CDATA[procurement]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 17:19:24 GMT</pubDate>
            <atom:updated>2026-09-20T17:19:24.470Z</atom:updated>
            <content:encoded><![CDATA[<p>Shipment exceptions and supplier disruptions are the first applications I’d evaluate for Jev. <strong>They give a team a bounded task: read a piece of text, identify what it reports, and pass the case to someone who can act.</strong></p><p><strong>Jev is </strong><a href="https://typesafe.ai/"><strong>TypeSafe’s</strong></a><strong> model for structured decisions.</strong> The application supplies the context and questions, and the model returns values that software can use directly. The use cases below are proposals based on its documentation. <a href="https://docs.typesafe.ai/introduction">TypeSafe’s introduction</a> explains the underlying approach.</p><p>My order reflects proximity to an operational decision and fit with the model’s documented capabilities. <strong>A team with a catalog backlog and well-labeled product data might reasonably start with matching instead.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kQgKXGYr-kVwfxe2dhSYww.png" /><figcaption>Generated with Codex</figcaption></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdatawrapper.dwcdn.net%2Fx13lm%2F1%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fdatawrapper.dwcdn.net%2Fx13lm%2F1%2F&amp;image=https%3A%2F%2Fdatawrapper.dwcdn.net%2Fx13lm%2Fplain-s.png%3Fv%3D1&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="479" frameborder="0" scrolling="no"><a href="https://medium.com/media/8b2f5fbb4dedc1e5d142d247d209a60d/href">https://medium.com/media/8b2f5fbb4dedc1e5d142d247d209a60d/href</a></iframe><h3>1. Shipment exceptions give you a concrete first test</h3><p>Imagine a carrier sends this message: <strong>“The container is still at the terminal. We cannot dispatch it until the shipper sends a corrected packing list.”</strong></p><p>A coordinator reading it would <strong>need to locate the shipment record, identify who owns the packing list, and determine whether the delay affects the customer’s delivery</strong>. Those steps require different information.</p><p>Ask <strong>Jev to classify the reported problem as documentation, capacity, damage, or something else, and separately check whether the carrier explicitly says dispatch is blocked.</strong> Include an unclear outcome for messages that leave the issue unresolved.</p><p>TypeSafe provides three question forms. <a href="https://docs.typesafe.ai/primitives/choice"><strong>Choice</strong></a><strong> selects from named options. </strong><a href="https://docs.typesafe.ai/primitives/noul"><strong>Noul</strong></a><strong> returns the probability that a yes/no statement is true. </strong><a href="https://docs.typesafe.ai/primitives/score"><strong>Score</strong></a><strong> places the input on a descriptive scale</strong>. These can <strong>support a classification, a blocked-dispatch check, and a separate assessment of the interruption described. A severity Score does not tell you how many days late a shipment will be.</strong></p><p>Independent questions can share a call. The application combines the answers afterward and decides which ones are relevant. For the packing-list example, that might mean proposing the documentation queue and attaching the original message. TypeSafe’s <a href="https://docs.typesafe.ai/patterns/intent-routing">routing pattern</a> shows how classifications can select a handler.</p><p><strong>Lateness needs the committed delivery time, the latest operational record, and a calculation. Keep that calculation in code. A routine-sounding message must not erase an overdue flag established from authoritative records.</strong></p><p>There is also a useful baseline here. If existing rules already route a class of messages correctly, adding a model needs a measurable benefit. Include those easy cases alongside ambiguous messages when evaluating Jev.</p><h3>2. Supplier disruptions require a link to what you buy</h3><p>A report that a plant has suspended production may deserve attention. Before escalating it,<strong> a procurement team needs to establish which plant it concerns and whether that plant supplies anything the business depends on.</strong></p><p>Jev <strong>could assess whether the report describes an actual stoppage, a warning about a possible stoppage, or unrelated activity. Give it the relevant passage and a precise question</strong>. TypeSafe lists this kind of text-to-risk-indicator workflow among its <a href="https://docs.typesafe.ai/concepts/use-case-map">example use cases</a></p><p>The <strong>supplier-to-part relationship comes from the team’s records</strong>. So do <strong>open orders, approved alternatives, and exposure. A correct reading of the article can still produce a poor alert if those relationships are wrong.</strong></p><p>I would retain the source passage with every proposed escalation. <strong>A buyer should be able to check what the report says and which supplier record the application connected it to.</strong></p><h3>3. Document checks can focus on one disputed field</h3><p>Suppose an i<strong>nvoice contains a subtotal, tax, a previous balance, and an amount due. An extraction system picks the wrong number.</strong></p><p>TypeSafe <strong>documents a workflow that finds candidate values first, asks Jev to select the relevant one, and lets code copy and normalize that value</strong>.</p><p>A separate cookbook uses Jev to check fields extracted by another model and escalate questionable results. <a href="https://docs.typesafe.ai/cookbooks/pre_parsed_value_extraction_cookbook">Candidate-value selection</a> and <a href="https://docs.typesafe.ai/cookbooks/sde_cascade">extraction verification</a> provide the examples.