<?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 Josphat mwangi on Medium]]></title>
        <description><![CDATA[Stories by Josphat mwangi on Medium]]></description>
        <link>https://medium.com/@josphatwanjiruw?source=rss-87acc5d3d7ad------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*DTCB1DEdRJg2vmFnBhOmsQ.png</url>
            <title>Stories by Josphat mwangi on Medium</title>
            <link>https://medium.com/@josphatwanjiruw?source=rss-87acc5d3d7ad------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 09:48:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@josphatwanjiruw/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[We Guarded the House. The Road Was Never Ours.]]></title>
            <link>https://medium.com/@josphatwanjiruw/we-guarded-the-house-the-road-was-never-ours-ce25aa421688?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/ce25aa421688</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[business-continuity]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 19:54:07 GMT</pubDate>
            <atom:updated>2026-11-13T19:54:07.482Z</atom:updated>
            <content:encoded><![CDATA[<p>Friday, nine o’clock. Your systems are running. Your backups are current. Nothing has been breached.</p><p>And nobody in the building can send an email.</p><p>That was the morning in 74 countries this week. We had guarded the house. The road outside closed.</p><p>What are you depending on that you have never planned to lose?</p><h3>What happened</h3><p>At 19:17 Nairobi time on Thursday 12 November, Microsoft <a href="https://socradar.io/blog/microsoft-outlook-outage/">opened an incident</a> on its email service. Within hours, people in <a href="https://www.aljazeera.com/news/2026/11/13/microsoft-hit-by-global-outages-what-happened-and-why">74 countries</a> were reporting the same blank inbox. More than a day later, service was still being restored region by region.</p><p><a href="https://nairobibusinessmonthly.com/microsoft-outage-hits-kenyan-banks/">Companies across Kenya</a> switched to phone lines. The NHS moved some services to paper, a shipping giant warned of delayed updates, and at least one listed company <a href="https://gvwire.com/2026/11/13/hours-long-microsoft-outage-disrupts-businesses-governments-globally/">postponed its results</a>.</p><p>Microsoft says it has found <a href="https://www.techradar.com/news/live/outlook-outage-nov-2026">no sign of an attack</a>. Its explanation so far is a fault traced back to a network change. The full report is still to come.</p><p>So nothing was broken inside any of those organisations. Their servers were fine. Their backups were fine.</p><p>Still, nobody could work.</p><p>The internet coped the way it always does. It made jokes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/706/1*w4T1rLuiwyG2MSM1gjY9lA.png" /></figure><p>The jokes are good. The question underneath them is serious.</p><h3>The house and the road</h3><p>We have spent years guarding the house. Locks. Alarms. A spare key. A second house, in case the first one burns.</p><p>That is what a disaster recovery plan is. It answers one question: what if we lose our own systems?</p><p>But the house was never the whole of it. Everything we do travels on a road. Email is a road. Sign-in is a road. The service in front of our website is a road. We do not own a single one of them.</p><p>The things we own get a recovery plan. The things we rent get a contract, an uptime promise and a row in a vendor list. The unspoken plan for their failure is to wait.</p><p>Waiting is what most of the world did this week. Organisations still running Microsoft’s software on their own servers were <a href="https://gvwire.com/2026/11/13/hours-long-microsoft-outage-disrupts-businesses-governments-globally/">somewhat spared</a>. That is no argument for going back to server rooms. Cloud providers run these systems better than most of us could. It is an argument for honesty. When we moved the system out of the building, we moved the risk out of our sight.</p><p>Not our mistake. Not our outage. Still our loss.</p><h3>What did it cost, and who paid?</h3><p>These are first-day estimates and they will move.</p><ul><li><strong>Lost productivity:</strong> about US$7 billion, according to Anderson Economic Group, <a href="https://www.tomsguide.com/news/live/microsoft-outlook-down-live-updates-outage-11-12-26">as reported</a>.</li><li><strong>Economic loss:</strong> about US$16 billion, according to <a href="https://www.brecorder.com/news/40448782/microsoft-outage-could-generate-700-million-in-insured-losses-kovrr">Kovrr</a>.</li><li><strong>Insured loss:</strong> about US$700 million, within a range of US$300 million to US$1.4 billion (Kovrr). A second estimate, from <a href="https://gvwire.com/2026/11/13/hours-long-microsoft-outage-disrupts-businesses-governments-globally/">Morningstar DBRS</a>, is US$300 million to US$500 million.</li><li><strong>For comparison:</strong> the 2024 CrowdStrike failure cost the Fortune 500 an estimated US$5.4 billion in direct losses, according to <a href="https://parametrixinsurance.com/in-the-news/crowdstrike-to-cost-fortune-500-5-4-billion-insured-loss-range-of-540-million-to-1-08-billion">Parametrix</a>.</li></ul><p>Read the second line and the third together. Sixteen billion lost. Under one billion insured.</p><p>Why so little? Some policies pay only when there is an attack. Others start counting only after eight or twelve hours. An honest mistake at somebody else’s company falls straight through the gap. The service credit in the contract refunds a slice of the subscription. It does not refund the day.</p><p>Then there is the loss no estimate captures. The approval that waited. The quote that arrived late. The customer who tried a competitor because nobody answered. One small business in Queensland <a href="https://www.abc.net.au/news/2026-11-13/microsoft-outlook-outage-cost-to-businesses/107381570">put a single lost day</a> at A$30,000 or more.</p><p><em>Pause here. What does one day without your main tools cost your organisation? Has anyone ever written the number down?</em></p><h3>Has this happened before?</h3><p>Five times in under two and a half years, a routine change at one provider has stopped work for millions of people who did nothing wrong. None of these was an attack.</p><ul><li><strong>12–13 November 2026, </strong><a href="https://www.techradar.com/news/live/outlook-outage-nov-2026"><strong>Microsoft 365</strong></a><strong>.</strong> Outlook, Teams and Copilot in 74 countries, for more than a day. Preliminary cause: a network change.</li><li><strong>23 February 2026, </strong><a href="https://www.decryptiondigest.com/blog/microsoft-365-mfa-outage-conditional-access-lockout-break-glass"><strong>Microsoft Entra ID</strong></a><strong>.</strong> Sign-ins that needed a second factor timed out when the verification service failed.</li><li><strong>18 November 2025, </strong><a href="https://blog.cloudflare.com/18-november-2025-outage/"><strong>Cloudflare</strong></a><strong>.</strong> Sites and apps behind it returned errors for about three hours, after a database permission change produced an oversized configuration file.</li><li><strong>29 October 2025, </strong><a href="https://virtualizationreview.com/articles/2026/05/26/resilient-identity-foundation.aspx"><strong>Microsoft Entra ID</strong></a><strong>.</strong> Sign-in was disrupted for several hours by a fault in Azure Front Door.</li><li><strong>19 July 2024, </strong><a href="https://parametrixinsurance.com/in-the-news/crowdstrike-to-cost-fortune-500-5-4-billion-insured-loss-range-of-540-million-to-1-08-billion"><strong>CrowdStrike</strong></a><strong>.</strong> A faulty software update crashed Windows machines worldwide.</li></ul><p>A different company each time. The same story each time.</p><p>This is not bad luck. One long-running count finds that outside providers account for <a href="https://www.servnetuk.com/research/data-centre-outage-statistics-2026">about two-thirds</a> of publicly reported outages. Our own systems are getting more reliable. The risk has moved to the ones we rent.</p><p><em>Pause here. Which of your providers could do this to you? When did you last ask them how?</em></p><h3>What if next time it is the login?</h3><p>Email we survived. We picked up the phone.</p><p>Now change one detail. Suppose the service that fails is the one that signs everyone in. That one screen opens the accounts, the customer records, the cloud console, and often the chat tool you would use to organise a response.</p><p>When that door jams, every room behind it locks at once.</p><p>It has already happened in small doses. In February, Microsoft’s verification step timed out and people simply could not get in, because access rules are built to refuse entry when a check cannot run. That is the right security design. It is also a total stop. During the Cloudflare outage, customers could not sign in to the very dashboard where they might have fixed things.</p><p>Providers do build safety nets. Microsoft keeps <a href="https://virtualizationreview.com/articles/2026/05/26/resilient-identity-foundation.aspx">recent sessions alive for up to three days</a> if its main sign-in system fails. A new device, or a new starter on Monday morning, is a different matter.</p><p>When the road to the front door closes, how do you get into your own house?</p><p><em>Pause here. If sign-in failed right now, who in your organisation could still get in? How do you know?</em></p><h3>The road your software arrives on</h3><p>An outage is at least loud. Everyone knows within minutes. This next road makes no noise at all.</p><p>Almost nothing is written from scratch any more. We build on packages. Small pieces of code, maintained by strangers, downloaded on trust. If your systems run on Node or PHP, they travel this road every day.</p><p>Look at the last three weeks alone.</p><p>On Tuesday, an attacker used <a href="https://thehackernews.com/2026/11/4837-hijacked-github-accounts-ran-ci.html">4,837 stolen GitHub accounts</a> to test more than 61,000 code repositories, and ran a secret-stealing script in at least 1,806 of them. In 771, it was wired to fire on an ordinary npm install. At least 50 of the hijacked accounts belong to people who publish packages for Node, PHP and other languages. One PHP package on Packagist was already carrying it.</p><p>A week before that, researchers reported <a href="https://safeguard.sh/resources/blog/ethersphere-bee-js-swarm-desktop-compromise-npm-loaders-without-install-scripts">ten hijacked npm packages</a> whose malware needs no install step at all. It wakes up when the code is simply loaded.</p><p>And Node.js itself shipped a fix on 21 October for a <a href="https://www.konfirmity.com/advisories/nodejs-october-2026-security-releases">critical flaw</a> in a compression library it bundles. A working exploit was demonstrated the next day. Within the week, US authorities had listed the flaw as actively exploited.</p><p>None of that made the evening news. All of it sits underneath software your organisation may be running tonight.</p><h3>Is AI making this worse?</h3><p>It deserves a fair hearing, so here are three precise claims.</p><p>First, AI assistants now install things for us. Researchers documented a developer machine infected simply by <a href="https://unit42.paloaltonetworks.com/monitoring-npm-supply-chain-attacks/">opening a project in an AI coding tool</a>, which loaded a poisoned package without anyone typing an install command.</p><p>Second, AI invents dependencies. In one large study, <a href="https://www.securityweek.com/ai-hallucinations-create-a-new-software-supply-chain-threat/">roughly one in five</a> of the packages that code-writing models suggested did not exist. An attacker only has to register the name and wait.