<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
<title>Ryan Morrison AI</title>
<link>https://www.ryanmorrisonai.com/</link>
<atom:link href="https://www.ryanmorrisonai.com/rss.xml" rel="self" type="application/rss+xml"/>
<description>Essays on practical AI for organisations, written in Amsterdam.</description>
<language>en</language>
<item><title>Why teams push back on AI, and what actually helps</title><link>https://www.ryanmorrisonai.com/essays/teams-and-ai</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/teams-and-ai</guid><pubDate>Tue, 17 Nov 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Resistance is usually a sign of unclear goals and missing time. Both are fixable.</description><content:encoded><![CDATA[<p>Every AI rollout reaches the same moment. The tools are in place, the kick-off went well, and a few weeks later usage is flat. Leadership concludes that people are resisting change. Sometimes that is true. More often, the organisation has asked people to change without telling them why, what for, or when.</p>
<p>Here are the five reasons I see most often, and what helps with each.</p>

<h2>1. "Will this replace me?"</h2>
<p>If leadership does not answer this question, people answer it themselves, usually with the worst case. Silence reads as confirmation.</p>
<p><strong>What helps:</strong> say clearly what AI is meant to do in your organisation, and what will happen with the time it frees up. If the plan is to grow without hiring, say so. If roles will change, explain how people will be supported. You can't promise that nothing will change, but you can be honest about the direction.</p>

<h2>2. "Use AI" is not a task</h2>
<p>Telling a team to "start using AI" is like telling them to "start using the internet". Without a concrete task, most people try it once, get an average result, and go back to the old way.</p>
<p><strong>What helps:</strong> attach AI to specific tasks people already do, preferably ones they dislike. "Use this prompt to draft the first version of the monthly report" gets adopted. "Explore how AI can help your work" does not.</p>

<h2>3. The tool adds work instead of removing it</h2>
<p>If using AI means copying data between three windows and checking every output line by line, people will rightly decide it is not worth it.</p>
<p><strong>What helps:</strong> measure the whole task, not just the AI step. Sometimes the fix is a better prompt, sometimes a small integration, sometimes the conclusion that this task is not a good fit yet. All three are useful answers.</p>

<h2>4. Worries about quality and responsibility</h2>
<p>Professionals care about their work. When they don't know whether they are allowed to put client data into a tool, or who is responsible if the AI makes a mistake, the safe choice is not to use it.</p>
<p><strong>What helps:</strong> a short, clear rule sheet on approved tools and data, plus a simple principle: the person who sends the work owns it, whoever drafted it. Clarity makes people braver.</p>

<h2>5. No time to learn</h2>
<p>Learning a new way of working takes a few hours of trying, failing and adjusting. If those hours have to come on top of a full week, they won't happen.</p>
<p><strong>What helps:</strong> give people protected time, even two hours a week for a month. Appoint one or two AI champions per team, people who are curious rather than necessarily technical, and let them share what works in short sessions. Peer examples persuade far more than vendor demos.</p>

<h2>The role of leadership</h2>
<p>The single strongest signal is leaders using AI visibly themselves and talking about it honestly, including where it failed. When a manager says "I tried this for my board update, here's what worked and what didn't", it gives everyone permission to experiment.</p>
<p>Finally, measure what matters. Count hours given back and what they were used for, not the number of logins. Adoption follows usefulness. If people are not using AI, the first question is not "why are they resisting?" but "what have we given them that is worth using?"</p>]]></content:encoded></item>
<item><title>When the AI gets it wrong, who is accountable?</title><link>https://www.ryanmorrisonai.com/essays/who-is-accountable</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/who-is-accountable</guid><pubDate>Tue, 20 Oct 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>&quot;The AI did it&quot; is not an answer a customer or a regulator will accept. Decide who owns each AI-assisted decision before something goes wrong.</description><content:encoded><![CDATA[<p>Sooner or later, an AI system in your organisation will get something wrong. A summary will leave out the one sentence that mattered. A draft reply will promise something you can't deliver. A screening step will rank a good candidate too low. When it happens, the first question will be: whose fault is this?</p>
<p>If you only start answering that question after the mistake, the answer is usually "nobody's", which is the worst possible answer for a customer, for an employee, and increasingly for a regulator.</p>