</p><p>For procurement, I’d test <strong>whether a selected field is supported by the purchase order, acknowledgment, or invoice.</strong> Scanned pages need text e<strong>xtraction first because Jev accepts text input.</strong> <a href="https://docs.typesafe.ai/models">The model documentation</a> describes that restriction.</p><p>Reconciliation remains a calculation. <strong>Multiply quantities and prices in code, check tolerances against policy, and leave payment authorization with the established workflow.</strong></p><h3>4. Catalog matching needs a review outcome</h3><p>Two suppliers may describe similar products differently. <strong>Before comparing their descriptions, shortlist plausible pairs using identifiers or search.</strong></p><p>TypeSafe’s <a href="https://docs.typesafe.ai/cookbooks/entity_alignment">entity-alignment example</a> <strong>evaluates candidate product pairs and allows ambiguous cases to go to a reviewer</strong>. That is a useful pattern to adapt to catalog maintenance.</p><p>Similar descriptions are only part of the decision. <strong>Pack size, material grade, and manufacturer identifiers can change whether items are interchangeable. </strong>An incorrect merge<strong> can attach purchasing history to the wrong product, so I would measure false matches separately from missed duplicates and require approval for merges during a pilot.</strong></p><h3>5. Demand signals belong in a separate forecasting experiment</h3><p>Customer inquiries and sales notes can contain evidence of purchase intent or supply concerns. TypeSafe proposes extracting such signals and adding them to a forecasting model alongside historical data. Its <a href="https://docs.typesafe.ai/concepts/use-case-map">use-case map</a> describes this division of work.</p><p>That creates a testable question: <strong>does the same forecast improve when those text features are included?</strong> Use<strong> unseen outcomes and preserve the timing of the information. A report published after a demand change cannot be used as advance evidence of that change.</strong></p><p>I would put this behind the earlier applications because it needs a downstream model and outcome data before its value can be assessed.</p><h3>What would justify a pilot ?</h3><p>Jev’s <a href="https://docs.typesafe.ai/model-jaggedness/jev-1.13">published limitations</a> support <strong>keeping numerical and date logic outside the model. They also make input selection worth testing: irrelevant material can reduce accuracy.</strong></p><p>Confidence can help decide which cases to review,<strong> but a confident answer can still be wrong</strong>. Thresholds n<strong>eed evaluation on the actual workload.</strong> <strong>TypeSafe’s </strong><a href="https://docs.typesafe.ai/confidence"><strong>confidence guidance</strong></a><strong> makes the review policy an application decision.</strong></p><p><strong>For a shipment pilot, collect past messages with agreed labels, reserve an untouched evaluation set, and compare Jev with current routing</strong>.</p><p>Track missed critical events, wrong queues, and the time reviewers spend correcting suggestions. Keep the model version fixed during the comparison.</p><p><strong>Expand the pilot only if those results justify the additional system and review work.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=11a69d2d6fa1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Announced Mineral Capacity Is Not Yet Supply]]></title>
            <link>https://medium.com/@kishorkukreja/announced-mineral-capacity-is-not-yet-supply-bf1ed2640084?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf1ed2640084</guid>
            <category><![CDATA[critical-minerals]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[ai-workflow]]></category>
            <category><![CDATA[rare-earth-element]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 15:54:05 GMT</pubDate>
            <atom:updated>2026-09-20T15:54:05.741Z</atom:updated>
            <content:encoded><![CDATA[<p>Critical-mineral resilience depends on the conversion system around the material</p><p>A mine, pilot line or policy pledge <strong>cannot protect production until its material works in a named application and reaches a buyer at production scale.</strong></p><p>That gap is where a bounded AI workflow can help.</p><ol><li>Supplier claims sit in presentations.</li><li>Feedstock records sit with sourcing or recycling.</li><li>Test results sit with engineering.</li><li>Qualification status sits with quality.</li><li>Purchase terms and receipts sit elsewhere.</li></ol><blockquote>The workflow connects that evidence without approving a material, supplier or commitment.</blockquote><p>The operating problem, Critical-mineral projects move through distinct states:</p><ol><li>Feedstock is secured.</li><li>Conversion runs repeatedly.</li><li>Output meets a defined specification.</li><li>The material works in the application.</li><li>Quality and the customer accept it.</li><li>Commercial volume is committed and delivered.</li></ol><p><strong>Announced tonnage usually describes an ambition somewhere before the last state</strong>. Treating every announcement as usable cover can delay redesign, qualification, scrap collection and commercial negotiation.</p><p>The workflow <strong>should convert scattered claims into a verified maturity assessment for one material, grade, site and application.</strong></p><p>It should not <strong>collapse gallium, germanium, rare earths and tungsten into one score. Their conversion processes, specifications and qualification paths differ.