</p><p>Third, the clock is faster. The evidence so far does not show that flaws found by AI are attacked more often than any others. What has changed is the pace. The typical gap between a flaw being published and being used <a href="https://cyberscoop.com/?p=89994">fell from about 120 days to about 80</a> in the first half of this year.</p><p>More code. More strangers on the road. Less time to react.</p><p><em>Pause here. Who decides which outside code runs inside your systems? A person, a policy, or nobody?</em></p><h3>The questions, by who should be asking</h3><p>“Are our backups good?” is no longer the question. Backups protect the house. Each of these protects the road, and each belongs to someone specific.</p><ul><li><strong>On the board, or the owner:</strong> Which three outside services could stop us working, and who is accountable for each?</li><li><strong>The chief executive:</strong> What does our bad-day version of the business look like, and have we ever run it?</li><li><strong>In charge of finance:</strong> What does one day without email or sign-in cost us, and have we ever priced it?</li><li><strong>In charge of technology:</strong> If our sign-in provider failed at nine tomorrow, who could still get in, and when was that last tested?</li><li><strong>Running operations or customer service:</strong> Do our customers already know a second way to reach us, before we need it?</li><li><strong>Leading engineering:</strong> Who approves the outside code we depend on, and do we wait before trusting a brand-new release?</li><li><strong>In risk, audit or procurement:</strong> Do our recovery tests cover the vendor failing, or only us failing?</li><li><strong>A lawyer, in-house or advising:</strong> When a provider’s honest mistake stops the business, what do our contracts give us back, and what do we owe our own customers for the delay?</li><li><strong>An insurer or broker:</strong> If thousands of policyholders lean on the same provider, what is our exposure on the day it fails, and do our wordings say plainly whether an accident is covered?</li></ul><h3>What to do on Monday</h3><p>None of this needs a big budget. A shop with five people can do it in an afternoon. So can a company of five thousand.</p><ol><li><strong>List what you cannot work without.</strong> Email, sign-in, payments, internet, the accounting tool, the code registry. Write down who provides each one, and who <em>they</em> depend on.</li><li><strong>Define your bad-day version.</strong> For each item, decide the least that must still work, and how. One analyst this week called it a <a href="https://www.aljazeera.com/news/2026/11/13/microsoft-hit-by-global-outages-what-happened-and-why">minimum viable service</a>.</li><li><strong>Publish your second channel before you need it.</strong> Announcing a backup phone line by email, during an email outage, does not work.</li><li><strong>Keep an emergency way in.</strong> One or two accounts for critical systems that do not depend on your normal sign-in, stored safely and tested every quarter.</li><li><strong>Treat packages as vendors.</strong> Pin versions. Wait a few days before adopting new releases. Limit what AI assistants may install on their own.</li><li><strong>Rehearse someone else’s outage.</strong> One hour, once a year, one scenario: sign-in is down for eight hours at month end. Note what breaks. Fix the top three.</li><li><strong>Read the fine print.</strong> Know what the contract refunds and what the insurance covers, before the day you need either.</li></ol><h3>The first hour</h3><p>The provider will apologise. It will publish a careful report and fix the fault. It always does.</p><p>Then the road will reopen. The inbox will fill. And the temptation will be to call it a bad week and move on.</p><p>That would waste the cheapest warning we are likely to get.</p><p>This week the road was email, and we walked around it. One day it will be a road with no way around.</p><p>We guarded the house well. Nobody doubts that. What remains is a smaller question, and a harder one.</p><p>Which road are you standing on? Who owns it? And what will you do in the first hour after it closes?</p><p>If you cannot answer in one sentence, you have found your next project.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ce25aa421688" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cost of What We Never Wrote Down]]></title>
            <link>https://medium.com/@josphatwanjiruw/the-cost-of-what-we-never-wrote-down-52dda4a95813?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/52dda4a95813</guid>
            <category><![CDATA[technical-debt]]></category>
            <category><![CDATA[documentation]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[engineering-leadership]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Tue, 15 Sep 2026 12:33:02 GMT</pubDate>
            <atom:updated>2026-09-15T12:33:02.402Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W3-aUXUAgpCoK7Mmwb68GA.png" /><figcaption>The Cost of What We Never Wrote Down</figcaption></figure><p>Every team I have joined, and every team I have watched grow, eventually runs into the same quiet wall. Someone asks why a system is built the way it is, and nobody in the room actually knows. The person who made that call moved on, changed teams, or simply never wrote it down, and what should be a two minute answer turns into two days of archaeology through old chats and half finished tickets.</p><p>I have sat on both sides of that wall enough times now, as the engineer asking the question and as the one who should have left an answer behind, to believe it is one of the more underrated problems in how we build software. Not the flashy kind of problem that gets a conference talk. The quiet, compounding kind that shows up as duplicated work, repeated arguments, and a codebase that slowly stops making sense to the people maintaining it.</p><p>My simplified insider view, after years of building enterprise systems and leading a developer community of a few thousand engineers, is that most of what we call technical debt is not really a technology problem. It is a memory problem. Teams do not lack the skill to build things well. They lack a reliable way of passing on why things were built the way they were, and that gap gets more expensive every time someone new joins or someone experienced leaves.</p><p>Here is the pattern I keep seeing, across very different teams and very different stacks. A hard decision gets made under pressure, in a meeting or a thread that nobody archives properly. It works, everyone moves on, and the reasoning behind it quietly disappears with the moment. Months later, someone new proposes undoing that decision, not out of carelessness, but because there was never any record explaining why it existed in the first place. The team spends hours re-arguing a settled question, or worse, reverses it and reintroduces the exact issue it was meant to prevent.</p><p>At the same time, and often in the same organisations, I have watched two teams build their own version of the same capability without ever realising it. An auth flow here, a notifications service there, each one built in good faith by people who simply did not know the other one existed. Nobody was lazy. Nobody was careless. The communication path that would have surfaced the overlap simply was not there, so each team did the rational thing and built what they needed on their own.</p><p>Here are the lessons I keep coming back to, both from building products and from watching this pattern play out across the wider engineering community I work with:</p><ol><li>Undocumented decisions do not disappear, they resurface as debt. The cheapest time to write down why a hard call was made is the moment it is made, while the reasoning is still fresh and the cost of writing it is a few minutes rather than a full afternoon of reconstruction later.</li><li>A short record beats a long wiki nobody reads. The teams that do this well are not the ones with the most documentation, they are the ones with a habit of capturing just enough, right at the point of decision, so a future engineer can tell the difference between debt that was chosen on purpose and debt that simply crept in.</li><li>Architecture needs to be visible across teams, not just within them. The organisations that avoid duplicated work are the ones that make it easy to check what already exists before building something new, even if that check is as simple as a shared map or a five minute conversation before a new service gets spun up.</li><li>Onboarding friction is a signal, not a rite of passage. When every new engineer hits the same wall asking the same question, that is not a training gap, it is a missing document, and it is worth treating it that way instead of accepting it as normal.</li><li>Never underestimate the value of a community that remembers. Some of the most useful knowledge transfer I have witnessed did not happen through formal documentation at all, it happened through people showing up for each other, sharing context freely, and treating institutional memory as something the whole team is responsible for, not just whoever happens to still be around.</li></ol><p>None of this asks teams to be more disciplined than they already are. It asks them to decide that the cheap moment, writing something down while it is still fresh, is worth more than the expensive one, months later, when someone has to rebuild a decision from memory that was never written down in the first place.</p><p>The teams and communities that get this right are not necessarily the most talented ones. They are the ones that treat knowledge as something worth passing on deliberately, rather than something that is simply assumed to stick around.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52dda4a95813" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Client Is Talking. You Just Can’t Hear Them.]]></title>
            <link>https://medium.com/@josphatwanjiruw/your-client-is-talking-you-just-cant-hear-them-01da654643a5?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/01da654643a5</guid>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[product-analytics]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[saas]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 21:15:38 GMT</pubDate>
            <atom:updated>2026-09-06T21:15:38.550Z</atom:updated>