<h2>The machine can't be accountable</h2>
<p>An AI model can produce a result, but it can't take responsibility for it. Accountability always lands with people: the person who used the output, the person who set up the system, or the person who decided to use AI for this task in the first place. The job is to decide in advance which of them it is.</p>

<h2>Three roles for every AI use case</h2>
<ul>
  <li><strong>The owner</strong> decided to use AI for this task and is accountable for the result overall. Usually the manager of the process, not someone in IT.</li>
  <li><strong>The builder</strong> set up the instructions, connections and checks, and is responsible for keeping them working.</li>
  <li><strong>The user</strong> acts on the output and is responsible for checking it before it matters.</li>
</ul>
<p>Write these names down for every AI use case that touches customers, money or people. In a small team one person may hold two roles. That is fine, as long as everyone knows.</p>

<h2>Match the check to the stakes</h2>
<p>Not every output needs the same scrutiny. An internal meeting summary can go out with a quick glance. A letter to a customer about their contract needs a careful read. A decision that affects someone's job, credit or access to a service needs a real human decision, with the AI as one input among several.</p>
<p>A simple rule: the more an error would cost the person on the other end, the more independent the check should be.</p>

<h2>Keep a trail</h2>
<p>For the use cases that matter, keep a record of what went in, what came out, and who approved it. You will need it the first time someone asks why a decision was made, and it is the only way to find out whether a mistake was a one-off or a pattern.</p>

<h2>What the AI Act adds</h2>
<p>The EU AI Act puts specific duties on organisations that use AI in high-risk areas, such as decisions about hiring, credit or access to essential services. These include human oversight by people with the right competence and authority. If any of your use cases come close to those areas, get proper advice. The role model above is a good starting point for that conversation, not a replacement for it.</p>

<h2>Make it safe to report mistakes</h2>
<p>The last part is cultural. If people fear blame, AI errors get quietly fixed and never reported, and the same error happens again next week. Treat the first reports as useful information, fix the instructions or the check, and tell people what changed. Accountability works best when it is about learning, not about finding someone to blame.</p>]]></content:encoded></item>
<item><title>Beyond the chatbot: what an AI agent actually is, and when you need one</title><link>https://www.ryanmorrisonai.com/essays/beyond-the-chatbot</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/beyond-the-chatbot</guid><pubDate>Tue, 15 Sep 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Chat, AI-powered workflows and agents are three different things. Most companies need the middle one first.</description><content:encoded><![CDATA[<p>"We need AI agents" has become the new "we need an app". The word gets used for almost anything, from a chatbot on a website to software that runs half a department. If you are deciding where to invest, it helps to separate three levels of AI, because they cost different amounts, carry different risks, and solve different problems.</p>

<h2>Level 1: chat</h2>
<p>This is what most people know: ChatGPT, Claude, Copilot, Gemini. You ask, it answers, and you copy the result somewhere else. The person stays in charge of every step.</p>
<p>Chat is excellent for drafting, summarising, brainstorming and learning. It is also where most companies stop. The limit is that the AI only knows what you paste in, and it can't do anything in your systems. Every gain depends on an individual remembering to use it.</p>

<h2>Level 2: workflows with an AI step</h2>
<p>Here the steps are fixed and designed by people, and AI handles one or two of them that need judgement. An incoming email arrives, the AI classifies it as a complaint, a quote request or an invoice, and the workflow sends it to the right queue with a draft reply attached.</p>
<p>This is where most practical value sits today. The process is predictable, you can test it, and when the AI is unsure you can send the case to a human. Many "agent" projects I see are better built as level 2.</p>