</strong></p><h3>Data the workflow needs</h3><p>Each assessment needs the original source and extracted passage, plus:</p><ul><li>material, grade, site and application</li><li>feedstock source, contract or collection record</li><li>current output, target output and the date attached to each</li><li>batch, yield and continuous-run history</li><li>certificates of analysis and application test results</li><li>qualification protocol, status, scope and approving owner</li><li>redesign, certification, warranty or customer-approval requirements</li><li>offtake terms, purchase orders and delivery receipts</li><li>export licences, government notices and measured shipment data</li></ul><p>Forecasts must remain forecasts. Company claims must remain labelled company claims. A laboratory purity result cannot substitute for repeatable application performance.</p><h3>A bounded evidence loop</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8p5kmEtMOILHHtAK7x4VxQ.png" /><figcaption>Generated with Codex</figcaption></figure><h4>1. Extract the claim</h4><p>On a recurring schedule, the workflow reads approved supplier, policy, test and delivery records. It extracts the material, grade, site, claimed capacity, current output, feedstock source, qualification stage, buyer and date.</p><p><strong>Low-confidence fields stay blank. Every populated field links to the source passage.</strong></p><h4>2. Test the claim against physical evidence</h4><p>Deterministic checks compare units, dates, grades and sites. The model then drafts contradictions for review:</p><ul><li>a 2030 target with no current continuous-run output</li><li>a purity result with no application test</li><li>a plant nameplate with no secured feedstock</li><li>a policy commitment with no licence, shipment or buyer receipt</li><li>an approved grade from a different site or process</li></ul><p><strong>The system describes the mismatch. It does not decide whether the risk is acceptable.</strong></p><h4>3. Assign the next evidence request</h4><p>The missing gate determines the owner. Sourcing or recycling owns feedstock. Supplier quality owns repeatability. Materials engineering owns performance and redesign. Product quality and the customer owner control qualification. Procurement owns commercial volume. Supply planning verifies repeated delivery.</p><p><strong>The workflow drafts a request with the missing field, supporting evidence and due date. A person accepts or changes the request.</strong></p><h4>4. Record human approval</h4><p>Materials engineering approves technical equivalence. Quality approves release status. Procurement approves supplier and commercial terms. Legal and trade compliance approve regulatory positions. Product and sales owners approve customer-facing changes.</p><p><strong>Each approval records the person, date, evidence, application and scope. An approval for one grade, site or product cannot silently spread to another.</strong></p><h4>5. Produce an application-specific decision</h4><p>The output should answer:</p><ul><li>Which sources are usable for which applications today?</li><li>Which sources are technically credible but lack commercial volume?</li><li>Which applications require redesign, certification or customer approval?</li><li>Which conversion projects need feedstock, testing or qualification investment?</li><li>Which policy or capacity claims lack physical receipts?</li></ul><p>Those answers support qualification funding, supplier development, inventory and customer-risk decisions. <strong>People retain every release, purchase, design and customer commitment.</strong></p><h3>Scenario stress test</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cj4RdYxs63JRj9GxC0tVBA.png" /><figcaption>Generated with Codex</figcaption></figure><h4>a. Base case: uneven project ramps</h4><p>If prices and policy support stay high while projects ramp unevenly, the workflow should keep Chinese exposure visible by grade and application. It should prioritise mature recyclers and qualification programmes with secured feedstock, repeatable output and named buyers. Procurement can propose investment or inventory options, but engineering, quality and commercial owners approve them.</p><h4>b. Upside case: physical flows improve</h4><p>If trade talks improve buyer receipts, public finance reaches conversion assets and selected qualification cycles shorten, evidence requirements do not disappear. The workflow should replace restrictions and forecasts with current licence, shipment, run and acceptance data. Owners can reduce contingency actions only after the improvement appears in the relevant grade and application.</p><h4>c. Downside case: projects miss commercial gates</h4><p>If flow restrictions persist and demand tightens while projects remain before qualification, the workflow should bring redesign, scrap recovery and customer-approval lead times into the decision. It can identify which applications have no accepted alternative and draft escalation options. People decide allocation, redesign, inventory, pricing and customer commitments.</p><p>The falsifier is equally important. If new projects produce qualified material faster than expected, trade flows normalise at buyer level and manufacturers substitute without long redesign or approval cycles, the conversion constraint has weakened. <strong>The workflow should record that evidence instead of defending its original risk rating.</strong></p><h3>Failure modes</h3><p>The workflow fails<strong> if it treats target output as current output, mixes mineral grades, accepts an old certificate after a site or process change, confuses purity with application fitness, or reads a diplomatic statement as physical delivery.