[{"data":1,"prerenderedAt":1854},["ShallowReactive",2],{"insights-list-insights_en":3},[4,241,484,725,914,1079,1288,1468,1710],{"id":5,"title":6,"author":7,"blobHue":11,"body":12,"category":215,"date":216,"description":217,"draft":218,"extension":219,"featured":220,"headline":221,"hue":222,"letter":223,"meta":224,"navigation":220,"path":225,"readMinutes":226,"related":227,"seo":231,"stem":232,"summary":233,"toc":234,"__hash__":240},"insights_en\u002Finsights\u002Fwhy-95-percent-of-ai-pilots-never-reach-production.md","Why 95% of AI pilots never reach production",{"name":8,"role":9,"bio":10},"André","Founder & CTO","André is the founder and CTO of WizardingCode. Eight years building the software companies run on, now putting agents into production.","blue",{"type":13,"value":14,"toc":206},"minimark",[15,24,27,32,37,43,47,50,73,77,80,167,171,193,197,200],[16,17,18,19,23],"p",{},"In 2025, MIT’s NANDA initiative looked at how companies were using generative AI and found that about 95% of enterprise pilots delivered no measurable impact on the P&L.",[20,21,22],"sup",{},"1"," Billions spent, demos applauded, and almost nothing changed in how the business actually ran.",[16,25,26],{},"We’ve seen the same pattern from the inside. Most of the companies that call us have already run a pilot. It worked in the demo. It never went live.",[28,29,31],"h2",{"id":30},"the-number-everyone-quotes","The number everyone quotes",[16,33,34,35],{},"The headline is easy to misread. It doesn’t say AI doesn’t work. It says pilots don’t turn into production. The same report found that projects built with specialised external partners reached deployment about twice as often as internal builds, and that the biggest returns came from unglamorous back-office work, not from customer-facing chatbots.",[20,36,22],{},[38,39,40],"blockquote",{},[16,41,42],{},"The model is rarely the problem. The plumbing is.",[28,44,46],{"id":45},"its-not-the-model","It’s not the model",[16,48,49],{},"When we look at why a pilot stalled, the answer is almost never “the AI wasn’t smart enough”. It is one of three things:",[51,52,53,61,67],"ol",{},[54,55,56,60],"li",{},[57,58,59],"strong",{},"No live data."," The pilot ran on an export. Connecting it to the real ERP, inbox or helpdesk was “phase two”, and phase two never came.",[54,62,63,66],{},[57,64,65],{},"No owner."," The innovation team built it. The team whose work it would take never asked for it, and never adopted it.",[54,68,69,72],{},[57,70,71],{},"No number."," Success was “a good demo”. Nobody agreed what had to move, so nobody could say it had worked.",[28,74,76],{"id":75},"four-things-the-5-do","Four things the 5% do",[16,78,79],{},"The projects that ship look different from day one. They connect to live systems in week one, not month six. The metric is agreed before a line of code is written, and it belongs to the team whose work changes. People approve the decisions that carry risk, so nobody has to trust the agent blindly. And there is a date: production in weeks, not a roadmap.",[81,82,83,98],"table",{},[84,85,86],"thead",{},[87,88,89,92,95],"tr",{},[90,91],"th",{},[90,93,94],{},"A PILOT",[90,96,97],{},"A SYSTEM IN PRODUCTION",[99,100,101,115,128,141,154],"tbody",{},[87,102,103,109,112],{},[104,105,106],"td",{},[57,107,108],{},"Data",[104,110,111],{},"A sample export",[104,113,114],{},"Your live systems",[87,116,117,122,125],{},[104,118,119],{},[57,120,121],{},"Success",[104,123,124],{},"A good demo",[104,126,127],{},"A number agreed on day one",[87,129,130,135,138],{},[104,131,132],{},[57,133,134],{},"Owner",[104,136,137],{},"The innovation team",[104,139,140],{},"The team whose work it takes",[87,142,143,148,151],{},[104,144,145],{},[57,146,147],{},"Humans",[104,149,150],{},"Watching",[104,152,153],{},"Approving what matters",[87,155,156,161,164],{},[104,157,158],{},[57,159,160],{},"Timeline",[104,162,163],{},"Open-ended",[104,165,166],{},"Weeks",[28,168,170],{"id":169},"a-checklist-before-you-start","A checklist before you start",[172,173,175],"prose-checklist",{"title":174},"Before you approve the next AI project, ask",[176,177,178,181,184,187,190],"ul",{},[54,179,180],{},"Which live system does it read and write, in week one?",[54,182,183],{},"Which number moves, and who owns it?",[54,185,186],{},"Which decisions stay with a person?",[54,188,189],{},"What happens on day 30?",[54,191,192],{},"Who owns the code, the prompts and the data?",[28,194,196],{"id":195},"what-this-means-for-you","What this means for you",[16,198,199],{},"If your last pilot is gathering dust, it probably wasn’t a bad idea. It was missing the boring parts. Start from one process that hurts, connect it to the real systems, agree the number and put a date on it. That’s the whole method. It isn’t magic; it’s engineering.",[201,202,203],"prose-footnotes",{},[16,204,205],{},"¹ MIT NANDA, “The GenAI Divide: State of AI in Business 2025”.",{"title":207,"searchDepth":208,"depth":208,"links":209},"",2,[210,211,212,213,214],{"id":30,"depth":208,"text":31},{"id":45,"depth":208,"text":46},{"id":75,"depth":208,"text":76},{"id":169,"depth":208,"text":170},{"id":195,"depth":208,"text":196},"strategy","2026-09-22","In 2025, MIT’s NANDA initiative looked at how companies were using generative AI and found that about 95% of enterprise pilots delivered no measurable impact on the P&L.1 Billions spent, demos applauded, and almost nothing changed in how the business actually ran.",false,"md",true,"Why 95% of AI pilots never reach ==production.==","magenta","95",{},"\u002Finsights\u002Fwhy-95-percent-of-ai-pilots-never-reach-production",7,[228,229,230],"what-an-agentic-os-is-and-what-it-isnt","every-agent-needs-a-judge","the-30-day-playbook-week-by-week",{"title":6,"description":217},"insights\u002Fwhy-95-percent-of-ai-pilots-never-reach-production","It’s rarely the model. It’s the data, the owner and the missing number. Here’s what the ones that ship do differently.",[235,236,237,238,239],{"id":30,"label":31},{"id":45,"label":46},{"id":75,"label":76},{"id":169,"label":170},{"id":195,"label":196},"9wJDwNO8uXgEwGK-cE1WYthNEN-J5Ic1e7kI_crM30g",{"id":242,"title":243,"author":244,"blobHue":245,"body":246,"category":215,"date":464,"description":250,"draft":218,"extension":219,"featured":218,"headline":465,"hue":466,"letter":467,"meta":468,"navigation":220,"path":469,"readMinutes":470,"related":471,"seo":474,"stem":475,"summary":476,"toc":477,"__hash__":483},"insights_en\u002Finsights\u002Fwhat-an-agentic-os-is-and-what-it-isnt.md","What an Agentic OS is, and what it isn’t",{"name":8,"role":9,"bio":10},null,{"type":13,"value":247,"toc":457},[248,251,254,258,261,264,268,274,306,312,318,324,390,395,399,405,411,417,421,424,427,431,454],[16,249,250],{},"Most people meet AI agents as a chat window. You ask, it answers, and then nothing happens. That is useful, but it isn’t how a business runs. A business runs on work that moves between systems, people and decisions, every day, whether anyone is watching or not.",[16,252,253],{},"An Agentic OS is what it takes to hand some of that work to agents and still sleep at night.",[28,255,257],{"id":256},"a-plain-definition","A plain definition",[16,259,260],{},"An Agentic OS is a team of AI agents trained on how your business works, running inside the tools you already use, following rules your people set. Each agent owns one job. They share what they know. They stop and ask when a decision is above their limits. And everything they do is visible on one screen.",[16,262,263],{},"The word “OS” is deliberate. An operating system is not an app you open. It is what sits underneath and keeps everything else running. The agents are the visible part; the layers around them are what make it safe.",[28,265,267],{"id":266},"the-four-layers","The four