<h2>Level 3: agents</h2>
<p>An agent gets a goal instead of a script. It has access to tools, such as search, your CRM, a spreadsheet or email, and it decides which steps to take, checks the result and tries again if needed. "Prepare a briefing on this customer before tomorrow's meeting" is an agent task: the AI chooses what to look up and in which order.</p>
<p>Agents are powerful for work that varies a lot from case to case. They are also harder to test, because the path is different every time.</p>

<h2>The building blocks, in plain language</h2>
<p>Whatever level you build at, the same few pieces decide whether it works:</p>
<ul>
  <li><strong>The model.</strong> The language model doing the thinking. Important, but rarely the deciding factor any more.</li>
  <li><strong>Context.</strong> Everything the model can see at that moment: instructions, documents, the customer record. Most poor results come from poor context, not a poor model.</li>
  <li><strong>Tools.</strong> What the AI is allowed to do: read a database, send an email, create a ticket. Each tool is also a risk to manage.</li>
  <li><strong>Memory.</strong> What it keeps between tasks, such as earlier conversations or customer preferences.</li>
  <li><strong>Evaluation.</strong> A set of real examples with known good answers, used to check that the system still works after every change. This is the piece most often missing.</li>
</ul>

<h2>Where projects get stuck</h2>
<p>In practice, AI projects rarely fail because the model is not smart enough. They stall on access (the AI can't reach the data it needs), on exceptions (the ten percent of odd cases nobody designed for), on missing evaluation (nobody can prove it works, so nobody dares to switch it on), and on ownership (the pilot team moves on and the system has no home).</p>

<h2>So which one do you need?</h2>
<p>Start by asking how predictable the work is. If the steps are the same every time, build a workflow with an AI step. If every case is different and a skilled person would have to think about what to do next, an agent may be worth it. And if people simply need help writing and thinking, give them good chat tools and training.</p>
<p>When a vendor offers you an agent, ask four questions. What tools can it use? How do you test it? What happens when it is unsure? Who looks after it in six months? The answers tell you more than any demo.</p>]]></content:encoded></item>
<item><title>How to tell whether AI is actually helping</title><link>https://www.ryanmorrisonai.com/essays/is-ai-helping</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/is-ai-helping</guid><pubDate>Tue, 21 Jul 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Hours saved is the number everyone asks for and the one that misleads most. Measure the work, not the tool.</description><content:encoded><![CDATA[<p>Six months into an AI programme, someone on the leadership team asks the obvious question: is it working? Usually the answer is a mix of licence usage numbers, a survey where people say they feel more productive, and a few good stories. None of it is wrong, but none of it really answers the question.</p>
<p>Here is a simple way to measure that holds up when the budget conversation comes around.</p>

<h2>Usage is not value</h2>
<p>The number of active users tells you people opened the tool. It doesn't tell you whether anything got better. High usage can mean a tool is genuinely useful, or that people were told to use it. Low usage can mean it is useless, or that the three people using it are saving the company a lot of time on one important task.</p>
<p>Track usage, but treat it as a signal to investigate, not as the result.</p>

<h2>Be careful with "hours saved"</h2>
<p>Self-reported time savings are almost always too high. People remember the time a task used to take on its worst day and compare it with the AI's best day. And saved time only matters if it goes somewhere: to more customers, faster answers or work that was previously left undone. Twenty minutes saved and spent on extra email is not a business result.</p>

<h2>Measure the work, before and after</h2>
<p>For each use case, pick two or three measures of the work itself, and record them before you start. Good measures are things the team already cares about:</p>
<table class="rubric">
  <thead><tr><th>Measure</th><th>Example</th></tr></thead>
  <tbody>
    <tr><td>Speed</td><td>Time from customer question to first useful answer</td></tr>
    <tr><td>Volume</td><td>Number of quotes sent per week by the same team</td></tr>
    <tr><td>Quality</td><td>Share of reports sent back for corrections</td></tr>
    <tr><td>Backlog</td><td>Number of open tickets at the end of the week</td></tr>
  </tbody>
</table>
<p>Measure the same things for a few weeks before the change and a few weeks after. It doesn't need to be a scientific study. It needs to be the same measure, taken the same way, before and after.</p>