</strong></p><p>It also fails when a <strong>polished assessment hides blank fields. Missing feedstock rights, scale data, customer acceptance or delivery history must remain visible</strong>. Supplier-provided evidence needs clear attribution and, where possible, an independent or official check.</p><h3>A 30-day start</h3><blockquote>Start with one material and one application.</blockquote><ol><li>In week one, define the six gates and the owners who can approve each one.</li><li>In week two, load supplier documents, test records and delivery history, then correct the extraction schema.</li><li>In week three, add policy, licence and shipment evidence and test contradiction detection against known cases.</li><li>In week four, run one review with engineering, quality and procurement.</li></ol><p>Measure the time spent finding evidence, the number of unresolved fields and the decisions that reached a named approver. T<strong>he workflow succeeds when the team can distinguish announced, demonstrated, qualified and delivered capacity quickly enough to act before a shortage sets the schedule.</strong></p><h3>Bottom line</h3><p>Critical-mineral resilience depends on the conversion system around the material: <strong>feedstock, process control, application engineering, qualification, buyer acceptance and repeated delivery.</strong></p><p><strong>AI can assemble and challenge the evidence. People decide whether the material, supplier and commercial commitment are acceptable.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf1ed2640084" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Backup Route Reprices the Customer Promise]]></title>
            <link>https://medium.com/@kishorkukreja/when-the-backup-route-reprices-the-customer-promise-8e3e170784db?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e3e170784db</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[logistics]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[freight]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 15:53:49 GMT</pubDate>
            <atom:updated>2026-09-13T15:53:49.708Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A backup route can keep product moving while invalidating the cost and service promise built around it. </strong>That is the operating problem behind redundancy repricing.</p><p>Hormuz was already impaired when Reuters reported attacks affecting Saudi Arabia’s East-West pipeline and Houthi forces reaching Perim island in Bab el-Mandeb.</p><p>The alternate corridors were no longer passive spare capacity. They had their own security, reliability, insurance and fuel consequences.</p><p>The cost signals also disagreed.</p><ol><li><strong>Reuters reported that marine-fuel availability had improved at major bunkering hubs while Singapore VLSFO remained more than 60% above pre-war levels</strong>.</li><li>E<strong>IA put U.S. on-highway diesel at $5.967 per gallon for the week of September 7.</strong></li><li><strong>FreightWaves, citing Freightos, reported falling Asia-U.S. ocean spot rates after frontloading.</strong></li></ol><p>An operator could therefore see a<strong> cheaper ocean quote, working fuel supply and worsening delivered cost at the same time.</strong> A single freight indicator cannot resolve that contradiction.</p><h3>The workflow’s job</h3><p>The workflow <strong>should prepare a repricing decision without making it.</strong></p><p>It collects route and cost evidence, preserves each measure’s definition, traces the change into exposed commitments and asks a named person to approve the response.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nvrc4GXLU7B8FwtFiDXfkg.png" /><figcaption>Generated with Codex</figcaption></figure><blockquote>The workflow turns an alternate-route assumption into a current, component-level delivered-cost case tied to a customer or replenishment promise.</blockquote><h3>Five workflow steps</h3><h4>1. Capture each signal with its measurement rules</h4><p>Ingest route advisories, EIA diesel data, marine-fuel reporting, carrier quotes, insurance changes, internal transport extracts and supplier notices. Store the date, geography, unit, method, scope, caveat and source.</p><p><strong>Crude, VLSFO, diesel, ocean spot rates, drayage, port volume and inventory carry stay separate. The workflow may compare them, but it must not average them into a synthetic freight trend.</strong></p><h4>2. Match the signal to a live exposure</h4><p>Connect each change to the lanes, modes, carriers, facilities, categories and customer commitments that use the affected corridor or fuel assumption. Internal data should include shipment legs, fuel-surcharge clauses, contracted indexes, product weight and cube, refrigeration needs, inventory days and promised service levels.</p><p><strong>The output is a traceable statement such as: this replenishment promise assumes an ocean saving that is smaller than the current diesel and drayage increase. A generic risk score lacks the cost components and commitment needed for action.</strong></p><h4>3. Recalculate the delivered-cost components</h4><p>Recalculate only the components supported by current evidence. Show the old assumption, current observation, source date, affected scope and unresolved gap. If a route event is temporary or the cost pass-through is unknown, say so.</p><p><strong>The calculation should expose which variable changed the decision. A lower linehaul quote may survive the test. It may also be offset by fuel, insurance, additional transit time or inventory finance.