layers",[16,269,270,273],{},[57,271,272],{},"Agents"," do the work. We use five kinds, and most companies go live with two or three:",[51,275,276,282,288,294,300],{},[54,277,278,281],{},[57,279,280],{},"Responder"," answers customers, suppliers and colleagues, in your tone, from your data.",[54,283,284,287],{},[57,285,286],{},"Classifier"," reads what comes in, from emails to tickets to documents, and sends it to the right place.",[54,289,290,293],{},[57,291,292],{},"Scraper"," watches the sources you care about and brings back what changed.",[54,295,296,299],{},[57,297,298],{},"Orchestrator"," runs work that spans several systems: the refund, the CRM update, the courier booking.",[54,301,302,305],{},[57,303,304],{},"Analyst"," reads the numbers and has the report ready before the meeting.",[16,307,308,311],{},[57,309,310],{},"Shared memory"," is what the agents know about your business: policies, price lists, customer history, past decisions. Every agent reads from the same source, and it stays current as you work. No more “ask Maria, she knows”.",[16,313,314,317],{},[57,315,316],{},"Approval rules"," decide what an agent may do alone. They are written in plain language and set by your team. Above the limit, the agent pauses and a person decides, in the tools they already use.",[16,319,320,323],{},[57,321,322],{},"The command center"," shows every run, cost and result on one screen. You can replay any decision step by step, pause any agent, and follow the number you agreed on, every day.",[81,325,326,339],{},[84,327,328],{},[87,329,330,333,336],{},[90,331,332],{},"LAYER",[90,334,335],{},"WHAT IT ANSWERS",[90,337,338],{},"WITHOUT IT",[99,340,341,353,365,377],{},[87,342,343,347,350],{},[104,344,345],{},[57,346,272],{},[104,348,349],{},"Who does the work?",[104,351,352],{},"Nothing gets done",[87,354,355,359,362],{},[104,356,357],{},[57,358,310],{},[104,360,361],{},"What do they know?",[104,363,364],{},"Every agent guesses",[87,366,367,371,374],{},[104,368,369],{},[57,370,316],{},[104,372,373],{},"What can they do alone?",[104,375,376],{},"Nobody dares switch them on",[87,378,379,384,387],{},[104,380,381],{},[57,382,383],{},"Command center",[104,385,386],{},"What did they do, and did it work?",[104,388,389],{},"Nobody can prove it",[38,391,392],{},[16,393,394],{},"Agents alone are a demo. What makes them safe is everything around them.",[28,396,398],{"id":397},"what-it-isnt","What it isn’t",[16,400,401,404],{},[57,402,403],{},"It isn’t a chatbot."," A chatbot waits for a question. An Agentic OS works through a queue. It reads the inbox overnight, reconciles the stock file, drafts the replies, and leaves the three decisions that need a person at the top of the list in the morning.",[16,406,407,410],{},[57,408,409],{},"It isn’t a platform licence."," You don’t rent seats in someone else’s product and bend your process to fit it. The agents are built around your process, inside your systems, and you own the code, the prompts and the data.",[16,412,413,416],{},[57,414,415],{},"It isn’t a pilot."," A pilot runs on an export, for a demo, with nobody accountable for the result. An Agentic OS runs on live systems from the first week, against a number that the team whose work it takes has signed off. If it isn’t in production, it isn’t an OS yet.",[28,418,420],{"id":419},"where-to-start","Where to start",[16,422,423],{},"Nobody builds all four layers for the whole company at once. You start with one process that hurts, usually one with volume, clear rules and a number: the support inbox, the supplier files, the weekly report. You build the first squad of agents with its memory, its rules and its screen. Then you add a squad at a time.",[16,425,426],{},"The layers are what make the second squad cheaper than the first. The memory is already there. The rules have a format the team knows. The command center already shows the first squad’s numbers, so the next one is judged on the same screen, against the same kind of target.",[28,428,430],{"id":429},"how-to-tell-if-you-have-one","How to tell if you have one",[172,432,434],{"title":433},"You have an Agentic OS if the answer is yes to all of these",[176,435,436,439,442,445,448,451],{},[54,437,438],{},"Do the agents work inside the systems your team already uses?",[54,440,441],{},"Can you say, in one sentence, which decisions they may not take alone?",[54,443,444],{},"Do they all read from the same, current source of policies and history?",[54,446,447],{},"Can you open one screen and see what they did today, what it cost and what is waiting for you?",[54,449,450],{},"Has a number moved, and does the team that owns it agree?",[54,452,453],{},"Do you own the code, the prompts and the data?",[16,455,456],{},"If any answer is no, you have agents. That is a good start. The rest is the part that lets you rely on them.",{"title":207,"searchDepth":208,"depth":208,"links":458},[459,460,461,462,463],{"id":256,"depth":208,"text":257},{"id":266,"depth":208,"text":267},{"id":397,"depth":208,"text":398},{"id":419,"depth":208,"text":420},{"id":429,"depth":208,"text":430},"2026-09-15","What an Agentic OS is, and what it ==isn’t.==","violet","OS",{},"\u002Finsights\u002Fwhat-an-agentic-os-is-and-what-it-isnt",4,[472,229,473],"why-95-percent-of-ai-pilots-never-reach-production","what-agents-should-never-do-alone",{"title":243,"description":250},"insights\u002Fwhat-an-agentic-os-is-and-what-it-isnt","Not a chatbot, not a platform licence. A plain-language definition, with the four layers that make agents safe to run a business.",[478,479,480,481,482],{"id":256,"label":257},{"id":266,"label":267},{"id":397,"label":398},{"id":419,"label":420},{"id":429,"label":430},"O4B58r8Hq-cONZEYpn0ps8PFaK7Sr4eN71hJvwSQj0c",{"id":485,"title":486,"author":487,"blobHue":245,"body":488,"category":707,"date":708,"description":492,"draft":218,"extension":219,"featured":218,"headline":709,"hue":710,"letter":711,"meta":712,"navigation":220,"path":713,"readMinutes":470,"related":714,"seo":715,"stem":716,"summary":717,"toc":718,"__hash__":724},"insights_en\u002Finsights\u002Fwhat-agents-should-never-do-alone.md","What agents should never do alone",{"name":8,"role":9,"bio":10},{"type":13,"value":489,"toc":700},[490,493,496,500,506,512,518,521,525,528,531,534,538,541,561,564,568,571,679,682,687,691,694,697],[16,491,492],{},"The question we hear most from operations leads is not “can the agent do this?”. It is “what stops it doing something stupid?”. The honest answer is not a better prompt. It is a short list of rules, written by the people who own the risk, that the agent cannot talk its way around.",[16,494,495],{},"Writing those rules is less work than it sounds. Almost everything that should stay with a person falls into three kinds of decision.",[28,497,499],{"id":498},"three-kinds-of-decisions","Three kinds of decisions",[16,501,502,505],{},[57,503,504],{},"Money."," Refunds, credits, discounts, prices, payments. Anything that moves money in or out of the business, or changes what a customer pays.",[16,507,508,511],{},[57,509,510],{},"Promises."," A delivery date, a replacement, an exception to the policy, anything that commits the company to a customer or a supplier. A wrong answer is fixable. A wrong promise has to be honoured or broken.",[16,513,514,517],{},[57,515,516],{},"Anything you can’t undo."," Deleting records, sending to your whole customer base, cancelling or changing an order that is already moving. If the mistake cannot be reversed by clicking something, a person decides.",[16,519,520],{},"Everything else, like reading, classifying, drafting, looking up, summarising and routing, is usually safe for an agent to do alone, provided it is visible afterwards.",[28,522,524],{"id":523},"draft-then-apply","Draft, then apply",[16,526,527],{},"The