<h2>Ask one honest question</h2>
<p>Next to the numbers, ask the people doing the work one question: "If we took this away tomorrow, would you mind?" It cuts through politeness. Tools that people would fight to keep are delivering something, even if the numbers are still noisy. Tools that nobody would miss can be stopped.</p>

<h2>Stop things on purpose</h2>
<p>A healthy AI programme stops things. If a use case hasn't improved the work after a fair test, end it and say so. It frees budget and attention, and it makes the remaining results more believable.</p>

<h2>A one-page report</h2>
<p>Every quarter, put each use case on one line: what it is for, the before-and-after measures, the answer to the "would you mind" question, and a decision (scale, keep, fix or stop). Leadership can read it in five minutes, and it turns "is AI working?" from a feeling into a set of decisions.</p>]]></content:encoded></item>
<item><title>How to pick AI champions, and what to give them</title><link>https://www.ryanmorrisonai.com/essays/ai-champions</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/ai-champions</guid><pubDate>Tue, 23 Jun 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Champions are the most effective way to spread AI in a company, and the most often wasted. Pick for curiosity and trust, not job titles.</description><content:encoded><![CDATA[<p>Almost every AI adoption plan includes "a network of champions". It is a good idea. People learn new ways of working from colleagues they trust far more than from a training video or an email from IT. But in practice the champions are often chosen in a hurry, given nothing, and forgotten after the kick-off.</p>
<p>Here is how to do it so it works.</p>

<h2>Pick for curiosity and trust</h2>
<p>The best champions are not always the most technical people or the most senior. Look for three things:</p>
<ul>
  <li>They are already curious and experimenting, without being asked.</li>
  <li>Colleagues go to them for help with everyday questions.</li>
  <li>They know the actual work of their team in detail.</li>
</ul>
<p>Ask team leads "who do people in your team go to when something doesn't work?" That person is usually a better champion than the one who volunteers first.</p>

<h2>One per team, not one per company</h2>
<p>A champion works because they sit next to the work. One central "AI person" can't know how finance, sales and customer service each spend their day. Aim for one champion per team of ten to twenty people, and make sure the teams that do the most repetitive text work are covered first.</p>

<h2>Give them time, in writing</h2>
<p>This is where most programmes fail. Being a champion is added to a full job, so it is the first thing to go when things get busy. Agree a fixed amount of time, for example two to four hours a week, and get their manager to confirm it. If there is no time, there is no programme.</p>

<h2>Give them better access and a direct line</h2>
<p>Champions should get the approved tools first, a slightly bigger budget for trying things, and a direct line to whoever owns AI in the organisation. When they hit a question about data or rules, they need an answer within a day, not a ticket that sits for three weeks.</p>

<h2>Give them a job, not a title</h2>
<p>"Spread AI" is not a job. Give each champion something concrete for the first three months:</p>
<ul>
  <li>Find the three tasks in their team that waste the most time.</li>
  <li>Test AI on one of them with two or three colleagues.</li>
  <li>Share what worked and what didn't, with real examples, in a short session.</li>
</ul>

<h2>Bring them together</h2>
<p>Run a short call every two weeks where champions share what they tried. This is where the real learning happens: the finance champion's way of checking invoices turns out to be useful for procurement, and a prompt that failed in sales saves marketing a week. Keep it informal, keep it short, and write down the good ideas.</p>