</strong></p><h4>4. Draft options for the responsible owner</h4><p>Route the case to the person who controls the next action:</p><ul><li>logistics reviews mode, lane, carrier and service assumptions;</li><li>procurement reviews surcharge and index language;</li><li>finance reviews margin and working-capital effects;</li><li>category teams review price, pack, assortment, formulation or sourcing;</li><li>sales reviews customer language and service caveats.</li></ul><p><strong>The workflow can draft alternatives and state the evidence behind each one. It cannot select the commercial outcome.</strong></p><h4>5. Record the approval and the falsifier</h4><p>A human approves any price change, supplier commitment, mode change, customer promise, hedge, compliance position or public claim. The approved action should include a review trigger, such as the next EIA diesel release or confirmation that East-West pipeline flows normalized.</p><p><strong>Every case also needs a falsifier. For this mechanism, a rapid diesel retreat, normalized route flows and the disappearance of fuel or route risk from operator guidance would weaken the need for continued repricing.</strong></p><h3>Data requirements</h3><p>The minimum credible data set combines external evidence and internal exposure:</p><ul><li>official weekly diesel prices;</li><li>current route and security reporting;</li><li>marine-fuel availability and price observations;</li><li>carrier quotes and advisories, separated by lane and method;</li><li>shipment legs, modes, accessorials and fuel-surcharge terms;</li><li>customer service commitments and replenishment frequency;</li><li>category exposure by weight, cube, perishability and refrigeration;</li><li>inventory days, working-capital assumptions and margin thresholds;</li><li>source dates, caveats and approval history.</li></ul><p>Missing data must remain visible. The system should never infer a contract term, pass-through rate, route closure or product effect because a nearby measure moved.</p><h3>Scenario stress test</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bLpoSUy1AIITF3Iw-9c_xg.png" /><figcaption>Generated with Codex</figcaption></figure><h4>Base case: serious disruption without sustained export loss</h4><p>If route pressure remains live and diesel stays elevated for at least one more EIA print, the workflow keeps route events, marine fuel, diesel and ocean rates separate. Logistics and finance review the exposed promises, but a person decides whether the evidence supports a surcharge, service caveat or inventory change. The next diesel release remains the immediate review trigger.</p><h4>Upside case: routes normalize and diesel retreats</h4><p>If Hormuz diplomacy reduces perceived risk, the East-West pipeline returns cleanly and diesel falls, the workflow should close or downgrade repricing cases whose assumptions no longer hold. It must preserve the earlier evidence without treating W37 as proof of lasting inflation. Humans decide whether to restore prior promises or wait for another confirming observation.</p><h4>Downside case: pressure spreads and adaptation becomes price rationing</h4><p>If diesel stays near or above $6 and marine-fuel adaptation shifts toward price rationing, the workflow expands the review from transport assumptions to surcharges, replenishment promises and SKU or pack decisions. It should identify the affected category and responsible owner, show which costs changed and ask for explicit approval. It must not generate an automatic price, assortment or sourcing action.</p><h3>Failure modes</h3><ul><li>Blending diesel, bunker fuel, ocean rates and port volumes into one direction.</li><li>Treating an analyst estimate as official flow data.</li><li>Applying a national diesel average to a lane without checking the contract and regional exposure.</li><li>Assuming improved availability restored the old cost.</li><li>Treating a temporary route event as a permanent planning base.</li><li>Changing customer language from an external headline without internal exposure data.</li><li>Hiding missing evidence behind a confidence score.</li><li>Letting the system approve a commercial, legal or supplier commitment.</li></ul><h3>The approval boundary</h3><p><strong>AI can collect, classify, compare, calculate from approved inputs and draft options. </strong>People retain authority over customer promises, pricing, sourcing, capacity, modes, suppliers, contracts, hedges, compliance and public statements.</p><p>The useful output is a sourced answer to a narrow question: <strong>does this backup route still support the cost and service promise attached to it?</strong></p><p>The workflow <strong>earns trust when it shows the conflicting measures, the exposed commitment, the responsible owner and the evidence that would reverse the case.</strong></p><h3>Bottom line</h3><p><strong>Do not approve a route change because movement remains possible.</strong></p><p><strong>Reprice the full path, show the evidence behind each component and put the customer or replenishment decision in front of the person who owns it.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e3e170784db" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build the cause claim before AI recommends a Supply chain capacity move]]></title>
            <link>https://medium.com/@kishorkukreja/build-the-cause-claim-before-ai-recommends-a-supply-chain-capacity-move-44745764978e?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/44745764978e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[workflow]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[food]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 16:49:19 GMT</pubDate>
            <atom:updated>2026-09-06T16:49:19.257Z</atom:updated>