most useful distinction in approval rules is between preparing an action and executing it. An agent can do all the work of a refund: find the order, check the policy, calculate the amount, write the reply. Then it stops, and a person applies it with one tap.",[16,529,530],{},"This keeps most of the time saving and almost none of the risk. The person is not doing the work. They are checking a finished proposal, with the order, the policy and the amount on one screen.",[16,532,533],{},"Over time, some drafts earn the right to apply themselves. When a person has approved the same kind of action enough times without changes, you can move it below the threshold. That is a decision for the team that owns it, not for the agent.",[28,535,537],{"id":536},"thresholds-not-feelings","Thresholds, not feelings",[16,539,540],{},"“Ask me when it’s important” is not a rule. An agent cannot know what feels important to you. A rule needs a number, a category or a list:",[176,542,543,549,555],{},[54,544,545,548],{},[57,546,547],{},"A number."," Refunds up to a limit are applied; above it, they go to the ops lead.",[54,550,551,554],{},[57,552,553],{},"A category."," Any change to a contract goes to legal, whatever the value.",[54,556,557,560],{},[57,558,559],{},"A list."," New suppliers, key accounts and anything to the press always go to a person.",[16,562,563],{},"Every rule names who decides and where they are asked: in Slack, Teams or email, wherever that person already works. A rule that sends approvals to a dashboard nobody opens is a rule that stops the business.",[28,565,567],{"id":566},"a-worked-example","A worked example",[16,569,570],{},"This is the shape of a rule set for an online store’s support and commerce agents. The limits are yours to set; the structure is what matters.",[81,572,573,586],{},[84,574,575],{},[87,576,577,580,583],{},[90,578,579],{},"ACTION",[90,581,582],{},"AGENT ALONE",[90,584,585],{},"GOES TO A PERSON",[99,587,588,601,614,627,640,653,666],{},[87,589,590,595,598],{},[104,591,592],{},[57,593,594],{},"Order status, tracking, policy questions",[104,596,597],{},"Answers",[104,599,600],{},"Never, unless the customer asks for one",[87,602,603,608,611],{},[104,604,605],{},[57,606,607],{},"Refund",[104,609,610],{},"Drafts every one, applies small ones",[104,612,613],{},"Above €500 in this example, the ops lead",[87,615,616,621,624],{},[104,617,618],{},[57,619,620],{},"Voucher or discount",[104,622,623],{},"Proposes, inside the margin floor",[104,625,626],{},"Anything outside the agreed limits",[87,628,629,634,637],{},[104,630,631],{},[57,632,633],{},"Price change",[104,635,636],{},"Never writes a price",[104,638,639],{},"The pricing owner, after the margin check",[87,641,642,647,650],{},[104,643,644],{},[57,645,646],{},"Delivery date",[104,648,649],{},"Quotes what the courier data says",[104,651,652],{},"Any exception or guarantee",[87,654,655,660,663],{},[104,656,657],{},[57,658,659],{},"Order change or cancellation",[104,661,662],{},"Drafts",[104,664,665],{},"Always applied by a person",[87,667,668,673,676],{},[104,669,670],{},[57,671,672],{},"Campaign to the whole base",[104,674,675],{},"Prepares and simulates",[104,677,678],{},"Always sent by a person",[16,680,681],{},"Start with the right-hand column. The right-hand column is the list of things your team has decided to keep. Everything to its left is work they no longer have to do.",[38,683,684],{},[16,685,686],{},"If you can’t undo it, a person does it.",[28,688,690],{"id":689},"the-veto-that-isnt-a-model","The veto that isn’t a model",[16,692,693],{},"Some rules are too important to leave to a language model, even one that is being checked. Margin is the clearest case.",[16,695,696],{},"In the marketplace system we built, no agent ever writes a price. Agents can propose a discount or a clearance offer, but every proposal passes through a margin guardrail first. The guardrail is plain code, not a model. It calculates the floor for that product after returns, fees and shipping, and anything below it is vetoed. There is nothing to persuade and no prompt to get wrong.",[16,698,699],{},"That is the pattern we use wherever a rule can be expressed as arithmetic or a lookup. The model proposes, deterministic code checks, and a person approves what the rules say a person approves. Each layer does the thing it is good at, and none of them is asked to be the last line of defence alone.",{"title":207,"searchDepth":208,"depth":208,"links":701},[702,703,704,705,706],{"id":498,"depth":208,"text":499},{"id":523,"depth":208,"text":524},{"id":536,"depth":208,"text":537},{"id":566,"depth":208,"text":567},{"id":689,"depth":208,"text":690},"playbooks","2026-09-08","What agents should never do ==alone.==","sun","!",{},"\u002Finsights\u002Fwhat-agents-should-never-do-alone",[229,230,228],{"title":486,"description":492},"insights\u002Fwhat-agents-should-never-do-alone","A practical way to write approval rules: money, promises and anything you can’t undo.",[719,720,721,722,723],{"id":498,"label":499},{"id":523,"label":524},{"id":536,"label":537},{"id":566,"label":567},{"id":689,"label":690},"mWnCbu3ZKOI9CJrpjxbakDDnQwkJuenNM8EI7HSjdMA",{"id":726,"title":727,"author":728,"blobHue":245,"body":729,"category":896,"date":897,"description":733,"draft":218,"extension":219,"featured":218,"headline":898,"hue":11,"letter":899,"meta":900,"navigation":220,"path":901,"readMinutes":470,"related":902,"seo":904,"stem":905,"summary":906,"toc":907,"__hash__":913},"insights_en\u002Finsights\u002Fevery-agent-needs-a-judge.md","Every agent needs a judge",{"name":8,"role":9,"bio":10},{"type":13,"value":730,"toc":889},[731,734,737,741,744,747,750,754,757,825,828,831,835,838,858,861,865,868,871,874,879,883,886],[16,732,733],{},"A language model is very good at sounding right. That is the problem. An answer with a wrong refund amount reads exactly as confidently as one with the right amount, and a person skimming a queue of drafts will not catch it every time.",[16,735,736],{},"So we don’t ask people to catch it. In the marketplace system we built, 90 agents work across eight departments, and every team has a judge: 9 judges in all. A judge is a second model whose only job is to check another agent’s answer before anyone relies on it.",[28,738,740],{"id":739},"why-a-second-model","Why a second model",[16,742,743],{},"Asking the same model to review its own work helps less than you would think. It tends to agree with itself, and it shares the blind spots that produced the mistake in the first place.",[16,745,746],{},"That is why our judges run on a different model family from the agents they check. Different training, different habits, different failure modes. When two unrelated models agree that an answer is grounded and within policy, that means much more than one model agreeing with itself twice.",[16,748,749],{},"A judge also sees the answer differently. The agent was trying to be helpful. The judge is only trying to find what is wrong. Giving it a narrow brief, and nothing else to do, is what makes it useful.",[28,751,753],{"id":752},"block-or-grade-later","Block or grade later",[16,755,756],{},"There are two ways to put a judge in the path, and choosing between them is the main design decision.",[81,758,759,771],{},[84,760,761],{},[87,762,763,765,768],{},[90,764],{},[90,766,767],{},"BLOCK AND REPAIR",[90,769,770],{},"SHIP, THEN GRADE",[99,772,773,786,799,812],{},[87,774,775,780,783],{},[104,776,777],{},[57,778,779],{},"When the judge runs",[104,781,782],{},"Before the answer leaves",[104,784,785],{},"After it has gone out",[87,787,788,793,796],{},[104,789,790],{},[57,791,792],{},"If it fails",[104,794,795],{},"Sent back once with the reasons, then shipped with reservations",[104,797,798],{},"Flagged, counted, fed into