<h2>Thank them visibly</h2>
<p>Mention their work in leadership updates, include it in their performance review, and let them present results themselves. Champions who feel seen stay champions. Champions who feel used become the first sceptics.</p>]]></content:encoded></item>
<item><title>Prompts are not a strategy: moving from chat to workflows</title><link>https://www.ryanmorrisonai.com/essays/from-prompts-to-workflows</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/from-prompts-to-workflows</guid><pubDate>Tue, 26 May 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Chat makes individuals faster. Workflows make the organisation faster. The step between them is smaller than it looks.</description><content:encoded><![CDATA[<p>Most organisations meet AI through a chat window. Someone types a question, gets a useful answer, and saves twenty minutes. Multiply that across a few hundred people and it adds up, and it is a perfectly good place to start.</p>
<p>But chat has a ceiling. Every person writes their own prompts, checks their own results and copies the output back into another system by hand. The knowledge of what works stays in individual chat histories. The organisation as a whole doesn't learn anything.</p>

<h2>What a workflow is</h2>
<p>A workflow is a task where the AI step is built into the way the work already flows. The input arrives from where it normally lives (an inbox, a form, a folder, a ticket system), the AI does one defined step with fixed instructions, and the result lands where people need it, ready for a check.</p>
<p>Take customer complaints. In chat, an employee copies a complaint into a tool and asks for a summary. In a workflow, every complaint that arrives is summarised, tagged by topic and urgency, and put into the right queue, with a draft reply attached. Same model, very different effect.</p>

<h2>How to spot a workflow in your chat use</h2>
<p>Look for prompts that people write again and again. If three people in a team ask the AI to do the same thing every week, with small variations, you have found a workflow waiting to be built. A quick survey asking "which prompts do you reuse?" usually turns up a handful of strong candidates.</p>
<p>Good candidates share three traits: the input is always the same kind of thing, the output has a clear shape, and someone can check the result quickly.</p>

<h2>The four parts to get right</h2>
<ul>
  <li><strong>The trigger:</strong> what starts it. A new email, a filled-in form, a file in a folder, a time of day.</li>
  <li><strong>The instructions:</strong> written once, tested on real examples, kept in one place and owned by one person.</li>
  <li><strong>The check:</strong> who looks at the result before it matters, and how they correct it.</li>
  <li><strong>The destination:</strong> where the output goes, so nobody has to copy and paste.</li>
</ul>
<p>Most of the effort goes into the first and last parts, the connections to your existing systems. The AI step itself is often the easy bit.</p>

<h2>You may not need developers</h2>
<p>Many of the tools you already pay for (your email and office suite, your CRM, your helpdesk) now include AI steps you can configure. Automation tools can connect them without writing code. Start there before commissioning anything custom, and only build when the off-the-shelf route hits a real limit.</p>

<h2>Where this leads</h2>
<p>Once you have a few workflows running, a pattern appears: some steps need judgement about what to do next, not just a fixed transformation. That is where agents come in, and it is a natural next step. But it only goes well when the basics are in place: clear instructions, a human check and a clear owner. Workflows are how you learn those basics on work that matters.</p>]]></content:encoded></item>
<item><title>The first 30 days of AI in a mid-size company</title><link>https://www.ryanmorrisonai.com/essays/first-30-days</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/first-30-days</guid><pubDate>Tue, 21 Apr 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Skip the licences and the hackathon. Start with an owner, a list of real tasks and three small tests.</description><content:encoded><![CDATA[<p>Most companies start their AI journey in the same place. Someone buys a batch of ChatGPT or Copilot licences, a few enthusiasts run a lunchtime demo, and three months later nobody can say what changed. Usage drops, the licences get questioned at budget time, and AI quietly becomes "something we tried".</p>
<p>The problem is rarely the technology. It is that nobody owned the question of what AI was supposed to do for the business. Here is the 30-day plan I would run in a company of 100 to 1,000 people that wants to get past that point.</p>