            <content:encoded><![CDATA[<p>No news isn’t good news. Most of the time it just means you haven’t gone looking yet.</p><p>Product teams treat “we haven’t gotten any complaints” like a green light. I’d compare it to a smoke detector with a dead battery. The room isn’t quiet because nothing’s burning, it’s quiet because nothing’s listening anymore.</p><p>Most unhappy users don’t bother filing a ticket. They just make a mental note and stop opening the app. By the time “no complaints” shows up anywhere in your metrics, you’ve usually already lost them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_kpvLN2Ldt9xKmzNEHU35w.jpeg" /></figure><h3>The math nobody wants to look at</h3><p>There’s a number that gets thrown around a lot in customer experience circles, so much that it’s basically a cliché at this point: for every customer who actually complains, something like 26 others hit the same problem and said nothing. It traces back to old TARP research from the 70s and 80s, and honestly, after being requoted for fifty years, the original methodology is murky enough that I wouldn’t treat it as exact. Take it as directionally true rather than a hard measurement. What every version of the research agrees on, though, is the basic shape of the problem: the complaints you hear are a small, self-selected slice of a much bigger, angrier silence.</p><p>Your dashboard doesn’t show heartbreak. It shows a small dip that gets waved off as seasonality in the next meeting, and it keeps getting waved off right up until it isn’t small anymore.</p><h3>Your frontend is monitored. It’s just not listening for the right thing.</h3><p>Ask any engineering team if the frontend is monitored and they’ll say yes without thinking twice: uptime checks, error rates, a nice green dashboard. All of that answers one question, and one question only. Is the site broken? That’s almost never the question a product manager actually needs answered. What they need to know is whether the user is succeeding, and that’s a completely different question that most tooling was never built to answer.</p><p>Here’s what that gap usually looks like in practice, and if you’ve run a checkout funnel before, this will feel familiar: completions quietly drop a few points. No tickets, no alerts, no complaints. Nothing broke, technically speaking. The page loads fine, the server returns 200 OK, and users just quietly give up halfway through, again and again, until somebody finally sits down and watches session recordings and finds a cart bug that’s been bleeding revenue for weeks.</p><p>Nothing was on fire. Nobody got paged. Money just leaked out the whole time, quietly, while everything looked fine.</p><h3>Backend visibility has the same problem, wearing a suit</h3><p>The backend version of this looks more thorough, so it’s easy to mistake it for actual visibility: latency graphs, error rates, throughput charts. It looks like you can see everything. You usually can’t, not in the way that actually helps someone make a product call.</p><p>A spike in server errors on one endpoint could mean one stubborn power user hammering retry over and over, or it could mean ten thousand new signups silently failing to onboard. The log genuinely can’t tell you which. And usually nobody’s job is to translate that log into what it actually means for the roadmap, so the PM ends up going with whoever complained loudest in the last meeting instead of what actually happened to real users.</p><p>It’s not really a data problem when you get down to it. It’s a translation problem. The data’s there. Nobody ever handed the PM a version of it they could actually use.</p><h3>What listening actually looks like</h3><p>The fix isn’t another dashboard that nobody opens. It’s connecting what happened to why it happened, in one place, so a product manager isn’t filing a support ticket just to understand their own product.</p><p>This is the loop that tools like PostHog are built around. You notice people dropping off at a certain funnel step, and from that same drop-off you can jump straight into a recording of what a real user actually did. Instead of “conversion dropped last week, someone go look into it,” it turns into “here are twelve sessions where a button looked clickable but wasn’t, and it’s only happening on mobile Safari.” That’s something you can act on today, not something that sits in the backlog until next sprint.</p><p>And the scale of what gets missed here is bigger than most teams assume. Pendo ran a study in 2019 across more than 600 customer products and found that 80% of features in the average software product are rarely or never used, which they estimated added up to around $29.5 billion in wasted engineering spend across publicly traded software companies alone. Teams aren’t just missing complaints. They’re missing entire features going untouched, and nobody ever finds out, because a feature nobody uses doesn’t generate a support ticket. It just generates silence.</p><h3>Where this goes next</h3><p>This is the first piece in a short series. The rest of it walks through, practically, how a product manager (not an engineer) can actually go and listen: building a first funnel to find where the real drop-off is happening, pairing it with session replay to see why, using feature flags to test a change without betting the whole release on it, and tying feedback to the exact moment it happens instead of a quarterly survey nobody remembers filling out.</p><p>The point isn’t to turn you into a data analyst. It’s to make sure that next time your client is talking without saying a word, whether that’s a rage click, a silent drop-off, or a feature they opened once and never touched again, somebody on your side is actually in a position to hear it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=01da654643a5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Africa’s Talking Voice API: A Complete Flow Guide for Building a Call Routing System]]></title>
            <link>https://medium.com/@josphatwanjiruw/africas-talking-voice-api-a-complete-flow-guide-for-building-a-call-routing-system-1073df28e5da?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/1073df28e5da</guid>
            <category><![CDATA[ivr]]></category>
            <category><![CDATA[design-systems]]></category>
            <category><![CDATA[voice-api]]></category>
            <category><![CDATA[africastalking]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Wed, 05 Aug 2026 07:50:22 GMT</pubDate>
            <atom:updated>2026-08-05T07:50:22.694Z</atom:updated>
            <content:encoded><![CDATA[<p>This is a flow-level reference for building a call center / support-line integration on top of Africa’s Talking’s Voice API the kind of thing you need once “answer one inbound call” isn’t enough and you’re routing to agents, queueing overflow, holding, transferring, and mixing browser softphones with real SIP phones.</p><p>It’s flow-focused on purpose: sequence diagrams and decision points, not a full code walkthrough. Everything here reflects behavior actually confirmed against live test calls and Africa’s Talking’s own documentation/support not assumptions from field names, which, as you’ll see, don’t always mean what they look like they mean.</p><h3>The core concept: one webhook, every event</h3><p>Africa’s Talking gives you exactly <strong>one</strong> voice callback URL per account. Every call-related event inbound calls, outbound call status, hold-music playback, browser-softphone dial-outs, SIP-phone dial-outs hits that same URL. Your handler has to branch on which fields are present to figure out what’s actually happening:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ksSFgVtp2fXNxNXNQmxuQg.png" /></figure><p>The single most important habit this API demands: <strong>every response must contain a real action element.</strong> There’s no such thing as “just acknowledging, no action needed” an empty &lt;Response&gt;&lt;/Response&gt; gets the call marked <strong>Aborted</strong> with &quot;No actions were found in the input&quot;. If you&#39;re not sure what to do yet, send something safe like &lt;Say&gt; never nothing.</p><h3>Resolving which tenant/account a call belongs to</h3><p>If you’re running this for more than one customer/organization off a single AT account, every call needs to resolve back to <em>whose</em> call it is. In priority order:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*33Zwy6s6_ShGfPrtBrewyw.png" /></figure><h3>Inbound call → route to an available agent</h3><p>The bread-and-butter flow: a customer dials your number, you find someone free to take it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3THe2thnmIzjoFhjjGd9Wg.png" /></figure><p>Two schema gotchas worth knowing up front:</p><ul><li><strong>&lt;Dial&gt; has no </strong><strong>clientName attribute.</strong> To address a connected browser softphone, put &lt;your-AT-username&gt;.&lt;clientName&gt; <em>inside</em> the phoneNumbers comma-list, not as a separate attribute: phoneNumbers=&quot;myaccount.agent123,+254700000000&quot;. Guessing at a clientName attribute gets you Missing phoneNumbers attribute in Dial action.</li><li><strong>&lt;Enqueue&gt; takes attributes, not text content.</strong> &lt;Enqueue&gt;hold&lt;/Enqueue&gt; does nothing useful; it&#39;s &lt;Enqueue holdMusic=&quot;https://...&quot; name=&quot;support&quot; /&gt;.</li></ul><p>&lt;Dial&gt;&#39;s real, confirmed attribute set: callerId, phoneNumbers (required), record (bool), sequential (bool), maxDuration (int), ringBackTone (an https URL). Nothing else.</p><h3>No agent available → queue → dequeue</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x5VTRb9z48F-Op-z0CU-_w.png" /></figure><p>The staleness check matters more than it looks: without it, a caller who queued and gave up hours ago can still get silently re-assigned to the next agent who frees up, leaving that agent stuck “busy” for a call that isn’t actually happening.</p><h3>Agent dials out via a browser softphone (WebRTC)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*26eeZkmbAsIs3dp6QXkarw.png" /></figure><p>This is the simplest of all the flows one webhook hit, one response, done.</p><h3>Agent’s SIP phone receiving a call (inbound)</h3><p>If an agent is on a real SIP phone (a desk phone or softphone client like Zoiper) instead of the browser, and they’re the one selected in the “Inbound call → agent” flow above, nothing changes on your end:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4hoMFwN1imfNWOKFcjx0Yw.png" /></figure><p>This stays inside the <em>same</em> webhook session as any other inbound call no separate event fires for the SIP leg specifically. If your &lt;Dial&gt; targets a SIP address correctly, this flow just works.</p><h3>Agent’s SIP phone placing a call (outbound) the flow that’s easy to miss</h3><p>This is the one that will cost you time if you don’t know it up front. When a SIP phone under your account places an <strong>outbound</strong> call to a customer, or even just to your own virtual number as a test AT does <strong>not</strong> treat it like a normal outbound dial. It hits your webhook as a fresh, standalone session:</p><pre>sequenceDiagram<br>    participant SIP as Agent&#39;s SIP phone<br>    participant AT<br>    participant You</pre><pre>    SIP-&gt;&gt;AT: Dials some number<br>    AT-&gt;&gt;You: POST direction=Outbound,&lt;br/&gt;destinationNumber=the SIP phone&#39;s OWN address,&lt;br/&gt;callerNumber=the number that was actually dialed<br>    You--&gt;&gt;AT: &lt;Dial phoneNumbers=&quot;{callerNumber}&quot;&lt;br/&gt;callerId=&quot;your virtual number&quot; record=&quot;true&quot;/&gt;<br>    AT-&gt;&gt;Note: Bridges to the actually-dialed number</pre><p>Two things about this payload will mislead you if you’re used to the normal inbound semantics:</p><ul><li><strong>destinationNumber is the SIP phone&#39;s own address</strong> not a real number, and not who&#39;s being called. You can reliably use it to look up which agent/tenant this call belongs to, since it&#39;ll match whatever SIP address you provisioned for that agent.</li><li><strong>callerNumber is the number the SIP phone actually dialed.</strong> Not who&#39;s calling. This is the field you need to bridge to, and it&#39;s genuinely counter-intuitive given every <em>other</em> webhook hit in this same callback uses callerNumber the normal way (who&#39;s calling you).</li></ul><p>If you send an empty response here, AT marks the call <strong>Aborted</strong> (&quot;No actions were found in the input&quot;) a required action element, same as everywhere else, so at minimum acknowledge with something. But acknowledging without a &lt;Dial&gt; just lets the call ring for a couple of seconds and die; nothing was ever told to connect anywhere.</p><h3>Hold and Unhold</h3><p>AT has <strong>no native “pause this call” primitive.