the next lessons",[87,800,801,806,809],{},[104,802,803],{},[57,804,805],{},"Cost to the user",[104,807,808],{},"Some extra seconds",[104,810,811],{},"None",[87,813,814,819,822],{},[104,815,816],{},[57,817,818],{},"Use it for",[104,820,821],{},"Anything a customer sees, anything with money",[104,823,824],{},"High volume, low risk, easy to correct",[16,826,827],{},"In the blocking path, a failed answer is not simply rejected. It goes back to the agent once, with the judge’s reasons, and the agent gets a chance to repair it. If the repaired answer still fails, it goes out marked with the judge’s reservations, or, where the rules require it, to a person. Nothing leaves without a verdict.",[16,829,830],{},"In the grading path, the answer goes out and the judge scores it afterwards. The scores show where an agent drifts, and they feed the lessons the system learns overnight. Those lessons are not applied on their own: a person approves each one before agents use it.",[28,832,834],{"id":833},"what-a-judge-checks","What a judge checks",[16,836,837],{},"A judge with a vague brief (“is this a good answer?”) is expensive noise. Ours check specific things, and each check can fail on its own:",[172,839,841],{"title":840},"What a judge checks, every time",[176,842,843,846,849,852,855],{},[54,844,845],{},"Can every number in the answer be traced to data the agent actually read in this run?",[54,847,848],{},"Does it contradict a policy in the shared memory?",[54,850,851],{},"Does it answer the question that was asked, completely?",[54,853,854],{},"Does it propose an action the approval rules do not allow?",[54,856,857],{},"Is the tone right for who will read it?",[16,859,860],{},"The first check is the one that matters most. A refund amount, a delivery date or a stock figure that does not appear in any tool result is treated as invented, however plausible it looks.",[28,862,864],{"id":863},"what-it-costs-and-when-to-skip-it","What it costs, and when to skip it",[16,866,867],{},"A judge is another model call on every answer. It adds tokens, and in the blocking path it adds seconds. For a customer reply that is cheap insurance. For some work, it is waste.",[16,869,870],{},"We skip the judge, or move it to grading later, when the output is low stakes, reversible and cheap to check. An internal tag suggestion that a person sees anyway, or a draft that someone always edits before sending, does not need a second model standing in front of it.",[16,872,873],{},"We also never ask a model to check what code can check. Whether a price is above its floor, whether a date is in the future or whether an order exists are deterministic questions. They get deterministic answers, which are faster and do not have opinions.",[38,875,876],{},[16,877,878],{},"A judge is for judgement. If a rule can be written as code, write it as code.",[28,880,882],{"id":881},"fail-loudly-not-silently","Fail loudly, not silently",[16,884,885],{},"Judges fail too. They time out, or their provider is down for a few minutes. The worst thing a system can do then is pretend the check happened.",[16,887,888],{},"When a judge cannot run, the answer is shown with a visible “not validated” mark, and the person reading it decides whether to rely on it. A silent pass is how a system loses the trust of the people who work with it, and that trust is much harder to rebuild than a timeout is to fix.",{"title":207,"searchDepth":208,"depth":208,"links":890},[891,892,893,894,895],{"id":739,"depth":208,"text":740},{"id":752,"depth":208,"text":753},{"id":833,"depth":208,"text":834},{"id":863,"depth":208,"text":864},{"id":881,"depth":208,"text":882},"engineering","2026-09-01","Every agent needs a ==judge.==","J",{},"\u002Finsights\u002Fevery-agent-needs-a-judge",[473,228,903],"testing-a-campaign-on-customers-who-dont-exist",{"title":727,"description":733},"insights\u002Fevery-agent-needs-a-judge","Why we put a second model in front of every answer, and when it isn’t worth the cost.",[908,909,910,911,912],{"id":739,"label":740},{"id":752,"label":753},{"id":833,"label":834},{"id":863,"label":864},{"id":881,"label":882},"XzL5diXOtcDslArv3P74XXRkQd_4qZoVSP7GUPDaC8E",{"id":915,"title":916,"author":917,"blobHue":245,"body":918,"category":1060,"date":1061,"description":922,"draft":218,"extension":219,"featured":218,"headline":1062,"hue":1063,"letter":1064,"meta":1065,"navigation":220,"path":1066,"readMinutes":470,"related":1067,"seo":1069,"stem":1070,"summary":1071,"toc":1072,"__hash__":1078},"insights_en\u002Finsights\u002Ftesting-a-campaign-on-customers-who-dont-exist.md","Testing a campaign on customers who don’t exist",{"name":8,"role":9,"bio":10},{"type":13,"value":919,"toc":1053},[920,923,926,930,933,936,940,943,963,966,970,973,976,979,984,988,991,1037,1040,1043,1047,1050],[16,921,922],{},"Our founder also runs a fashion marketplace. We built its Agentic OS the way we build one for any client, inside its own CRM: 90 agents in eight departments, each team checked by a judge. There, a campaign to the whole base is the kind of action we keep with a person. Once it is out, it is out. A tone that lands badly or an offer that confuses reaches everyone at once, and the unsubscribes do not come back.",[16,924,925],{},"So before a send, we test it on customers who don’t exist. These are notes on how the synthetic customers are built, what they are good for, and where we have learned not to trust them.",[28,927,929],{"id":928},"why-simulate-a-send","Why simulate a send",[16,931,932],{},"A\u002FB tests are the honest way to compare messages, but they test on real people. Half your audience gets the weaker version, and you learn after the fact. For a weekly campaign in four languages, across email, SMS and push, there are more variants than there is audience to test them on.",[16,934,935],{},"A simulation is a cheap first filter. It does not replace the real result. It helps decide which variants deserve to reach real people at all, and it catches the obvious mistakes before anyone sees them.",[28,937,939],{"id":938},"how-a-synthetic-customer-is-built","How a synthetic customer is built",[16,941,942],{},"We do not invent personas from a marketing brief. Each one is built from real buyers.",[51,944,945,951,957],{},[54,946,947,950],{},[57,948,949],{},"Split the base into strata."," Groups of buyers that behave alike: how often they buy, what they buy, how they respond to discounts, which channels they use, which language they read.",[54,952,953,956],{},[57,954,955],{},"Describe each stratum from its data."," Order history, categories, return behaviour, past campaign responses. The description is written from what these buyers did, not from what we imagine they want.",[54,958,959,962],{},[57,960,961],{},"Give each persona a voice."," A persona model reads the description and answers as that kind of buyer would, when shown a subject line, a message and an offer.",[16,964,965],{},"Every stratum gets its own persona, so the simulation reflects the mix of your actual audience, not an average customer who does not exist either.",[28,967,969],{"id":968},"two-models-one-prediction","Two models, one prediction",[16,971,972],{},"Each persona is played by two different persona models, working as an ensemble. For every variant of a campaign, both predict whether that buyer would open, click, convert or unsubscribe, and each gives the objection it would have: “the discount isn’t worth the shipping”, “I bought this last week”, “this doesn’t sound like you”.",[16,974,975],{},"Using two models matters for the same reason judges run on a different model family from the agents they check. When the two agree, the signal is stronger. When they disagree, the disagreement is itself useful: it usually points at a message that could be read two