            <content:encoded><![CDATA[<p>A factory stops. One source blames a <strong>component shortage; another blames weak model demand.</strong> Both explanations describe the s<strong>ame symptom and demand different decisions.</strong></p><p>That is the right boundary for a supply-chain AI workflow. <strong>The system should assemble and challenge a cause claim before a person moves capacity, cash, inventory, supplier priority, or customer pricing.</strong></p><p>The pattern applies whenever headline pressure hides several possible drivers.<strong> It is especially useful when official releases, market data, suppliers, carriers, and internal operations disagree.</strong></p><h3>The operating transformation</h3><p>The workflow converts each symptom into an approval request with six connected elements:</p><ol><li>the observed symptom and its effective date</li><li>the evidence family that supports it</li><li>the proposed cause</li><li>the operational scope affected</li><li>the action that would follow</li><li>the evidence that would disprove the cause</li></ol><p>The output is useful only when those elements stay connected.</p><ol><li>A price increase supported by a supplier quote is different from one supported by a commodity index.</li><li>A booking decline is different from final port throughput.</li><li>A power request is different from a financed project with grid access.</li></ol><h3>Inputs the workflow needs</h3><p>External evidence may include official releases such as ISM and FAO, current reporting, carrier notices, port data, booking data, commodity prices, and supplier statements.</p><p>Internal evidence should include purchase orders, supplier commits, lead times, inventory by SKU and location, bookings, customer allocations, price terms, plant schedules, and approved capital plans. Every input needs a source, observation date, applicable geography, unit, and caveat.</p><p><strong>The workflow should keep forecasts, observed volumes, prices, and company claims in separate evidence families. Blending them produces confident answers from incompatible measures.</strong></p><h3>Workflow steps</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rQB97H1rTo6Nw_mEkj7ldQ.png" /><figcaption>Generated with Codex</figcaption></figure><h4>1. Capture the symptom</h4><p>Record what changed without naming the cause. Examples include a price rise, slower delivery, lower bookings, a production stop, or a longer equipment lead time.</p><h4>2. Separate the evidence families</h4><p>Classify each observation as official data, forecast, market price, operator report, supplier claim, or internal transaction data. Preserve the original wording and caveats.</p><h4>3. Generate competing cause claims</h4><p>The AI proposes plausible causes supported by the available evidence. For a plant stop, those could include component output, model demand, scheduling, or a combination. It must show what supports each cause and what remains unknown.</p><h4>4. Test the operational scope</h4><p>Match the cause claim to the affected SKU, category, supplier, lane, plant, customer promise, or capital project. If the scope cannot be identified, the workflow stops before recommending action.</p><h4>5. Name the owner and falsifier</h4><p>Assign the person who can test the cause. Define the next observation that would confirm or reject it: a supplier commit, order change, port release, category quote, utility study, or production schedule.</p><h4>6. Prepare the approval request</h4><p>Only then may the system propose a capacity reservation, inventory buy, premium-freight move, hedge, supplier escalation, customer-price change, or capital commitment. The request carries the evidence, caveats, alternatives, and expiry time.</p><h4>7. Record the outcome</h4><p>After the decision, compare the expected result with what happened. Use that result to improve source weighting and cause selection, never to remove the human approval boundary.</p><h3>Stress-test the workflow with three scenarios</h3><p>These scenarios come from the current trigger set and serve as planning tests rather than forecasts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n47SH12lLRGyIMXGBhTL5Q.png" /><figcaption>Generated with Codex</figcaption></figure><h4>Base case: pressure stays mixed</h4><p>Manufacturing, services, food, freight and infrastructure evidence continue to point in different directions. The workflow keeps each evidence family separate and sends cause-specific approval requests to the relevant owner. The aim is to avoid a blanket capacity move when the constraint may sit in one category, lane or project.</p><h4>Upside case: the pressure separates cleanly</h4><p>Port and freight data confirm a controlled slowdown after tariff front-loading, food-category pressure stabilises and services price pressure eases. The workflow checks each improvement against dated operational evidence before proposing lower inventory cover, released freight capacity or revised price assumptions. A person approves any change to commitments.</p><h4>Downside case: several constraints deepen together</h4><p>Sugar and cereal risks rise, fuel disruption holds up freight costs, and long power-equipment lead times crowd out ordinary industrial projects. The workflow must resist collapsing those events into one severity score. Category owners test crop and supplier exposure, logistics and finance test delivered-cost assumptions, and infrastructure procurement tests transformer and grid timing. Each response keeps its own evidence and approval path.</p><h3>Human approval boundaries</h3><p><strong>People remain accountable for any action that changes money, commitments, compliance posture, or customer terms.