</strong> The only mid-call control it exposes is callTransfer (a separate REST call: POST https://voice.africastalking.com/callTransfer, form body username, apiKey, sessionId, phoneNumber, optional callLeg &quot;caller&quot; for the PSTN/customer side, &quot;callee&quot; for the agent side, defaulting to &quot;callee&quot; and optional holdMusicUrl, which only plays <em>while ringing</em>, not on an indefinite loop). Real, sustained hold is built from two primitives glued together:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4P8aJ-OEDUSAJXfxRSk40g.png" /></figure><p>This needs one dedicated AT number purchased solely to act as the hold-music destination a real recurring cost, not just a code path. callLeg=&quot;caller&quot; is used for both Hold and Transfer because both relocate the <em>customer&#39;s</em> leg; you never need to move the agent&#39;s side.</p><h3>Transfer</h3><p>Same callTransfer primitive as Hold, just pointed at a different destination a real phone number, or another agent:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IVGTmLzne8XeWJ-aTXj0xg.png" /></figure><p>CustomerATYouCustomerATYoucallTransfer(sessionId, new destination, callLeg=caller)Leg moved to the new destination</p><h3>Recording</h3><p>Set record=&quot;true&quot; on your &lt;Dial&gt; and AT records the bridged call automatically no separate action needed to start it. The finished recording&#39;s URL arrives as a field on the Completed webhook hit (log the full raw POST body on that event once while testing; the exact field name is worth confirming against your own account, since this can vary test and confirm rather than assuming). From there it&#39;s a normal async download-and-store step on your end.</p><h3>The one habit worth taking away</h3><p>Every flow above was arrived at the same way: <strong>assume nothing about field semantics, respond with something, then check the actual result</strong> in AT’s own dashboard (Voice Session Details shows you exactly what request hit your webhook and what your response was interpreted as). Field names that look self-explanatory (callerNumber, destinationNumber) mean different things depending on which call shape you&#39;re in it&#39;s not always obvious from the names alone, worth confirming against a live test call. Guessing at schema costs you a broken production call; a live test call with the dashboard open costs you two minutes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1073df28e5da" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The One Field That Breaks Africa’s Talking SIP Outbound Calls (And How to Fix It)]]></title>
            <link>https://medium.com/@josphatwanjiruw/the-one-field-that-breaks-africas-talking-sip-outbound-calls-and-how-to-fix-it-42cb013a811b?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/42cb013a811b</guid>
            <category><![CDATA[voice-api]]></category>
            <category><![CDATA[ivr]]></category>
            <category><![CDATA[africastalking]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 09:18:58 GMT</pubDate>
            <atom:updated>2026-08-04T09:18:58.081Z</atom:updated>
            <content:encoded><![CDATA[<p>If you’ve provisioned a SIP phone Zoiper, Linphone, a desk phone, whatever under your Africa’s Talking Voice account and tried to let it place outbound calls through your own webhook, you’ve probably hit this: the call rings, plays fine on your end, and then just… ends. No error in your logs. Your webhook returns 200. And yet nothing ever actually connects.</p><p>This post documents exactly why, because the fix isn’t where you’d look first, and AT’s own docs don’t make the field semantics obvious.</p><h3>The setup</h3><p>Standard flow: a SIP client is registered under your AT Voice account (agent1.yourclientname@ke.sip.africastalking.com), provisioned from AT&#39;s own console. You&#39;ve already got a webhook handling <strong>inbound</strong> calls to your virtual number that part works fine, textbook AT Voice XML:</p><pre>&lt;Response&gt;<br>  &lt;Say&gt;Please hold while we connect you.&lt;/Say&gt;<br>  &lt;Dial phoneNumbers=&quot;agent1.yourclientname@ke.sip.africastalking.com&quot; record=&quot;true&quot; /&gt;<br>&lt;/Response&gt;</pre><p>Customer calls in, gets bridged to the SIP phone, everyone’s happy.</p><p>Now you want the <em>reverse</em>: someone picks up that same SIP phone and dials <strong>out</strong> to a customer, or to test by dialing your own virtual number. You’d assume AT treats this the same way any outbound call works: you place it via their REST API (POST https://voice.africastalking.com/call), or it just works because the SIP trunk is two-way.</p><p>It doesn’t. Here’s what actually happens.</p><h3>What AT actually sends you</h3><p>The moment the SIP client places an outbound call, AT hits your <strong>same</strong> voice callback URL a fresh, standalone webhook session, unrelated to anything you initiated:</p><pre>{<br>  &quot;callSessionState&quot;: &quot;Ringing&quot;,<br>  &quot;direction&quot;: &quot;Outbound&quot;,<br>  &quot;callerNumber&quot;: &quot;+254711082116&quot;,<br>  &quot;destinationNumber&quot;: &quot;agent1.yourclientname@ke.sip.africastalking.com&quot;,<br>  &quot;sessionId&quot;: &quot;ATVId_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx&quot;<br>}</pre><p>If you’re used to AT’s normal inbound semantics callerNumber is the person calling you, destinationNumber is the number they dialed this payload looks backwards. And it is, for this one specific call shape:</p><ul><li>destinationNumber is <strong>the SIP identity that placed the call</strong> not a real virtual number, and not who&#39;s being called.</li><li>callerNumber is <strong>the number the SIP phone actually dialed</strong> the thing you actually need to connect to.</li></ul><p>Nowhere in the payload is there an obviously-named field like numberDialed or to. Inspecting it cold, callerNumber reads like &quot;who&#39;s calling me,&quot; so it&#39;s natural to either ignore this webhook hit entirely or acknowledge it with an empty response, assuming it&#39;s just a status notification about a call your own code already knows about. Both are wrong, and both fail differently.</p><h3>Three attempts, in order (so you don’t have to burn them yourself)</h3><h3>1. Empty response, on the assumption this hit needs no action</h3><pre>&lt;Response&gt;&lt;/Response&gt;</pre><p><strong>Result:</strong> AT’s dashboard reports the call as <strong>Aborted</strong>, with the error:</p><blockquote><em>No actions were found in the input [</em><em>&lt;Response&gt;&lt;/Response&gt;]</em></blockquote><p><strong>Lesson:</strong> Africa’s Talking requires a real action element in <em>every</em> webhook response, regardless of call direction or leg. There’s no such thing as a “just acknowledging, no action needed” response in their XML dialect.</p><h3>2. A &lt;Say&gt; to at least clear the abort</h3><pre>&lt;Response&gt;&lt;Say playBeep=&quot;false&quot;&gt;Connecting your call.&lt;/Say&gt;&lt;/Response&gt;</pre><p><strong>Result:</strong> status flips to <strong>Success</strong> in AT’s dashboard but the call still lasts about 2 seconds and goes nowhere. Obvious in hindsight: &lt;Say&gt; speaks a message and then, with nothing else in the response, the call simply ends. It was never told to connect to anyone.</p><h3>3. The actual fix bridge to callerNumber</h3><pre>&lt;Response&gt;<br>  &lt;Dial phoneNumbers=&quot;+254712345678&quot; callerId=&quot;+254711082116&quot; record=&quot;true&quot; /&gt;<br>&lt;/Response&gt;</pre><p>Where +254712345678 is the value from the inbound callerNumber field (the number the SIP phone dialed), and +254711082116 is your own virtual number, passed as callerId so the person you&#39;re calling sees your business number instead of AT&#39;s shared default.</p><p>This is, in fact, exactly what Africa’s Talking’s own help center documents buried in “How do I make outbound calls from my SIP phone or any IP phone,” easy to miss if you’re not specifically looking for SIP-outbound behavior:</p><pre>callerNumber = request.values.get(&quot;callerNumber&quot;, None)</pre><pre>&lt;Response&gt;<br>  &lt;Dial phoneNumbers=&quot;{callerNumber}&quot; callerId=&quot;your-virtual-number&quot;/&gt;<br>&lt;/Response&gt;</pre><p>Once you see it stated plainly like that, it’s obvious. Getting there from the raw webhook payload alone is not.</p><h3>A minimal handler (Go, adapt freely)</h3><pre>func VoiceWebhook(c *fiber.Ctx) error {<br>    destinationNumber := c.FormValue(&quot;destinationNumber&quot;)<br>    callerNumber := c.FormValue(&quot;callerNumber&quot;)<br>    state := c.FormValue(&quot;callSessionState&quot;)</pre><pre>    // A SIP client under your account placed an outbound call.<br>    if isKnownSIPIdentity(destinationNumber) {<br>        c.Set(&quot;Content-Type&quot;, &quot;application/xml&quot;)</pre><pre>        if strings.EqualFold(state, &quot;Ringing&quot;) {<br>            if callerNumber == &quot;&quot; {<br>                // Nothing to bridge to - don&#39;t send a &lt;Dial&gt; with no target.<br>                return c.SendString(`&lt;Response&gt;&lt;Say playBeep=&quot;false&quot;&gt;Connecting your call.&lt;/Say&gt;&lt;/Response&gt;`)<br>            }<br>            return c.SendString(fmt.Sprintf(<br>                `&lt;Response&gt;&lt;Dial phoneNumbers=&quot;%s&quot; callerId=&quot;%s&quot; record=&quot;true&quot; /&gt;&lt;/Response&gt;`,<br>                callerNumber, yourVirtualNumber,<br>            ))<br>        }<br>        // Completed/other terminal states don&#39;t need an action.<br>        return c.SendString(`&lt;Response&gt;&lt;/Response&gt;`)<br>    }</pre><pre>    // ... your existing inbound-call handling ...<br>}</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*55UOK6Kg9uftXZQLRAsPHg.jpeg" /></figure><h3>The general lesson</h3><p>This is really a lesson about verifying third-party API behavior empirically rather than trusting field names or inferring semantics from a payload shape that looks familiar. callerNumber means something different depending on which call flow you&#39;re in, and every webhook response genuinely does need to describe a real action there&#39;s no passive &quot;OK, got it&quot; option.</p><p>If you’re integrating AT Voice with SIP phones and something is quietly failing after 2–3 seconds, check whether you’re responding to the right field for the call shape you’re actually in.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42cb013a811b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stability Is a Lagging Indicator]]></title>
            <link>https://medium.com/@josphatwanjiruw/stability-is-a-lagging-indicator-2970bc32ea50?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/2970bc32ea50</guid>
            <category><![CDATA[jobs]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[stability]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 07:19:18 GMT</pubDate>
            <atom:updated>2026-07-06T07:19:18.408Z</atom:updated>