ways.",[16,977,978],{},"The predictions are calibrated against real results, stratum by stratum: what the personas predicted is compared with what actual buyers did, and the simulation is adjusted. A simulation that is never checked against reality drifts into fiction.",[38,980,981],{},[16,982,983],{},"A synthetic customer is a hypothesis about real ones. It has to be tested against them.",[28,985,987],{"id":986},"what-they-get-right-and-wrong","What they get right, and wrong",[16,989,990],{},"After enough sends, the pattern is clear.",[81,992,993,1003],{},[84,994,995],{},[87,996,997,1000],{},[90,998,999],{},"GOOD AT",[90,1001,1002],{},"BAD AT",[99,1004,1005,1013,1021,1029],{},[87,1006,1007,1010],{},[104,1008,1009],{},"Ranking variants against each other",[104,1011,1012],{},"Predicting absolute open or click rates",[87,1014,1015,1018],{},[104,1016,1017],{},"Catching an obvious tone miss",[104,1019,1020],{},"Anything genuinely new to the audience",[87,1022,1023,1026],{},[104,1024,1025],{},"Spotting a confusing offer or subject line",[104,1027,1028],{},"Price sensitivity in the moment",[87,1030,1031,1034],{},[104,1032,1033],{},"Surfacing the objection nobody wrote down",[104,1035,1036],{},"Events outside the data, like the weather or the news",[16,1038,1039],{},"Ranking is what they are built for: which variant is likely to do better, not by how much.",[16,1041,1042],{},"The absolute rates are the weak point. Models are not good at saying exactly how many people will open something, and we do not use them that way. Novelty is another: a type of campaign the audience has never seen has no history for the personas to draw on. And price sensitivity is hard. What a buyer says about a discount and what they do on a Friday night are not the same.",[28,1044,1046],{"id":1045},"the-rule-that-blocks-a-launch","The rule that blocks a launch",[16,1048,1049],{},"The simulation is part of the launch. A newsletter or SMS launch without a recent simulation behind it is flagged, and it can be blocked until a new one has run. If the content has changed since the last run, the simulation is out of date.",[16,1051,1052],{},"The personas do not decide whether a campaign goes out. A person does, with the predictions, the objections and the disagreements on one screen. The synthetic customers only make sure that nobody sends a campaign to real ones without asking the question first.",{"title":207,"searchDepth":208,"depth":208,"links":1054},[1055,1056,1057,1058,1059],{"id":928,"depth":208,"text":929},{"id":938,"depth":208,"text":939},{"id":968,"depth":208,"text":969},{"id":986,"depth":208,"text":987},{"id":1045,"depth":208,"text":1046},"case-notes","2026-08-25","Testing a campaign on customers who don’t ==exist.==","aqua","S",{},"\u002Finsights\u002Ftesting-a-campaign-on-customers-who-dont-exist",[1068,229,473],"cart-recovery-with-agents-notes-from-our-own-store",{"title":916,"description":922},"insights\u002Ftesting-a-campaign-on-customers-who-dont-exist","Synthetic personas built from real buyers: what they predict well, and what they get wrong.",[1073,1074,1075,1076,1077],{"id":928,"label":929},{"id":938,"label":939},{"id":968,"label":969},{"id":986,"label":987},{"id":1045,"label":1046},"FOFAwqeMsuY8POjF0OqFXdwG1XrLGeDBiMZwOLpdldQ",{"id":1080,"title":1081,"author":1082,"blobHue":245,"body":1083,"category":707,"date":1270,"description":1087,"draft":218,"extension":219,"featured":218,"headline":1271,"hue":222,"letter":1272,"meta":1273,"navigation":220,"path":1274,"readMinutes":470,"related":1275,"seo":1277,"stem":1278,"summary":1279,"toc":1280,"__hash__":1287},"insights_en\u002Finsights\u002Fthe-30-day-playbook-week-by-week.md","The 30-day playbook, week by week",{"name":8,"role":9,"bio":10},{"type":13,"value":1084,"toc":1262},[1085,1088,1091,1095,1098,1101,1105,1108,1134,1137,1141,1144,1147,1150,1154,1157,1160,1242,1246,1249,1252,1256,1259],[16,1086,1087],{},"Thirty days sounds fast for putting AI agents into production. It is fast. It is also the reason most of our projects ship: a fixed date forces every decision that a pilot would postpone to be taken in the first week.",[16,1089,1090],{},"This is the playbook we follow. Every step ends with something written down and someone’s name next to it.",[28,1092,1094],{"id":1093},"week-0-diagnose","Week 0: diagnose",[16,1096,1097],{},"It starts with a free call. If there is a fit, we spend the next week inside your operations, most of it with the people who do the work. We sit in on the inbox, the report, the supplier file. We want to see where the hours go and where the money leaks, not how the process is described in a slide.",[16,1099,1100],{},"The artefact is a map of your operations with the value of every agent we would build, and a fixed price. The sign-off is yours: which process goes first. We push for one with volume, clear rules and a number someone already cares about.",[28,1102,1104],{"id":1103},"week-1-map","Week 1: map",[16,1106,1107],{},"This is the week most pilots skip, and the week that decides whether yours ships.",[51,1109,1110,1116,1122,1128],{},[54,1111,1112,1115],{},[57,1113,1114],{},"Access."," Real credentials to the real systems: the CRM, the helpdesk, the ERP, the inbox. Not an export.",[54,1117,1118,1121],{},[57,1119,1120],{},"Real cases."," A set of past emails, tickets or files, with what a good answer looked like. They become the tests.",[54,1123,1124,1127],{},[57,1125,1126],{},"One metric, signed off."," First response time, time to spot a new listing, partner hours on admin. One number, owned by the team whose work changes, signed by the person who owns it.",[54,1129,1130,1133],{},[57,1131,1132],{},"First approval rules."," What the agents may do alone, and what always goes to a person, written in plain language.",[16,1135,1136],{},"If week 1 ends without access and a signed metric, we say so, because it is cheaper for both sides than finding out on day 29.",[28,1138,1140],{"id":1139},"week-2-build","Week 2: build",[16,1142,1143],{},"Now we build the agents, the rules, the shared memory and the command center, wired into your tools. The agents live where your team already works; nobody gets a new app to learn.",[16,1145,1146],{},"The real cases from week 1 become evals: automated tests that run every answer against what a good answer looked like. Customer-facing answers get a judge. Tone is agreed with the people who own it, often in a single working session with a pile of past replies.",[16,1148,1149],{},"The sign-off at the end of the week is a walkthrough with the team lead, on their own cases.",[28,1151,1153],{"id":1152},"week-3-prove","Week 3: prove",[16,1155,1156],{},"First the evals, on real cases the agents have never seen. Then shadow mode on live work: the agents draft, people send. Every edit a person makes is a signal, and we read them daily.",[16,1158,1159],{},"This is also where the handoff rules are tuned with the team. Which cases go straight to a person, at what threshold, through which channel. The team should be able to change a rule themselves, and in week 3 they practise doing it.",[81,1161,1162,1175],{},[84,1163,1164],{},[87,1165,1166,1169,1172],{},[90,1167,1168],{},"STEP",[90,1170,1171],{},"ARTEFACT",[90,1173,1174],{},"WHO SIGNS",[99,1176,1177,1190,1203,1216,1229],{},[87,1178,1179,1184,1187],{},[104,1180,1181],{},[57,1182,1183],{},"Week 0",[104,1185,1186],{},"Operations map, value per agent, fixed price",[104,1188,1189],{},"You: which process first",[87,1191,1192,1197,1200],{},[104,1193,1194],{},[57,1195,1196],{},"Week 1",[104,1198,1199],{},"Access, real cases, first approval rules",[104,1201,1202],{},"The owner of the metric",[87,1204,1205,1210,1213],{},[104,1206,1207],{},[57,1208,1209],{},"Week 2",[104,1211,1212],{},"Agents, memory, rules, command