<h2>Week 1: one owner, one question</h2>
<p>Name one person who owns AI for the next three months. This person does not need to be technical. They need time (a day a week at least), access to the leadership team, and permission to say no to ideas.</p>
<p>Then agree on one question, in business language. "How do we use AI?" is too vague. "Where are we losing the most hours to repetitive work with text, documents or email?" is a question people can answer.</p>

<h2>Week 2: map the work, not the tools</h2>
<p>Talk to five or six teams for 45 minutes each. Ask what they do every week that is boring, slow, or error-prone. You are looking for tasks, not tools. Aim for a list of 20 to 30 concrete tasks, such as "summarising customer complaints for the weekly report" or "checking supplier invoices against purchase orders".</p>
<p>Score each task from 1 to 5 on four things:</p>
<table class="rubric">
  <thead><tr><th>Criterion</th><th>The question to ask</th></tr></thead>
  <tbody>
    <tr><td>Volume</td><td>How often does this happen, and how many people do it?</td></tr>
    <tr><td>Pain</td><td>How much do people dislike it, or how costly are the mistakes?</td></tr>
    <tr><td>Data ready</td><td>Is the input already digital and reachable, or stuck in someone's head or on paper?</td></tr>
    <tr><td>Risk</td><td>What happens if the AI gets it wrong? (Score low risk high.)</td></tr>
  </tbody>
</table>
<p>The tasks with the highest total are your shortlist. They are almost never the most exciting ideas, and that is fine. Exciting ideas make good keynotes. High-volume, low-risk tasks make good first projects.</p>

<h2>Week 3: three small tests</h2>
<p>Pick the top three and test each one with the people who actually do the work. Keep it small: one team, real examples, and existing tools where possible. Many first tests need nothing more than a well-written prompt, a shared folder of examples and an AI tool you already pay for.</p>
<p>Before you start, write down how the task works today: how long it takes, how often it goes wrong. Without that baseline you will end up arguing about opinions instead of numbers.</p>

<h2>Week 4: measure, decide, and write the rules down</h2>
<p>For each test, ask three questions. Did it save time? Did quality hold up? Would the team keep using it without being pushed? Kill anything that fails two of the three. Keep the rest and plan how to make them part of the normal way of working.</p>
<p>In the same week, write a one-page rule sheet: which tools are approved, what data may go into them, and who to ask when unsure. People are much more willing to experiment when they know where the edges are.</p>

<h2>What you should have on day 30</h2>
<ul>
  <li>One owner with a mandate</li>
  <li>A scored list of 20 to 30 real tasks</li>
  <li>Three tested use cases, with before-and-after numbers</li>
  <li>A decision on which ones to scale and which to stop</li>
  <li>A one-page rule sheet everyone can read in five minutes</li>
</ul>
<p>None of this needs a large budget or a data science team. It needs someone to own the question, and the discipline to start with the work instead of the technology. Once you have that, the bigger moves (agents, integrations, new products) stop being guesses.</p>]]></content:encoded></item>
<item><title>A one-page AI policy your people will actually read</title><link>https://www.ryanmorrisonai.com/essays/one-page-ai-policy</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/one-page-ai-policy</guid><pubDate>Tue, 24 Mar 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Your people are already using AI. A short, clear rule sheet does more for safety than a long document nobody opens.</description><content:encoded><![CDATA[<p>If you have not written any rules for AI yet, your people are not waiting. Surveys keep finding the same thing: a large share of employees use AI tools at work, and many of them don't tell their manager. They paste emails, reports and sometimes customer data into whatever tool is open in their browser.</p>
<p>The answer is not a ban. Bans push the use out of sight, where you can't guide it. The answer is a short policy that tells people what is fine, what is not, and who to ask. One page, written in plain language, that someone can read over a coffee.</p>