</strong> The AI can gather evidence, compare explanations, identify missing data, and draft an approval request. It cannot approve supplier commitments, premium freight, inventory builds, hedges, tariff treatment, public claims, or customer pricing.</p><p>Named owners should also reject an approval request when the cause is unresolved, the evidence families conflict, or the proposed action reaches beyond the proven scope.</p><h3>Failure modes</h3><ul><li>The system treats a headline index as the cause instead of a symptom.</li><li>It combines forecasts, bookings, observed throughput, and rates into one freight measure.</li><li>It accepts the loudest supplier or vendor claim without internal confirmation.</li><li>It applies a category signal to every SKU in the budget.</li><li>It hides caveats when summarising evidence.</li><li>It recommends an action before naming what would disprove the cause.</li><li>It lets an old observation survive after its decision window has closed.</li></ul><h3>A practical test</h3><p>Take one current pressure signal and ask the workflow to produce two competing cause claims. Require a different owner, falsifier, and proposed action for each. If both paths produce the same recommendation, the system has probably skipped attribution.</p><h3>Bottom line</h3><blockquote><strong>AI earns a role before the approval meeting by separating symptoms from causes and making weak claims visible. The accountable operator decides whether the evidence supports the proposed move.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44745764978e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 6:12 a.m. supply-chain problem]]></title>
            <link>https://medium.com/@kishorkukreja/the-6-12-a-m-supply-chain-problem-1af7d8a18320?source=rss-98105f25f905------2</link>
            <guid isPermaLink="false">https://medium.com/p/1af7d8a18320</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[skills]]></category>
            <category><![CDATA[procurement]]></category>
            <dc:creator><![CDATA[Kishor Kukreja]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 20:39:15 GMT</pubDate>
            <atom:updated>2026-09-05T20:39:15.623Z</atom:updated>
            <content:encoded><![CDATA[<p>At 6:12 a.m., a supplier delay lands in a planner’s inbox. The signal is clear enough. The work is not. Which customer orders are now exposed? How much inventory remains? Is an alternative supplier viable? Who has the authority to accept higher cost, shift production, or tell a customer?</p><p>That is not primarily an AI-agent problem. It is a decision problem: a repeatable operational moment in which evidence sits across systems, ownership is diffuse, and the next action becomes late because the team has to assemble the case before it can make the call.</p><p>Before asking an agent to do anything, leaders need to name the decision they are trying to improve. Until they can name it, define the person responsible for it, and describe what is at risk, the exception is only a scramble with a fashionable label.</p><p>Most agent conversations skip this diagnosis. They begin with a capability: monitor suppliers, optimise inventory, reroute freight. But the management problem usually comes earlier. The facts that should inform a decision live in different places, their timeliness and credibility are uneven, and nobody has agreed which variance deserves attention first.</p><p>The result is not a lack of automation. It is a weak problem statement. A late purchase order, shortage, or shipment delay gets treated as a message to clear rather than a decision to make.</p><h3>A skill for the work before the workflow</h3><p>To make this diagnosis practical, we built the <a href="https://github.com/kishorkukreja/awesome-supply-chain/blob/main/skills/supply-chain-decision-to-delegation">supply-chain-decision-to-delegation skill</a>. It is designed for the work that happens before a team selects an AI agent, designs a workflow, or funds a pilot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VprKP3ZEYWNKh-oDfmROvw.png" /></figure><p>The skill does not fix the supply-chain problem. It helps a leader or planner identify the decision trapped inside it. Starting with one recent exception, it guides the team through the trigger, decision owner, evidence, time window, stakes, reversibility, and consequences. It separates gathering facts from comparing options and from making a commitment that somebody must own.</p><p>That distinction allows the skill to recommend one of three delegation boundaries: AI drafts and executes inside pre-approved rules; AI drafts and a human approves; or AI drafts while a human owns the decision. It can also conclude that the use case is not ready for AI because ownership, policy, process, or data must be repaired first.</p><p>The output is a scoped problem statement and a structured recommendation — not an automatic technology prescription. Depending on the situation, the skill can produce a decision-to-delegation brief, use-case shortlist, workshop summary, bounded-pilot charter, or not-ready-for-AI report.</p><p>Invoke $supply-chain-decision-to-delegation with one recent operating event and ask: what decision is trapped inside this problem, and what—if anything—should be delegated to AI?</p><p>A useful problem statement starts with questions like these:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdatawrapper.dwcdn.net%2Ff3dsw%2F1%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fdatawrapper.dwcdn.net%2Ff3dsw%2F1%2F&amp;image=https%3A%2F%2Fdatawrapper.dwcdn.net%2Ff3dsw%2Fplain-s.png%3Fv%3D1&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="413" frameborder="0" scrolling="no"><a href="https://medium.com/media/60db99993be2a69583bb154b330acd38/href">https://medium.com/media/60db99993be2a69583bb154b330acd38/href</a></iframe><p>The test changes the question. The same late-supplier alert may contain several kinds of work: collecting facts, comparing options, and accepting consequences. Lumping them together is what makes teams say either “AI can handle it” or “AI cannot be trusted.” Neither is a useful diagnosis.