center, evals",[104,1214,1215],{},"The team lead, after a walkthrough",[87,1217,1218,1223,1226],{},[104,1219,1220],{},[57,1221,1222],{},"Week 3",[104,1224,1225],{},"Eval results, shadow-mode log, tuned handoffs",[104,1227,1228],{},"The team lead: ready to go live",[87,1230,1231,1236,1239],{},[104,1232,1233],{},[57,1234,1235],{},"Day 30",[104,1237,1238],{},"Production, monitoring, trained team",[104,1240,1241],{},"Both of us, against the metric",[28,1243,1245],{"id":1244},"day-30-live","Day 30: live",[16,1247,1248],{},"On day 30 the agents are in production, on live work, monitored from the command center. The team is trained on it: how to read a run, how to approve, how to pause an agent, how to change a rule.",[16,1250,1251],{},"Production means the agents answer, route or update on their own inside the rules, and the number from week 1 is tracked every day from then on. It does not mean “available on request” or “ready for phase two”.",[28,1253,1255],{"id":1254},"why-the-date-holds","Why the date holds",[16,1257,1258],{},"Three things keep the date honest. The scope is one process, not the company. The metric is agreed before a line of code, so nobody can move the goalposts later. And we put our money on it: live in 30 days, or your money back.",[16,1260,1261],{},"After day 30 we run the fleet: monitoring, tuning and, when you are ready, the next squad. On our largest project that became one new squad a month, each built on the memory and rules of the ones before it.",{"title":207,"searchDepth":208,"depth":208,"links":1263},[1264,1265,1266,1267,1268,1269],{"id":1093,"depth":208,"text":1094},{"id":1103,"depth":208,"text":1104},{"id":1139,"depth":208,"text":1140},{"id":1152,"depth":208,"text":1153},{"id":1244,"depth":208,"text":1245},{"id":1254,"depth":208,"text":1255},"2026-08-18","The 30-day playbook, week by ==week.==","30",{},"\u002Finsights\u002Fthe-30-day-playbook-week-by-week",[472,473,1276],"build-buy-or-both",{"title":1081,"description":1087},"insights\u002Fthe-30-day-playbook-week-by-week","From the first call to production: the exact steps, artefacts and sign-offs we use.",[1281,1282,1283,1284,1285,1286],{"id":1093,"label":1094},{"id":1103,"label":1104},{"id":1139,"label":1140},{"id":1152,"label":1153},{"id":1244,"label":1245},{"id":1254,"label":1255},"rT6HBQ-x40ciVZH-qHpcDH1wLuGv79MQR2E3ffHF_dY",{"id":1289,"title":1290,"author":1291,"blobHue":245,"body":1292,"category":1060,"date":1451,"description":1296,"draft":218,"extension":219,"featured":218,"headline":1452,"hue":710,"letter":1453,"meta":1454,"navigation":220,"path":1455,"readMinutes":470,"related":1456,"seo":1458,"stem":1459,"summary":1460,"toc":1461,"__hash__":1467},"insights_en\u002Finsights\u002Fcart-recovery-with-agents-notes-from-our-own-store.md","Cart recovery with agents: notes from our own store",{"name":8,"role":9,"bio":10},{"type":13,"value":1293,"toc":1444},[1294,1297,1300,1304,1307,1310,1314,1317,1320,1340,1343,1347,1350,1353,1357,1360,1363,1431,1435,1438,1441],[16,1295,1296],{},"Our founder also runs a fashion marketplace. We built its Agentic OS the way we build one for any client, inside its own CRM: 90 agents in eight departments, each team checked by a judge. Cart recovery is one of the places where the data was already there and the problem was obvious. People fill a cart, and then life happens.",[16,1298,1299],{},"These are notes on what changed, what we measure, and what did not work. We have left out revenue and percentages on purpose. They depend on the season, the catalogue and the traffic, and a number without that context would tell you less than the method does.",[28,1301,1303],{"id":1302},"where-we-started","Where we started",[16,1305,1306],{},"Before agents, cart recovery was like most stores’: a fixed sequence. One email after a few hours, another a day later, sometimes a discount in the third. The same words for everyone, in the language of the store, whatever the customer had been looking at.",[16,1308,1309],{},"It worked. It just spoke to nobody in particular, and got the results of that.",[28,1311,1313],{"id":1312},"one-person-one-message","One person, one message",[16,1315,1316],{},"The first real change was writing each message for the person who left the cart. Every night, the system refreshes a profile for every buyer: a persona, a churn risk and a next best action. The recovery agent reads that profile, the cart and the customer’s history before it writes anything.",[16,1318,1319],{},"That changes three things.",[51,1321,1322,1328,1334],{},[54,1323,1324,1327],{},[57,1325,1326],{},"Timing."," A customer who usually buys late in the evening is not reminded at nine in the morning. Someone who comes back within the hour is left alone for a while.",[54,1329,1330,1333],{},[57,1331,1332],{},"Channel."," Email, SMS, push or WhatsApp, depending on what that person has agreed to and has actually responded to before. Not every channel, and not the loudest one.",[54,1335,1336,1339],{},[57,1337,1338],{},"Tone."," A first-time visitor and a customer with years of orders do not get the same message. The first needs reassurance about sizes and returns. The second needs a short reminder, in their language, about the pieces they chose.",[16,1341,1342],{},"A judge checks the messages: the facts about the product and the policy must match the catalogue and the shared memory, and the tone must fit the brand.",[28,1344,1346],{"id":1345},"vouchers-only-inside-limits","Vouchers, only inside limits",[16,1348,1349],{},"The easiest way to recover a cart is to give money away. That is also the easiest way to lose margin on a sale that might have happened anyway.",[16,1351,1352],{},"So the recovery agent can propose a voucher, but only inside limits the commerce team has set, and only after a margin guardrail has checked it. The guardrail is plain code, not a model. It works out the floor for each product after returns, fees and shipping, and anything below it is vetoed. No agent ever writes a price. A voucher is a tool for a specific person, not a default.",[28,1354,1356],{"id":1355},"a-proposal-every-week-a-verdict-28-days-later","A proposal every week, a verdict 28 days later",[16,1358,1359],{},"Once a week, the system proposes how to recover more: change the timing for a segment, try a different channel, adjust the wording, tighten or loosen a voucher rule. A person reviews each proposal and decides whether to apply it.",[16,1361,1362],{},"Then comes the part we care about most. Every applied change is measured 28 days later, against what was happening before. Not the next morning, when a change always looks good or bad by chance. Twenty-eight days is long enough to see whether customers actually came back and bought, and whether the orders held once returns had come in.",[81,1364,1365,1377],{},[84,1366,1367],{},[87,1368,1369,1371,1374],{},[90,1370,1168],{},[90,1372,1373],{},"WHO",[90,1375,1376],{},"WHEN",[99,1378,1379,1392,1405,1418],{},[87,1380,1381,1386,1389],{},[104,1382,1383],{},[57,1384,1385],{},"Propose a change",[104,1387,1388],{},"The recovery agent",[104,1390,1391],{},"Every week",[87,1393,1394,1399,1402],{},[104,1395,1396],{},[57,1397,1398],{},"Approve or reject",[104,1400,1401],{},"A person on the commerce team",[104,1403,1404],{},"Before anything changes",[87,1406,1407,1412,1415],{},[104,1408,1409],{},[57,1410,1411],{},"Measure the result",[104,1413,1414],{},"The system, against the baseline",[104,1416,1417],{},"28 days after it goes live",[87,1419,1420,1425,1428],{},[104,1421,1422],{},[57,1423,1424],{},"Keep, change or drop",[104,1426,1427],{},"The same person",[104,1429,1430],{},"After the measurement",[28,1432,1434],{"id":1433},"what-didnt-work","What didn’t work",[16,1436,1437],{},"Not every proposal earns its place. Some changes made no measurable difference after 28 days and were dropped. That is the point of measuring: without it, they would still be running, and we would still believe in them.",[16,1439,1440],{},"A profile is only as good as the history behind it. For a first-time visitor there is little to go on, and the message is closer to the generic one than we would like. We treat that as a limit to be honest about, not something to paper over with guesses.",[16,1442,1443],{},"And a better message does not fix a bad reason for leaving. When a cart is abandoned because of shipping costs, a sold-out size or a confusing returns policy, the right move is to fix the store. The recovery agent’s most useful output is sometimes not a message at all, but a pattern for someone to look at.",{"title":207,"searchDepth":208,"depth":208,"links":1445},[1446,1447,1448,1449,1450],{"id":1302,"depth":208,"text":1303},{"id":1312,"depth":208,"text":1313},{"id":1345,"depth":208,"text":1346},{"id":1355,"depth":208,"text":1356},{"id":1433,"depth":208,"text":1434},"2026-08-11","Cart recovery with agents: notes from our own ==store.