<h2>What goes on the page</h2>
<p>Five short sections are enough for most organisations.</p>
<table class="rubric">
  <thead><tr><th>Section</th><th>What it says</th></tr></thead>
  <tbody>
    <tr><td>Approved tools</td><td>Which AI tools are approved for work, and which account to use (the company one, not a personal one).</td></tr>
    <tr><td>Data</td><td>What may go in and what may not, with examples: public information yes, customer personal data no, confidential contracts only in the approved tool.</td></tr>
    <tr><td>Checking</td><td>You are responsible for anything you send or publish, whether a person or a machine wrote the first draft.</td></tr>
    <tr><td>Being open</td><td>When to say that AI was used, for example in client deliverables or decisions about people.</td></tr>
    <tr><td>Questions</td><td>One named person or channel for questions, and a promise of a quick answer.</td></tr>
  </tbody>
</table>

<h2>Use examples, not principles</h2>
<p>"Use AI responsibly" tells nobody anything. "You may use the approved assistant to draft a reply to a customer, but remove their name, address and order number first" tells people exactly what to do. Every rule on the page should come with one concrete example from your own work.</p>

<h2>Make the safe option the easy option</h2>
<p>Rules work best when the approved path is also the convenient one. If the company tool is slow, hard to reach or missing the features people need, they will keep using their personal accounts, whatever the policy says. Budget for a good approved tool before you write the rules.</p>

<h2>It is also part of the AI Act</h2>
<p>Under the EU AI Act, organisations that use AI systems are expected to take steps so that their staff have a sufficient level of AI literacy. A clear policy, together with a short training session, is a practical first step toward that. It won't cover everything the Act asks for, particularly if you use AI for decisions about people, but it shows you have started.</p>

<h2>Review it every quarter</h2>
<p>The tools change fast, so the page will go out of date. Put a review in the calendar every three months: which questions came in, which rules were unclear, which new tools people are asking for. Update the page and tell people what changed.</p>
<p>A policy that is short, current and full of examples gets read. A policy that is long, abstract and written once gets filed. Only one of them changes what people do on Monday.</p>]]></content:encoded></item>
<item><title>What I automated first, and what I should have automated first</title><link>https://www.ryanmorrisonai.com/essays/what-i-automated-first</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/what-i-automated-first</guid><pubDate>Tue, 17 Feb 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Notes from automating my own work with language models: the boring tasks paid off, the clever ones mostly didn't.</description><content:encoded><![CDATA[<p>Over the last few years I have automated a large part of my own working day with language models and agents. Not because I set out to, but because every time something worked, the next thing looked possible. Looking back, the order in which I did things was mostly wrong. Here is what I learned, in case it saves you some detours.</p>

<h2>I started with the impressive things</h2>
<p>My first projects were the ones that made good stories: a research assistant that would read everything about a market and write a report, an agent that would plan my week. They demoed well and I used them for about two weeks.</p>
<p>The problem was not the quality. It was that these were tasks I did occasionally, each time a little differently. Every run needed new instructions, new checking and new fixes. The automation saved less time than it cost to look after.</p>

<h2>The boring things paid off</h2>
<p>What actually changed my week were small, repetitive tasks that I did the same way every time:</p>
<ul>
  <li>Sorting incoming email into "needs me today", "needs me this week" and "for reference"</li>
  <li>Turning meeting notes into a short summary with decisions and next steps</li>
  <li>Writing first drafts of routine replies and updates in my own style</li>
  <li>Pulling the key points out of long documents before I read them properly</li>
</ul>
<p>None of these is impressive on its own. Together they gave me back hours every week, because they happen every day and they are always the same shape.</p>

<h2>Write the instructions down, like you would for a person</h2>
<p>The biggest improvement came when I stopped tweaking prompts in the moment and started writing instructions as if for a new colleague: what the task is for, what a good result looks like, two or three real examples, and what to do when unsure. Once written, they could be reused, improved and handed over.</p>
<p>This is also the most useful habit to bring into a company. Instructions that live in one person's chat history disappear when that person goes on holiday.</p>