</p><p>That distinction is becoming more important as AI agents move from demonstrations into operational settings. Recent supply-chain and agent-governance commentary converges on a simple idea: make authority explicit, limit access to consequential actions, and keep the activity observable. <a href="https://www.scmr.com/article/how-agentic-ai-changes-supply-chain-operations/automation">Supply Chain Management Review</a> and <a href="https://www.idc.com/resource-center/blog/agentic-ai-is-critical-infrastructure/">IDC</a> make the same broad case from different angles.</p><p>For a planner, that produces three materially different arrangements:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdatawrapper.dwcdn.net%2FsLRQ2%2F1%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fdatawrapper.dwcdn.net%2FsLRQ2%2F1%2F&amp;image=https%3A%2F%2Fdatawrapper.dwcdn.net%2FsLRQ2%2Fplain-s.png%3Fv%3D1&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="434" frameborder="0" scrolling="no"><a href="https://medium.com/media/fcc68e7198697c452390121aa48ac795/href">https://medium.com/media/fcc68e7198697c452390121aa48ac795/href</a></iframe><p>Take the 6:12 a.m. alert again. Imagine that a supplier has moved delivery of a critical component from Thursday to Monday. Before anyone can respond, somebody has to establish the facts: the affected purchase-order lines, the quantity and date variance, on-hand and in-transit inventory by location, the production orders that consume the part, the customer commitments downstream, and whether the new date is confirmed or merely an estimate.</p><p>Then comes a different kind of work: comparing the possible responses. Can production be resequenced? Is stock available in another warehouse or country? Is a substitute part qualified? Can the supplier expedite a partial shipment? What does each option do to cost, service, quality, and the promises already made to customers?</p><p>Finally comes the consequential decision. Paying an expedite premium, reallocating scarce stock from one customer to another, pausing a production line, or changing a delivery promise does more than clear an alert. It changes a commercial or operational commitment. All of this is often labelled “the supplier exception,” but it is a chain of distinct decisions with different evidence, stakes, reversibility, and owners. Until those differences are visible, asking whether AI should manage the exception is a category error.</p><blockquote>Note: <strong>Frame the problem before choosing the agent</strong></blockquote><blockquote>When [trigger] happens, [decision owner] needs to decide [specific decision] within [time window]. They need [evidence] to assess [stake]. A wrong or late decision affects [customer, commercial, or operational consequence] and is [reversible / hard to reverse].</blockquote><p>Only then define what AI may prepare, recommend, or execute, and what remains human-owned.</p><h3>The delegation map</h3><p>Once the decision is framed, the workflow becomes easier to see. The aim is not to give every stage to AI. It is to make the handoff and the authority boundary visible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9XH_dZUW-ILqQKOvgbD9WA.png" /></figure><h3>When the framing becomes the work</h3><p>Doing this properly takes time. A leader has to reconstruct a real event, find the decision hidden inside the noise, trace the evidence, identify who can accept the consequence, and decide what must remain human-owned. That is exactly why teams under pressure are tempted to jump from a broad pain point to an agent demo.</p><p>The work can become tedious, especially when the people framing the problem are also running the operation. So, help us help you.</p><p>Use the <a href="https://github.com/kishorkukreja/awesome-supply-chain/blob/main/skills/supply-chain-decision-to-delegation">supply-chain-decision-to-delegation skill</a> to:</p><ul><li>turn a broad symptom into a specific, owned decision;</li><li>separate fact-finding, option comparison, and consequential commitment;</li><li>make stakes, reversibility, and customer, commercial, or operational consequences explicit;</li><li>test whether the use case is ready for AI;</li><li>choose between bounded execution, human approval, and a human-owned decision;</li><li>produce a decision-to-delegation brief, use-case shortlist, workshop summary, or not-ready report.</li></ul><p>Bring one recent exception, not a polished AI brief. The skill will guide the interview, challenge vague assumptions, expose missing ownership or data, and help identify the smallest useful use case. It may recommend a bounded pilot. It may also conclude that process, governance, or data needs attention first. Both are valuable outcomes.</p><p>Your first AI-agent use case should not be “supplier management” or “inventory optimisation.” It should be a specific decision that arrives too late, requires too much evidence stitching, and has clear limits on what can be delegated.</p><p><strong>So ask your team: what is the biggest supply-chain problem we keep discovering too late? Then ask the harder question: what exact decision is trapped inside it?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1af7d8a18320" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>