            <content:encoded><![CDATA[<p>For decades, career advice came down to one sentence: get into government or a big corporate, and you’re set for life. It made sense. For a long time, it was true.</p><p>Here’s the uncomfortable part: it’s still <em>partly</em> true and that’s what makes it dangerous. Stability isn’t a lie. It’s a lagging indicator: a measurement of the past we keep mistaking for a guarantee about the future.</p><p>Look at where we were told to hide. In 2025, US employers announced 1.17 million job cuts the highest since the pandemic including roughly a quarter of a million federal government jobs. In the first half of 2026, tech layoffs jumped 83% year over year, with AI the most-cited reason for four consecutive months. Closer to home, Kenya merged 42 state corporations into 20 and dissolved nine. Standard Chartered Kenya has cut staff for eleven straight years from 2,048 employees in 2014 to 942 today. Branch laid off workers in Kenya and Nigeria while posting ~$30M in profit.</p><p>That last one matters most. <strong>Profitability no longer protects your job.</strong> Meta grew revenue 33% and cut 8,000 people the next month. The institution can be perfectly stable while your role inside it is not. When people say “the bank is stable,” they’re describing the institution. But you are not the institution you are a role, and roles are being repriced by technology faster than institutions ever have been. A big employer offers you a slower rate of change, not immunity from it. That slowness feels like safety. It’s just a longer fuse.</p><p>And there’s a quieter cost. In a large hierarchy, ideas don’t compete on merit they compete on altitude, surviving layers of sign-offs before reaching anyone who can say yes. Over years, that teaches you the safest idea is the one you never raise. You stop building the exact muscles the new market pays for: shipping, deciding, owning outcomes. The environment we choose for safety strips us of the skills we’ll need the day that safety runs out.</p><p>I’m not saying join a startup startups fail too. I’m saying stop buying stability as a product and start building it as a skill. Ship things with your name on them. Learn the AI tools reshaping your field before your employer schedules a workshop about them. Keep one foot somewhere decisions happen fast.</p><p>Stability was never a place you could sit down in. It was always a speed the speed at which you can adapt. The people who understand this early won’t need anyone’s guarantee. They’ll be the guarantee.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2970bc32ea50" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You’re Not Buying a Product. You’re Buying a Story.]]></title>
            <link>https://medium.com/@josphatwanjiruw/youre-not-buying-a-product-you-re-buying-a-story-a9668d8f1658?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9668d8f1658</guid>
            <category><![CDATA[product]]></category>
            <category><![CDATA[repackaging]]></category>
            <category><![CDATA[community]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Thu, 14 May 2026 21:10:09 GMT</pubDate>
            <atom:updated>2026-05-14T21:10:09.232Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YI3o8ZWeaIPgpOjT5U6-uQ.jpeg" /></figure><p>Picture two bottles of water sitting side by side at a petrol station. Same source. Same plastic. Same liquid inside. One costs twenty shillings. The other costs eighty. And the eighty shilling one keeps selling out first.</p><p>That image has stayed with me for years, because it describes something far bigger than water.</p><p>We live in an age where the raw material behind almost every product you use the loan app on your phone, the AI tool your team pays for monthly, the supplements on the wellness shelf is available to anyone willing to look. The factories are shared. The APIs are open. The ingredients are listed on the back of the tub. And yet we keep choosing the expensive bottle. We keep subscribing. We keep renewing. Not because the product is different, but because something about it feels different.</p><p>I used to think that was irrational. Now I think it’s the most rational thing consumers do.</p><p>Take financial products. In Kenya alone, Fuliza, Tala, Branch, and your SACCO’s emergency loan are all built on the same four-hundred-year-old primitive: one party extends capital, another repays it over time with interest. The product hasn’t changed since the Medicis were running Florence. What has changed is the story wrapped around it the speed, the name, the app icon, the WhatsApp message that says your funds are ready. That story is doing enormous work. It is, in many ways, the entire product.</p><p>The AI industry has made this even more transparent. Most of the tools your company is paying for right now the writing assistants, the coding copilots, the meeting summarisers are sitting on top of the same two or three foundation models. Strip away the interface and you would struggle to tell them apart. One founder I spoke to described building his “AI product” in a weekend: OpenAI API, a system prompt he spent three hours refining, a clean UI, and a Stripe checkout. He was charging $29 a month within the week. His users loved it. Not because the intelligence was his. Because the experience felt like it was made for them.</p><p>This is the world product builders are operating in now. The hard parts have been abstracted away. Infrastructure is rented. Models are borrowed. Factories are shared. And so the question that used to be “can we build it?” has quietly become something else entirely: “can we make people feel like this belongs to them?”</p><p>That shift is what David Spinks saw coming long before most. In <em>The Business of Belonging</em>, he tells the story of building CMX Summit a conference for community professionals at a time when nobody took community seriously as a business discipline. What he noticed was that the community managers attending his event were doing the same work, facing the same problems, and feeling completely alone in their companies. Nobody understood what they did. Nobody had a name for it.</p><p>When he put three hundred of them in a room together for the first time, something happened that no product feature could have produced. They wept. They felt seen. One attendee wrote years later that she remembered exactly what she was wearing that day, the colour of the scarf, the way the room looked because it was the first time in her career she hadn’t felt alone.</p><p>Spinks calls the mechanism behind that moment the Social Identity Cycle. Community creates shared identity. Shared identity creates belonging. Belonging creates the kind of loyalty that marketing budgets cannot buy and competitors cannot copy. It is, as he puts it plainly, the one thing they can’t replicate because it isn’t in the product. It’s in the people gathered around it.</p><p>Salesforce understood this. Airbnb understood this. Atlassian built an ecosystem of community advocates so deeply invested in the product that they became its most effective sales team. None of them won purely on features. They won because they made people feel like they were part of something a tribe, a movement, a room where everyone finally spoke the same language.</p><p>So here is what I keep coming back to, every time I sit in a product review or a roadmap session: the features we are debating will be cloned. They always are. The API we are wrapping will be available to our competitors tomorrow. The factory we are using already has ten other clients. None of that is our moat.</p><p>The moat is whether the people who use what we build feel, somewhere beneath all the utility and the convenience, that they belong to it.</p><p>The Medicis didn’t last because of their money. They lasted because people believed in what gathering around them meant. The best products have always worked the same way not by being irreplaceable, but by making the people around them feel like they are.</p><p>Your next sprint will ship. Your competitor will copy it. But the community that gathers around what you build the people who feel seen by it, who bring others into it, who defend it in rooms you’ll never enter that is the one thing no roadmap can produce and no funding round can buy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9668d8f1658" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Event-Driven Observability in AWS Detecting Root Account Logins in Real Time]]></title>
            <link>https://medium.com/@josphatwanjiruw/event-driven-observability-in-aws-detecting-root-account-logins-in-real-time-1926c412c5ed?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/1926c412c5ed</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[eventbridge]]></category>
            <category><![CDATA[aws-cloudtrail]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[aws-lambda]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Thu, 14 May 2026 07:36:27 GMT</pubDate>
            <atom:updated>2026-05-14T07:36:27.041Z</atom:updated>
            <content:encoded><![CDATA[<p>A while back, I came across an article by <a href="https://medium.com/u/f4dcb96c3cba">Paul Ciurean</a> about building event-driven observability on AWS to detect RDS storage auto-scaling events. The pattern was elegant EventBridge → Lambda → SNS — and it got me thinking.</p><p>What else in AWS happens silently, has no CloudWatch metric, and really deserves immediate attention?</p><p>The answer came quickly: <strong>root account logins.</strong></p><p>Your AWS root account is the most powerful identity in your entire AWS setup. It bypasses all IAM policies, can close the account, modify billing, and remove MFA. If someone logs into it, you want to know immediately — not hours later when you happen to check the console.</p><p>So the problem became: <em>How do we get notified in real time when a root account login occurs?</em></p><p>This led me to build the same event-driven pattern, and in this article I’ll walk you through the architecture, the Terraform implementation, and the lessons I picked up along the way.</p><h3>Understanding the Need</h3><p>The concern is straightforward. In a healthy AWS account, the root user should almost never log in. You create the account, set up IAM users with appropriate permissions, and lock the root credentials away.</p><p>So if root login fires — <strong>that itself is suspicious and worth investigating.</strong></p><p>The team needs to know immediately so they can:</p><ul><li>Verify whether the login was intentional</li><li>Investigate the source IP and timestamp</li><li>Respond quickly if credentials have been compromised</li><li>Maintain an audit trail for compliance</li></ul><h3>The Hidden Gap: No CloudWatch Metric for Root Logins</h3><p>When I looked through CloudWatch metrics for IAM and sign-in events, I ran into the same wall Paul described with RDS:</p><ul><li>AWS logs root login events internally</li><li>But CloudWatch has no metric for “root account just logged in”</li><li>You cannot create a standard CloudWatch alarm for this</li></ul><p>That is when it clicked — this is not a metrics problem. It is an <strong>event problem.</strong></p><p>Root login events exist in CloudTrail, and CloudTrail integrates natively with EventBridge. The solution was not metric-based monitoring. It was event-driven observability.</p><h3>The Solution: CloudTrail + EventBridge + Lambda + SNS</h3><p>Once the approach was clear, the architecture came together naturally.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Pk1gggmCLTlSWtqcpvumA.png" /></figure><p>The flow works like this:</p><ol><li>Someone logs into the AWS root account via the console</li><li>CloudTrail captures the login event and writes it to S3</li><li>EventBridge detects the event in real time and matches it against a filter rule</li><li>The rule triggers a Lambda function</li><li>Lambda formats the raw event into a clean, human-readable message</li><li>Lambda publishes the message to an SNS topic</li><li>SNS delivers an email alert to all subscribers</li></ol><p>Simple, reliable, and deployable in minutes.</p><h3>Why CloudTrail Is Required Here</h3><p>Unlike RDS events which EventBridge catches directly, IAM and sign-in events live in CloudTrail. EventBridge has native integration with CloudTrail — but CloudTrail must be enabled first.</p><p>This is also why there is an S3 bucket in the architecture. CloudTrail requires a destination bucket before it will activate. The S3 bucket is a passive filing cabinet — it sits in the background satisfying that requirement while EventBridge does the real-time work.