<h2>Keep a human check where mistakes are expensive</h2>
<p>I let the machine sort, summarise and draft. I don't let it send, pay or promise anything without me looking first. That line has moved over time as I learned where the tools were reliable, but it has never disappeared. A short review step costs seconds; a wrong email to the wrong person costs much more.</p>

<h2>What I would do in your place</h2>
<p>Make a list of everything you do more than three times a week that involves reading or writing text. Pick the most boring item. Write proper instructions for it, test them on last week's real examples, and use the result for two weeks before touching anything else.</p>
<p>It will feel too small. It isn't. Small, repeated wins are what build the trust, and the skill, to take on the bigger things later.</p>]]></content:encoded></item>
<item><title>Five questions to ask before you buy an AI tool</title><link>https://www.ryanmorrisonai.com/essays/before-you-buy</link><guid isPermaLink="true">https://www.ryanmorrisonai.com/essays/before-you-buy</guid><pubDate>Tue, 20 Jan 2026 08:00:00 GMT</pubDate><dc:creator>Ryan Morrison</dc:creator><description>Most AI purchases go wrong on questions nobody asked during the demo.</description><content:encoded><![CDATA[<p>AI vendors have become very good at demos. In twenty minutes you see a tool summarise a contract, draft a proposal and answer questions about a policy document. It is genuinely impressive, and that is exactly the problem: the demo answers the question "can this work?" when the question you need answered is "will this work here?"</p>
<p>Before you sign anything, ask these five questions. None of them are technical, and the vendor can't answer most of them for you.</p>

<h2>1. Which task, for whom, how often?</h2>
<p>Name the task in one sentence, the team that does it, and how often it happens. "Better knowledge management" is not a task. "Answering the same twenty customer questions about delivery times, about 300 times a week, in customer service" is.</p>
<p>If you can't write that sentence, you are not buying a solution. You are buying a hope, and hopes are expensive on a per-seat licence.</p>

<h2>2. What does the tool need to see?</h2>
<p>Every AI tool is only as good as the information it can reach. Ask which systems it needs to connect to, who has to approve that access, and how long that approval usually takes in your organisation. In many companies the honest answer is "three months and a security review", which changes the business case more than any feature.</p>
<p>Also ask what happens to the data you put in. Is it used to train the vendor's models? Where is it stored? Can you delete it? A clear answer on one page is a good sign. A link to a forty-page terms document is not.</p>

<h2>3. What do we do today, and what does it cost?</h2>
<p>Write down how the task works now: how long it takes, how often it goes wrong, and what a mistake costs. This is the baseline you will compare against. Without it, every review turns into an exchange of impressions, and the loudest opinion wins.</p>
<p>It also protects you from a common trap: automating a task that only takes ten minutes a week. A tool can be brilliant and still not be worth the money.</p>

<h2>4. Who owns it after go-live?</h2>
<p>Someone has to keep the instructions up to date, answer questions from users, notice when quality drops, and decide when to stop. If that person doesn't exist, the tool will slowly decay. Prompts go stale, documents it relies on change, and people quietly go back to the old way.</p>
<p>Ask the vendor what this ownership looks like at their best customers. Then ask yourself who that person would be here, and whether they have the time.</p>

<h2>5. How do we stop?</h2>
<p>This is the question nobody asks. How long is the contract? Can you export what you built, such as instructions, templates and conversation history? What would switching to another tool cost in a year, when the market has moved again?</p>
<p>The AI market changes every few months. A tool that is the best choice today may not be next spring. Prefer short contracts, open formats and tools that work with what you already have.</p>

<h2>A simple rule</h2>
<p>If you can answer all five questions clearly, run a small paid pilot with one team and real work. If you can't answer the first one, stop there: no tool will answer it for you. The good news is that answering it costs nothing but a few conversations with the people who do the work.</p>]]></content:encoded></item>
</channel>
</rss>