==","R",{},"\u002Finsights\u002Fcart-recovery-with-agents-notes-from-our-own-store",[903,473,1457],"getting-cited-by-chatgpt-and-claude",{"title":1290,"description":1296},"insights\u002Fcart-recovery-with-agents-notes-from-our-own-store","Timing, channel and tone. What changed when the message was written for one person.",[1462,1463,1464,1465,1466],{"id":1302,"label":1303},{"id":1312,"label":1313},{"id":1345,"label":1346},{"id":1355,"label":1356},{"id":1433,"label":1434},"KH3Tvoi1FzHdhgydQOS66KG25jDv5Q_dCBEpLT85Ztc",{"id":1469,"title":1470,"author":1471,"blobHue":245,"body":1472,"category":215,"date":1694,"description":1476,"draft":218,"extension":219,"featured":218,"headline":1695,"hue":466,"letter":1696,"meta":1697,"navigation":220,"path":1698,"readMinutes":470,"related":1699,"seo":1700,"stem":1701,"summary":1702,"toc":1703,"__hash__":1709},"insights_en\u002Finsights\u002Fbuild-buy-or-both.md","Build, buy or both?",{"name":8,"role":9,"bio":10},{"type":13,"value":1473,"toc":1687},[1474,1477,1480,1484,1487,1494,1501,1507,1510,1514,1517,1543,1546,1550,1553,1559,1565,1642,1645,1650,1654,1657,1660,1664,1684],[16,1475,1476],{},"Every company we talk to already pays for some AI. A writing assistant here, a meeting summariser there, a chatbot on the website. The question is rarely whether to use AI. It is which work deserves more than a subscription.",[16,1478,1479],{},"The honest answer is that most of it doesn’t. And the part that does usually needs both: things you buy and things you build.",[28,1481,1483],{"id":1482},"when-buying-is-enough","When buying is enough",[16,1485,1486],{},"An off-the-shelf tool is the right choice when three things are true.",[16,1488,1489,1490,1493],{},"The work is ",[57,1491,1492],{},"generic",". Summarising a meeting, rewriting an email, translating a document, drafting a first version of a job ad. Every company does it roughly the same way, so a product built for everyone fits you well.",[16,1495,1496,1497,1500],{},"It needs ",[57,1498,1499],{},"no live data",". The input is whatever the person pastes in, and the output goes back to that person. Nothing has to be read from your ERP or written into your CRM.",[16,1502,1503,1506],{},[57,1504,1505],{},"Nobody owns an outcome."," It makes individuals faster, but no team has signed up to move a number with it. If it gets worse next month, someone switches tools and nothing breaks.",[16,1508,1509],{},"For this kind of work, building your own is a waste of money. Buy the tool, set some sensible rules on what data goes into it, and move on.",[28,1511,1513],{"id":1512},"when-a-process-deserves-its-own-agent","When a process deserves its own agent",[16,1515,1516],{},"The picture changes when the work runs through your systems and your rules. A process deserves its own agent when:",[51,1518,1519,1525,1531,1537],{},[54,1520,1521,1524],{},[57,1522,1523],{},"It lives in your systems."," The agent has to read the order, the ticket history, the supplier file or the price list, and write back to them.",[54,1526,1527,1530],{},[57,1528,1529],{},"There is a number."," First response time, hours spent on admin, stock left unsold. A team owns it and wants it to move.",[54,1532,1533,1536],{},[57,1534,1535],{},"Some decisions need a person."," Refunds, promises, anything that cannot be undone. Those need approval rules that match how your company works, not how a vendor imagined companies work.",[54,1538,1539,1542],{},[57,1540,1541],{},"It is specific to you."," Your policies, your exceptions, your tone, the thing only Maria knows. A generic tool does not know any of it and you cannot teach it properly.",[16,1544,1545],{},"No product sold to everyone can do all four for you. It doesn’t know your systems, it can’t own your number, and its rules are someone else’s.",[28,1547,1549],{"id":1548},"the-both-pattern","The both pattern",[16,1551,1552],{},"In practice, the answer for most serious processes is both.",[16,1554,1555,1558],{},[57,1556,1557],{},"You buy the commodity."," The language models themselves, from the big model providers. Hosting, email delivery, search, the plumbing. Nobody should build their own model to answer supplier emails.",[16,1560,1561,1564],{},[57,1562,1563],{},"You build what is yours."," The agents, the approval rules and the shared memory. That is where your process, your policies and your judgement live, and it is the part that makes the difference between a demo and a system you can rely on.",[81,1566,1567,1579],{},[84,1568,1569],{},[87,1570,1571,1573,1576],{},[90,1572],{},[90,1574,1575],{},"BUY",[90,1577,1578],{},"BUILD",[99,1580,1581,1594,1606,1618,1630],{},[87,1582,1583,1588,1591],{},[104,1584,1585],{},[57,1586,1587],{},"Models",[104,1589,1590],{},"From the model providers",[104,1592,1593],{},"Never",[87,1595,1596,1601,1604],{},[104,1597,1598],{},[57,1599,1600],{},"Infrastructure",[104,1602,1603],{},"Hosting, email, storage",[104,1605,1593],{},[87,1607,1608,1612,1615],{},[104,1609,1610],{},[57,1611,272],{},[104,1613,1614],{},"Rarely a fit",[104,1616,1617],{},"Built around your process",[87,1619,1620,1624,1627],{},[104,1621,1622],{},[57,1623,316],{},[104,1625,1626],{},"Generic settings",[104,1628,1629],{},"Written with your team",[87,1631,1632,1636,1639],{},[104,1633,1634],{},[57,1635,310],{},[104,1637,1638],{},"Empty until you fill it",[104,1640,1641],{},"Your policies, history and decisions",[16,1643,1644],{},"This split also protects you. Models improve and prices change every few months. When the agents, rules and memory are yours, swapping the model underneath is an engineering task, not a migration. The marketplace system we built uses models from several providers at once, each chosen for the job, and the system does not depend on any single one.",[38,1646,1647],{},[16,1648,1649],{},"Buy the model. Build the part that knows your business.",[28,1651,1653],{"id":1652},"who-owns-what","Who owns what",[16,1655,1656],{},"Whatever you build, ask who owns it at the end. In our projects the answer is simple. The agents run in your environment and the memory is in your systems. You own the code, the prompts and the data, so you can run them without us.",[16,1658,1659],{},"That matters more than it seems. A process that runs on an agent you do not own is a process someone else can reprice, change or switch off.",[28,1661,1663],{"id":1662},"a-quick-test","A quick test",[172,1665,1667],{"title":1666},"Before you buy another AI tool, ask",[176,1668,1669,1672,1675,1678,1681],{},[54,1670,1671],{},"Does it need to read or write your live systems?",[54,1673,1674],{},"Is there a number that a team is expected to move with it?",[54,1676,1677],{},"Are there decisions in it that must stay with a person?",[54,1679,1680],{},"Does it depend on policies, exceptions or tone that are specific to you?",[54,1682,1683],{},"If the vendor changed its price or product tomorrow, would a process stop?",[16,1685,1686],{},"If every answer is no, buy it. If two or more are yes, the process probably deserves its own agent, built on bought models, with rules and memory that belong to you.",{"title":207,"searchDepth":208,"depth":208,"links":1688},[1689,1690,1691,1692,1693],{"id":1482,"depth":208,"text":1483},{"id":1512,"depth":208,"text":1513},{"id":1548,"depth":208,"text":1549},{"id":1652,"depth":208,"text":1653},{"id":1662,"depth":208,"text":1663},"2026-08-04","Build, buy or ==both?