</p><h3>Implementation</h3><p>The entire stack is deployed with Terraform, making it reproducible across multiple AWS accounts in minutes.</p><h3>CloudTrail</h3><pre>resource &quot;aws_cloudtrail&quot; &quot;main&quot; {<br>  name                          = &quot;${var.project_name}-trail&quot;<br>  s3_bucket_name                = aws_s3_bucket.cloudtrail_logs.id<br>  include_global_service_events = true<br>  is_multi_region_trail         = true<br>  enable_log_file_validation    = true<br><br>  depends_on = [aws_s3_bucket_policy.cloudtrail_logs]<br>}</pre><p>Two settings matter here. include_global_service_events = true is required because IAM is a global service, not region-specific. is_multi_region_trail = true ensures root logins are captured regardless of which region the session originates from.</p><h3>EventBridge Rule</h3><pre>resource &quot;aws_cloudwatch_event_rule&quot; &quot;root_login&quot; {<br>  name        = &quot;${var.project_name}-rule&quot;<br>  description = &quot;Captures AWS root account console login events&quot;<br><br>  event_pattern = jsonencode({<br>    source      = [&quot;aws.signin&quot;]<br>    &quot;detail-type&quot; = [&quot;AWS Console Sign In via CloudTrail&quot;]<br>    detail = {<br>      userIdentity = {<br>        type = [&quot;Root&quot;]<br>      }<br>    }<br>  })<br>}</pre><p>This is the brain of the entire stack. EventBridge watches every event flowing through your account and matches against this pattern. All three conditions must be true — sign-in service, console login event type, and user type of Root. A regular IAM user logging in will not trigger this rule.</p><h3>Lambda Function</h3><pre>def lambda_handler(event, context):<br>    detail      = event.get(&#39;detail&#39;, {})<br>    event_time  = event.get(&#39;time&#39;, &#39;Unknown time&#39;)<br>    source_ip   = detail.get(&#39;sourceIPAddress&#39;, &#39;Unknown IP&#39;)<br>    login_result = detail.get(&#39;responseElements&#39;, {}).get(&#39;ConsoleLogin&#39;, &#39;Unknown&#39;)<br>    account_id  = event.get(&#39;account&#39;, &#39;Unknown account&#39;)<br><br>    message = f&quot;&quot;&quot;<br>ROOT ACCOUNT LOGIN DETECTED<br><br>Account ID  : {account_id}<br>Time        : {event_time}<br>Source IP   : {source_ip}<br>Login Result: {login_result}<br><br>Root account usage is discouraged. If this was not you,<br>    immediately rotate your credentials and review your account.<br>    &quot;&quot;&quot;.strip()<br><br>    sns.publish(<br>        TopicArn=os.environ[&#39;SNS_TOPIC_ARN&#39;],<br>        Subject=&#39;AWS Root Account Login Detected&#39;,<br>        Message=message<br>    )</pre><p>Lambda transforms the raw CloudTrail JSON into something a human can act on immediately. The source IP is especially useful — a login from an unfamiliar country at 3am tells you everything you need to know.</p><h3>IAM Role — Least Privilege</h3><p>The Lambda function is granted only two permissions: publish to the specific SNS topic, and write its own execution logs to CloudWatch. Nothing else. It cannot read S3, touch IAM, or interact with any other AWS service.</p><pre>Statement = [<br>  {<br>    Effect   = &quot;Allow&quot;<br>    Action   = &quot;sns:Publish&quot;<br>    Resource = aws_sns_topic.root_login_alerts.arn<br>  },<br>  {<br>    Effect = &quot;Allow&quot;<br>    Action = [&quot;logs:CreateLogGroup&quot;, &quot;logs:CreateLogStream&quot;, &quot;logs:PutLogEvents&quot;]<br>    Resource = &quot;arn:aws:logs:*:*:*&quot;<br>  }<br>]</pre><h3>“But I Have 2FA — Why Does This Matter?”</h3><p>A fair question. 2FA is a strong control, but it is not the only one you need.</p><p>Root access keys have no 2FA at all. If someone creates root access keys and those keys are accidentally committed to a public GitHub repository — which happens more often than it should — an attacker can use them directly via the API without ever touching the console login flow.</p><p>More importantly, this stack is a <strong>detective control</strong>, not a preventive one. It does not stop the login. It ensures you know about it immediately so you can determine whether it was legitimate or an incident. Security without visibility is just hope.</p><h3>Lessons Learned</h3><p>This project reinforced a few things worth sharing:</p><ul><li>CloudWatch alarms are not the only observability tool. EventBridge gives you access to events that have no metric equivalent.</li><li>Root login should be a rare event in a healthy account. If it fires unexpectedly, that signal alone is worth investigating.</li><li>Least privilege on Lambda is not optional — it is a two-minute addition that meaningfully reduces blast radius if something goes wrong.</li><li>The entire stack costs effectively nothing at low volume. A root login alert that fires once or twice a year generates negligible CloudWatch logs and SNS costs.</li><li>Terraform makes this pattern deployable across ten AWS accounts in ten minutes. That repeatability is the real value for teams managing multiple environments.</li></ul><p>If you found this useful, feel free to share it or reach out with questions. Always happy to discuss DevOps and AWS automation. <a href="https://github.com/josphat-mwangi/IAM-Root-Login-Observability">github repo</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1926c412c5ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is Your Company Actually Productive — or Just Busy?]]></title>
            <link>https://medium.com/@josphatwanjiruw/is-your-company-actually-productive-or-just-busy-42e0fab58d2e?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/42e0fab58d2e</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[remote-work]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Sat, 04 Apr 2026 17:20:17 GMT</pubDate>
            <atom:updated>2026-04-05T01:39:37.701Z</atom:updated>
            <content:encoded><![CDATA[<h3>Is Your Company Actually Productive or Just Busy?</h3><p><em>Fuel costs are rising. AI agents are here. The office-first model is under real pressure. The companies that figure out how to measure, run, and scale their work intelligently before the disruption forces them to Will win. The ones that don’t will keep confusing motion with progress.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9VsndxxlncB5FJlJXYntNQ.png" /></figure><p>Most companies measure productivity by activity. Hours logged. Meetings held. Slack messages sent. It feels like control. It isn’t. Real productivity is simpler and a lot harder: did the right things get done, by the right people, in the shortest time possible? And can you actually prove it?</p><p>That question is getting more urgent by the day. Fuel prices are trending in one direction. Commutes are becoming a real cost in money, time, and energy that people don’t get back. Remote and hybrid work is no longer a conversation about flexibility. It’s quietly becoming a conversation about survival. And companies still anchored to presence as a proxy for performance are one bad quarter away from a very uncomfortable realisation.</p><p>Remove the office, and everything that was already broken becomes impossible to ignore.</p><p><strong>The foundation nobody talks about</strong></p><p>Company performance is not a company problem. It is the sum of individual performances and most individuals have never been given a clear system for measuring their own output honestly.</p><p>Here’s one that works. Run your week like a personal sprint. Three priorities, ranked. Not ten. Not a task list. Three things that, if done well, would make the week genuinely count. On Friday, score yourself did you deliver? What got in your way? That reflection is not a formality. It is where improvement actually happens, week by week.</p><p>When every person on a team does this and when managers create space for that conversation something real shifts. People stop filling their days with motion and start making deliberate choices about impact. That discipline, personal accountability for outcomes, is the foundation everything else is built on.</p><p>Scale it up: the team does the same with sprint planning. The company does the same with quarterly goals. Each layer feeds the next. A company cannot be efficient if its teams are unclear. Teams cannot flow if individuals don’t know what they are supposed to be delivering. Fix the foundation first.</p><p><strong>Agile is not just for developers</strong></p><p>There’s a rhythm that the best-run engineering teams figured out years ago, and most of the business world still hasn’t borrowed it. Plan in short cycles. Ship something real. Reflect honestly. Adjust and go again.</p><p>That’s agile thinking and it belongs everywhere, not just in product and tech.</p><p>Sprint planning at the start of every two weeks: what are we actually committing to deliver, not just work on? A daily standup of fifteen minutes what did I do yesterday, what am I doing today, what is blocking me? A sprint review at the end of the cycle where you demonstrate what shipped, not what was worked on. And a retrospective where the team asks, honestly, what do we fix next time?</p><p>The retrospective is where most teams cheat themselves. It becomes a formality. People say things are fine. Nothing changes. But the teams that use it properly the ones who are genuinely honest about what slowed them down those are the ones that actually get faster over time.</p><p>Apply this rhythm beyond engineering. Sales teams. Customer support. Operations. Marketing. The cycle is universal because the problem is universal: without a structured moment to reflect, teams repeat the same mistakes indefinitely.</p><p><strong>The tools that show productivity, not just activity</strong></p><p>The right tools don’t just help you work they make your output visible. That distinction matters because visibility builds trust, and in a remote or hybrid environment, trust is what holds teams together.</p><p><strong>Linear or Jira</strong> for sprint and backlog management. You see exactly what’s in progress, what’s done, and what’s blocked. The metrics that come out of this cycle time, sprint completion rate, throughput are your actual productivity data. Not the hours someone spent at their desk.</p><p><strong>Notion or Confluence</strong> for documentation and async updates. A good team doc replaces three meetings that didn’t need to happen. Decisions get written down. Context gets shared. New people can get up to speed without booking time with five different people.</p><p><strong>Slack or Teams</strong>, used well, kills unnecessary meetings. Used badly, it becomes one long meeting that never ends. The discipline is to keep it async by default and protect synchronous time for things that actually need it.</p><p><strong>Loom or Grain</strong> for async video. Walk someone through a decision, a demo, or a piece of work without scheduling a call. They watch it when they’re ready. They can replay it. It’s faster for everyone.</p><p><strong>Zendesk or Freshdesk</strong> for customer tickets. Every issue tracked, assigned, timed, and resolved. The numbers that matter here are first response time, full resolution time, and customer satisfaction score. Those three tell you more about your support team’s productivity than any headcount metric.</p><p>And then there’s AI ChatGPT, Claude, Gemini, whatever your team reaches for. First drafts, research summaries, decision support, meeting notes. The teams using these tools well are not working harder. They’re working faster, with more energy left for the things that actually require human judgment.</p><p>What you measure matters as much as which tools you use. Cycle time tells you if your team is getting faster. Sprint completion rate tells you if your planning is realistic consistently below 70% means something is wrong upstream, not with the people. Ticket resolution time tells you if your customers are actually being helped. And a simple weekly self-score did each person deliver their three priorities? surfaces blockers before they become crises.</p><p><strong>Kill the meeting. Keep the decision.</strong></p><p>This is where companies bleed hours they never get back, and it gets worse when teams go remote. Someone schedules a call because it feels like progress. Ten people join. Half of them didn’t need to be there. Forty-five minutes pass. Nothing gets decided. Everyone goes back to their actual work, now behind.</p><p>The agile ceremonies standup, planning, retro, review are short, purposeful, and time-boxed for a reason. The rest is supposed to be deep work. That principle doesn’t only apply to developers.</p><p>A meeting must have a decision to make, or it should not exist. Updates go in writing. Questions go in a thread. If it can be resolved async, it must be. When meetings do happen: thirty minutes maximum, only the people who actually need to be there, and one person owns the follow-up or it didn’t happen.</p><p>Protecting your team’s time is a cultural statement. It tells people that their deep work matters more than your need to feel informed.</p><p><strong>AI is already in the building. The question is whether you’re using it right.</strong></p><p>AI is not something coming in the future. It’s already changing what is possible in a working day, for teams that know how to use it. The intelligent company is not one that has replaced its people with agents. It’s one where every person is more capable because of the systems around them.</p><p>A developer using AI ships faster and with fewer blockers. A customer service agent with AI surfacing ticket history and suggested responses resolves issues in half the time. A manager whose team uses AI for first-draft thinking makes better decisions with less preparation time and fewer meetings needed to align on context.</p><p>Agentic AI systems that act, not just respond is the next layer that’s arriving right now. Ticket routing, follow-up scheduling, data gathering, status updates: agents can run these in the background while your team focuses on what only humans can do. But there’s a catch. Agents need clean processes to follow. If your workflows are messy right now, agents will automate the mess. The intelligent company fixes the process first, agile-style, then lets technology scale it.</p><p><strong>Speed closes tickets. Empathy builds loyalty.</strong></p><p>This is where most companies get the AI + customer service combination wrong. They reach for automation, they chase speed, and they quietly lose trust in the process. A ticket closed in three minutes by a system that clearly didn’t understand the issue is not a win. The customer raises it again. Now you have two tickets, a frustrated person, and a longer resolution than if someone had just handled it properly the first time.</p><p>Speed matters. Feeling heard matters more.</p><p>The companies that win are the ones that use AI to move fast on everything that can be handled consistently instant acknowledgement, smart routing to the right team, surfacing customer history and context so the agent already knows the story, first-response drafts for common issues, SLA tracking so nothing falls through. And they use humans for everything that can’t be scripted: the complex case, the frustrated customer who needs to feel that someone actually cares, the conversation that requires reading what isn’t being said.</p><p>The goal is not to make customers feel like they’re talking to a fast company. It’s to make them feel like they’re talking to a company that knows them one that remembers their history, responds quickly, resolves things properly, and leaves them feeling respected. AI makes that possible at scale. The human touch makes it feel real.</p><p>The company that gets this blend right doesn’t have to choose between efficiency and empathy. It delivers both faster resolution, genuine connection, and customers who trust it enough to come back and bring others with them.</p><p>The fuel pressure, the AI shift, the quiet end of presence as a performance metric these are not separate trends. They are one moment, arriving all at once. The intelligent company is not a technology project. It’s a people project, running on agile principles, powered by the right tools, and never letting speed become an excuse for losing the human touch that makes people trust you in the first place.</p><p>Start building it now before the disruption makes the decision for you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42e0fab58d2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your OTP Is Being Held Hostage.]]></title>
            <link>https://medium.com/@josphatwanjiruw/your-otp-is-being-held-hostage-a269926e17eb?source=rss-87acc5d3d7ad------2</link>
            <guid isPermaLink="false">https://medium.com/p/a269926e17eb</guid>
            <category><![CDATA[sms]]></category>
            <category><![CDATA[otp-sms]]></category>
            <category><![CDATA[otp-verification]]></category>
            <dc:creator><![CDATA[Josphat mwangi]]></dc:creator>
            <pubDate>Tue, 31 Mar 2026 04:12:48 GMT</pubDate>
            <atom:updated>2026-03-31T04:12:48.418Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>And Your Customers Are Paying the Price.</strong></p><p><em>Why Africa’s verification infrastructure is overdue for a rethink — and what product and engineering leaders can do about it today.</em></p><blockquote><strong><em>“We lost 40% of our weekend signups. Our OTP was blocked. By Monday, the moment was gone.”</em></strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vm4gb6_qrKGjjKqlTd2tNw.png" /></figure><p>That is a real conversation I had with a product manager at an African fintech company. Their SMS sender ID was flagged on a Friday evening — no alert, no fallback, just silence — right as their weekend promotion peaked.</p><p>This is not a one-off incident. It is a structural problem that repeats itself every week across African markets, quietly killing onboarding flows, blocking legitimate customers, and eroding trust in digital services.</p><p>The cause is almost always the same: we have built our entire verification layer on top of a single, increasingly unreliable pipe. And that pipe is controlled by parties whose commercial interests do not always align with ours.</p><p>This article is for decision-makers — CTOs, product heads, engineering leads, and compliance officers at banks, fintechs, insurers, telcos, betting &amp; gaming and SaaS platforms. The way we verify users is outdated, fragile, and exclusionary. We can do better. The tools to do it already exist.</p><h3>A Quick Note on Terminology: OTP and CVC</h3><p>Throughout this article, you will see two terms used: OTP (One-Time Password) and CVC (Customer Verification Code). In practice, many organisations use these terms interchangeably, and that is fine — the label matters far less than the purpose.</p><p>Both refer to the same underlying goal: confirming that the person initiating an action is who they claim to be, using a temporary credential that expires after use. Whether your team calls it an OTP, a CVC, a verification code, or a one-time PIN, the problem described in this article — and the solutions proposed — apply equally. Use the term that fits your context and your customers.</p><h3>The Problem Is Not Technical. It’s Structural.</h3><p>When most product teams build user verification, they reach for SMS OTP almost by reflex. It feels safe, familiar, and good enough. But in the African context, this default carries hidden costs that compound over time.</p><h3>International services face a gatekeeping problem</h3><p>Here is something that does not get discussed openly enough: when a global platform — a cloud provider, an email infrastructure company, a SaaS service — tries to send an OTP to a user on a local mobile network, that message may simply not arrive. Not because of a technical fault. Because of a commercial one.</p><p>Mobile operators in several African markets have implemented policies that treat internationally-originated SMS traffic differently from local traffic. Some require the sending party to pay international interconnect fees. Some route international SMS through content filtering systems that flag unfamiliar sender IDs. Some have outright blocked certain categories of traffic without public announcement.</p><p>The result is that a developer in Nairobi trying to verify their account on AWS, a marketing team using a global email platform, or a startup integrating an international payments provider may find that OTPs simply do not arrive — depending entirely on which network their SIM is registered to. The same message that delivers fine on one operator may silently fail on another.</p><p>This is not a hypothetical. It is a daily friction point for African developers and businesses trying to participate in the global digital economy.</p><h3>Sender ID blocks: silent, sudden, and timed badly</h3><p>Sender ID blocking happens when a regulatory body, an operator, or a content moderation system flags a particular sender name. These blocks can arrive at any time. They are rarely announced in advance. And they tend to hit hardest at the worst possible moments.</p><p>Gaming and financial platforms, for example, often see their peak engagement on weekends and event days. A sender ID blocked on a Friday evening may not be discovered until Monday morning. By then, an entire weekend of onboarding attempts, transaction confirmations, and password resets has failed silently. The customers who could not verify did not wait — they moved on.</p><p>What makes this particularly damaging is that the failures are invisible at the user level. The customer sees an error or simply never receives their code. They do not know why. They do not file a support ticket. They just leave.</p><h3>Regulatory change as commercial leverage</h3><p>It would be unfair to characterise every operator policy decision as purely commercial. Regulators have legitimate reasons to control SMS traffic — fraud prevention, spam reduction, consumer protection. These are real concerns.</p><p>But it is equally important to acknowledge the pattern: verification-related policies in several markets have changed in ways that happen to increase costs for businesses, particularly those routing traffic through international providers. When a new requirement is introduced — mandatory local sender ID registration, new interconnect fee structures, content filtering for unrecognised senders — the effect is the same regardless of the intent: African businesses and their customers bear the cost.</p><p>The strategic response is not to fight that dynamic. It is to architect around it.</p><h3>The 2G user is not a fringe case</h3><blockquote><strong><em>A significant proportion of active mobile users across African markets still operate primarily on 2G or basic feature phones — not because smartphones are unavailable, but because 2G is more affordable, more battery-efficient, and more reliable in many areas.</em></strong></blockquote><p>Designing verification flows exclusively for smartphone users with stable data connections is a quiet form of market exclusion. In a continent where financial inclusion is both a social good and a business opportunity, that exclusion is expensive. Every user who cannot complete verification because your flow assumes a modern device and a stable data connection is a customer you did not serve.