==","?",{},"\u002Finsights\u002Fbuild-buy-or-both",[228,472,230],{"title":1470,"description":1476},"insights\u002Fbuild-buy-or-both","When an off-the-shelf AI tool is enough, and when your process deserves its own agent.",[1704,1705,1706,1707,1708],{"id":1482,"label":1483},{"id":1512,"label":1513},{"id":1548,"label":1549},{"id":1652,"label":1653},{"id":1662,"label":1663},"XoWrr_D4QkTedkkrnrgccPyPPGBK8-ZFnFJfkxogvLc",{"id":1711,"title":1712,"author":1713,"blobHue":245,"body":1714,"category":896,"date":1838,"description":1718,"draft":218,"extension":219,"featured":218,"headline":1839,"hue":11,"letter":1840,"meta":1841,"navigation":220,"path":1842,"readMinutes":470,"related":1843,"seo":1844,"stem":1845,"summary":1846,"toc":1847,"__hash__":1853},"insights_en\u002Finsights\u002Fgetting-cited-by-chatgpt-and-claude.md","Getting cited by ChatGPT and Claude",{"name":8,"role":9,"bio":10},{"type":13,"value":1715,"toc":1831},[1716,1719,1722,1726,1729,1732,1735,1739,1777,1782,1786,1789,1796,1799,1805,1808,1812,1815,1818,1821,1825,1828],[16,1717,1718],{},"More and more product research starts with a question typed into ChatGPT, Claude or another assistant, not a search box. “Which linen shirt holds up after washing?” “Is this blender loud?” The assistant answers in a paragraph, and sometimes it names a source.",[16,1720,1721],{},"Generative-engine optimisation is the work of making your pages the source it names. It is not a new trick. Most of it is writing product pages the way a careful shop assistant would talk: plainly, specifically, and without hiding the facts.",[28,1723,1725],{"id":1724},"how-an-answer-engine-reads-your-page","How an answer engine reads your page",[16,1727,1728],{},"A search engine ranks pages. An answer engine assembles an answer from pieces of pages, and quotes the pieces it trusts. That changes what a good page looks like.",[16,1730,1731],{},"It prefers sentences that can be lifted out whole and still make sense. It prefers facts stated once, clearly, over facts implied by adjectives. It prefers sources that name things consistently, so it can be sure the product on your page is the same one a review talks about. And it reads structure: headings, lists, tables and structured data all help it find the piece it needs.",[16,1733,1734],{},"None of this replaces classic SEO. A page that is not indexed, fast and linked will not be read at all. Think of it as a floor and a ceiling: search engines are the floor, answer engines are the ceiling.",[28,1736,1738],{"id":1737},"six-habits-of-pages-that-get-quoted","Six habits of pages that get quoted",[51,1740,1741,1747,1753,1759,1765,1771],{},[54,1742,1743,1746],{},[57,1744,1745],{},"State facts plainly."," Material, dimensions, weight, care, compatibility, warranty, origin. In sentences, with units, not only in a spec table image.",[54,1748,1749,1752],{},[57,1750,1751],{},"Write answer-shaped paragraphs."," Open a section with the direct answer in one sentence, then explain. The first sentence is the one most likely to be quoted.",[54,1754,1755,1758],{},[57,1756,1757],{},"Add real questions and answers."," Use the questions customers actually ask, taken from support tickets and reviews, each with a short, complete answer.",[54,1760,1761,1764],{},[57,1762,1763],{},"Build entity pages."," A page for the brand and a page for each category, saying what they are, who they are for and how they differ. Answer engines use them to understand where a product fits.",[54,1766,1767,1770],{},[57,1768,1769],{},"Use structured data."," Product, offer, review and FAQ markup that matches the visible text exactly. Markup that contradicts the page does more harm than none.",[54,1772,1773,1776],{},[57,1774,1775],{},"Keep names consistent."," The same product name, brand name and category name on the page, in the feed, in the markup and in every language.",[38,1778,1779],{},[16,1780,1781],{},"Write the sentence you would want an assistant to quote about you.",[28,1783,1785],{"id":1784},"a-product-paragraph-before-and-after","A product paragraph, before and after",[16,1787,1788],{},"This is the kind of paragraph most catalogues are full of:",[16,1790,1791,1795],{},[1792,1793,1794],"em",{},"Before:"," “Discover effortless style with our iconic shirt. Crafted from premium fabrics for all-day comfort, it’s the perfect addition to any wardrobe. Elevate your look this season.”",[16,1797,1798],{},"There is nothing an assistant can quote. No material, no fit, no care, nothing that answers a question.",[16,1800,1801,1804],{},[1792,1802,1803],{},"After:"," “This shirt is made of 100% European linen, with a relaxed fit and a straight hem designed to be worn untucked. It is machine washable at 30 °C and softens with each wash. It runs true to size; if you are between sizes, choose the smaller one for a closer fit.”",[16,1806,1807],{},"The second version is not less appealing. It is more useful, and every sentence answers a question someone has asked. (The shirt and its details are an illustration, not a real listing.)",[28,1809,1811],{"id":1810},"doing-it-for-thousands-of-products","Doing it for thousands of products",[16,1813,1814],{},"Writing like this by hand works for ten products. It does not work for a catalogue that changes every week.",[16,1816,1817],{},"In the marketplace system we built, a product studio writes product, category, brand and page copy with SEO and AI-search content, in four languages. It works from the supplier data and the catalogue, not from imagination. Brand and category entity pages are written the same way, so the names match everywhere.",[16,1819,1820],{},"A person still reviews what goes live. The studio’s job is to make the careful version the default, not to publish without anyone looking.",[28,1822,1824],{"id":1823},"what-you-cant-control","What you can’t control",[16,1826,1827],{},"Nobody decides what an assistant cites. Models change, sources shift and the same question can get different answers on different days. Anyone who promises you a position in ChatGPT is promising something they do not control.",[16,1829,1830],{},"What you can control is whether your pages are worth quoting. Check your key products the simple way: ask the assistants the questions your customers ask, and see whose sentences come back. If they are not yours, the fix is usually on the page.",{"title":207,"searchDepth":208,"depth":208,"links":1832},[1833,1834,1835,1836,1837],{"id":1724,"depth":208,"text":1725},{"id":1737,"depth":208,"text":1738},{"id":1784,"depth":208,"text":1785},{"id":1810,"depth":208,"text":1811},{"id":1823,"depth":208,"text":1824},"2026-07-28","Getting cited by ChatGPT and ==Claude.==","G",{},"\u002Finsights\u002Fgetting-cited-by-chatgpt-and-claude",[1068,1276,229],{"title":1712,"description":1718},"insights\u002Fgetting-cited-by-chatgpt-and-claude","Generative-engine optimisation for product pages: facts, Q&A pairs and entity pages.",[1848,1849,1850,1851,1852],{"id":1724,"label":1725},{"id":1737,"label":1738},{"id":1784,"label":1785},{"id":1810,"label":1811},{"id":1823,"label":1824},"_uD17NODHcZyxNIpzQ8lLewrVkFutpXnh-B7MI0qwuQ",1790618463653]