</p><h3>Rethinking Verification: Format, Channel, and Trust</h3><p>The core insight is simple: verification does not have to mean a six-digit numeric code delivered over SMS. That is one implementation of a broader idea. The goal is to confirm identity using a temporary credential that works for the user, delivered through a channel that is reliable for their context.</p><p>When we expand that definition, we open up a much richer solution space.</p><h3>The format does not have to be a number</h3><p>Consider how some global platforms have moved away from numeric codes entirely. Instead of sending a six-digit number and asking the user to type it, the system displays a number on the user’s screen and asks them to tap the matching number on their phone — or vice versa. No typing. No risk of misreading digits. No copying from a notification.</p><p>A word-select approach takes this further: the user is shown three words and asked to confirm which one matches what was communicated to them. This format is speakable over a voice channel, legible on small screens, accessible for users with lower numeric literacy, and harder to phish than a traditional code because there is nothing to type into a fake form.</p><p>A short passphrase — something like “Blue-Mango-7” — is another option. It is memorable, human-readable, typeable on a feature phone keypad if needed, and degrades gracefully across channels: spoken over voice, confirmed with a keypress over USSD, tapped in an app. The format becomes channel-agnostic by design.</p><p>None of these alternatives are weaker than a six-digit number. Several are stronger. The important thing is that the format should be chosen to serve the user’s context, not inherited by default.</p><h3>Addressing user trust and familiarity</h3><p>Any new verification method has to earn user trust, and that requires intentional design. Two methods that often raise questions deserve specific attention.</p><p>Verification links sent via email or messaging apps are an excellent channel when implemented with clear, transparent UX. The concern — that users may not be able to distinguish a legitimate link from a phishing attempt — is valid, and it is a design problem, not a reason to avoid the channel. The solution is to display the link destination explicitly, show a preview of where it leads, and educate users inside the product about what legitimate links from your platform look like. Trust is built through consistency and clarity, not by avoiding the channel.</p><p>Call-based verification methods — where the system initiates a call to the user — can seem unfamiliar to users who are not expecting them. The solution here is also UX: the user should be told, before the call is initiated, exactly what to expect. “We are about to call you on this number. Listen for the code and enter it below.” Or: “You will receive a call. Press 1 to confirm this action.” When the user is prepared, a call is not alarming — it is reassuring. It is a human-scale interaction in a world full of impersonal digital flows.</p><p>The principle applies across all non-standard channels: unfamiliarity is a communication problem, not a channel problem. Solve it with clear in-product messaging, consistent branding, and user education built into the flow.</p><h3>The Multi-Channel Verification Stack</h3><p>A modern, Africa-ready verification system treats every channel as one option in a coordinated stack — selected based on the user’s context, with automatic fallback if a channel fails. Here is what that looks like in practice:</p><h3>For users with internet access</h3><p>• <strong>TOTP via authenticator app: </strong>Completely offline, no telco involved, works anywhere. A user who has set up TOTP is immune to SMS-related outages entirely.</p><p>• <strong>Email verification link: </strong>Independent of mobile networks. With clear, branded UX that shows the user exactly where they are going, this is fast and reliable.</p><p>• <strong>In-app word-select or number-tap: </strong>No SMS sent at all. The verification happens entirely inside the authenticated session, with a single tap.</p><p>• <strong>Push notification: </strong>For returning users who have verified before, a push to their registered device is fast and seamless.</p><h3>For 2G and feature phone users</h3><p>• <strong>Voice call OTP: </strong>The system calls the user and speaks the code clearly. Works on any network, any handset, any literacy level. Completely bypasses SMS infrastructure and its associated gatekeeping.</p><p>• <strong>USSD session verification: </strong>A verification prompt initiated through a short-code session. No internet required, works on 2G, and typically subject to far less content filtering than SMS. The user enters a code or presses a key to confirm.</p><p>• <strong>Messaging app OTP: </strong>For users with WhatsApp or similar apps, delivery through an OTT channel bypasses SMS gatekeepers entirely and tends to have higher delivery reliability in markets with complex SMS routing.</p><p>• <strong>Call-based confirmation: </strong>The user receives a call and presses a key to approve an action. No code to remember or type. Useful for users who find numeric input under pressure difficult.</p><h3>Building Resilience: The Fallback Cascade</h3><p>A multi-channel stack only delivers value if it is wired together intelligently. The goal is that no single channel failure should ever result in a user seeing a dead end.</p><p>A well-designed fallback cascade works like this: the system attempts the primary channel and monitors for a delivery confirmation. If none arrives within a defined window, it automatically tries the next channel in the cascade — moving from SMS to voice, from voice to USSD, from USSD to messaging app — until the verification is delivered. The user sees a single, clean experience. The complexity is handled entirely by the system.</p><p>What makes this operationally sustainable is a channel health monitoring layer. The system should track delivery rates per route in near-real time, automatically deprioritising routes that fall below a reliable threshold and rerouting traffic before a human even notices the problem. When a sender ID is blocked, the system detects the delivery drop and reroutes within minutes — not after a Monday morning incident review.</p><blockquote><strong><em>The goal is not zero channel failures. The goal is zero user-visible failures. Failures should be handled inside the system, not handed to the customer as an error message.</em></strong></blockquote><h3>On Security: Diverse Channels Are More Robust, Not Less</h3><p>A concern that comes up whenever this conversation happens is security. If we allow voice, USSD, or word-select formats instead of the familiar six-digit SMS code, are we weakening verification?</p><p>In most cases, the alternatives are more secure than traditional SMS OTP.</p><p>SMS-based verification is vulnerable to several well-documented attack vectors: SIM swap fraud, protocol-level interception, and social engineering of operator customer service agents. These are active threats in African markets, not theoretical ones. The same telco infrastructure that is subject to commercial policy changes is also subject to exploitation by bad actors with access to the right channels.</p><p>Voice OTP cannot be intercepted by SIM swap in the same way as SMS, because the code is not stored on the SIM. TOTP codes are generated locally on the user’s device and never transmitted over any network at all. In-app confirmation methods require an active, authenticated session. USSD sessions are terminated at the network level when the call ends, leaving no persistent message to intercept.</p><p>A diverse channel stack also removes the single point of failure that SMS currently represents. If one channel is compromised, blocked, or fails commercially, the others continue to function. Resilience and security reinforce each other here.</p><p>The appropriate response to security concerns is not to defend SMS OTP as the baseline — it is to design a verification architecture where multiple channels work together, each with its own threat model understood and mitigated.</p><h3>For Decision-Makers: Where to Start</h3><p>You do not need to rebuild your verification stack overnight. Here is a practical sequence:</p><h3>1. Audit your current failure rate</h3><p>Pull your OTP delivery logs for the past 90 days and segment by route. You will almost certainly find that delivery rates vary significantly across different operators and time windows. That data is your business case for change — and it is probably more compelling than you expect.</p><h3>2. Add one non-SMS channel as a fallback</h3><p>Start with voice call OTP or USSD as a fallback for failed SMS delivery. This single addition can recover a significant portion of failed onboarding attempts with relatively modest implementation effort. It also gives you a tested fallback channel before you need it urgently.</p><h3>3. Build sender ID health monitoring</h3><p>Before your sender ID gets blocked, put monitoring in place. Define what a delivery rate drop looks like, who gets alerted, and what the runbook is. A block discovered within minutes is a minor incident. A block discovered on Monday morning is a weekend of lost users.</p><h3>4. Design for the lowest-connectivity user from the start</h3><p>When building new verification flows, treat the feature phone user on a 2G connection as the primary persona. A flow that works in that context will work for everyone. The reverse is not true. This is also a good forcing function for simplicity: if your verification UX requires explanation, it needs to be redesigned.</p><h3>5. Educate users on any new channel you introduce</h3><p>Every new verification channel requires a user education moment built into the product. Tell users what to expect before they experience it. Consistent, clear in-product messaging is what turns an unfamiliar channel into a trusted one over time.</p><h3>The Bigger Picture</h3><p>Africa’s digital economy is growing faster than almost anywhere else in the world. The ambition is extraordinary. The infrastructure underneath it needs to match.</p><p>Verification and authentication are foundational infrastructure. They are the first interaction a new user has with your product, and the last line of defence for your most sensitive transactions. When they fail — silently, on a Friday evening, because of a commercial decision made somewhere else in the stack — the damage is invisible but real. A customer who did not onboard. A transaction that did not complete. A relationship that never began.</p><p>The dependence on a single channel, controlled by parties whose incentives are not always aligned with your users’ needs, is a risk that has been normalised. It should not be. The tools to build a more resilient, more inclusive, more user-centric verification layer exist. The question is whether we choose to use them.</p><blockquote><strong><em>Every verification failure is a customer who did not get through. Building a verification stack that works for all your users — on every network, every device, every channel — is not a technical nicety. It is a business imperative.</em></strong></blockquote><p>In a follow-up technical article, I will walk through the implementation of a multi-channel verification engine using open source tools — covering the channel abstraction layer, fallback cascade logic, health monitoring, and format adapters for SMS, voice, USSD, messaging apps, and TOTP. If you are building in this space, that piece is for you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a269926e17eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>