Service 04
Most companies have purchased AI licenses. A few people use them for drafting reports, someone in the org built a dashboard or two, and the automation stalled at the first number that came back wrong. Nobody mapped which work should move to software, nobody built it into the stack, and nothing was checking the output before a person read it. That’s the integration, and it’s the part we do together: map the work, build it with gates, hand it over running.
Every agency now claims to be AI-native and almost none will show you a running system, because a prompt library in a shared doc isn’t one. The honest numbers are humbling: a fifth of mid-market marketing teams have an agent in production, most pilots die at the integration, and the usual cause isn’t the model. It’s that nobody unified the data first, so the agent confidently answered from four systems with four versions of the truth.
So the order is fixed. The warehouse before the agent. The gate before the output. The mechanical half of the work, reporting, analysis, audits, first drafts, moved to software that checks its own numbers against the source before a person reads them, and a person before anything reaches a customer. The judgment half stays with your people, who now have the time for it. Built in your stack, documented, trained into your team, left running.
What your team actually does in a week, sorted: the mechanical half software should own, the judgment half that stays with people, and the risk tier of each, so low-stakes work runs on its own and anything touching a customer or a budget gets a human approving. The map is the plan. Most “AI strategies” skip it and buy a tool.
Sales, ads, catalog and inventory landed in one warehouse with a real catalog layer, so every downstream number has one source and one definition. An agent on top of fragmented reporting is four systems with a chat interface.
Agents that read the warehouse the way a growth marketer would, two windows, control groups, baseline checks, and write the analysis every morning, on a schedule, before anyone’s coffee. Frontier models for the reasoning, a local model where the data can’t leave the building.
The part nobody builds. Every output checked against the data underneath it; if a claim isn’t supported, it doesn’t reach a human, let alone a customer. Golden sets, escalation rules, and intervention and override rates tracked, because a system you have to correct often is a system that isn’t ready.
Judgment encoded as software: listing and tracking audits that scrape live pages, score them against a standard and return annotated reports; bid and budget logic that produces recommendations on a schedule rather than hunches in a meeting; forecasting that carries a stated confidence.
Weekly decks and scorecards generated from the warehouse instead of rebuilt by hand, with the numbers that matter given real estate rather than a footnote. Then the part that makes it integration rather than a demo: your team trained, the logic documented, the system running without me.
No agent gets built before the warehouse it reads from is correct. Most AI marketing projects fail here and blame the model.
New systems run alongside your existing process, producing output nobody acts on, until the outputs agree. Cut-over is a decision your team makes on evidence, not a launch date on a slide.
Grounding checks, golden sets, escalation rules. An automated system that can’t say “I don’t know” is a liability wearing a dashboard. When it can, it’s yours: in your stack, documented, running.
These aren’t a black box rented back to you. They get built in your stack, documented, and left behind running, whether the engagement continues or not. Bring me the Monday report that gets rebuilt by hand and I’ll tell you what it would take to never build it again.
What artificial intelligence actually is and how it works, machine learning and deep learning without the mystery, generative AI, AI agents and what makes them dangerous, how to use AI in marketing without it embarrassing you, the risks that are real, ChatGPT and the other tools, and the truth about AI detectors. Written by someone who builds and runs these systems for a living and has had to put gates on them.
Artificial intelligence is software that does things which, when a person does them, we'd say require intelligence: recognizing what's in a picture, understanding and producing language, predicting from data, driving a car. "Artificial" means built by people; the intelligence is the capability, not a mind. AI stands for artificial intelligence, and "A.I.," "a i" and "inteligencia artificial" are the same words.
Here's the deflating truth your vendor's deck leaves out: nearly everything called AI in 2026 is machine learning, a model that learned patterns from a very large amount of data and applies them to new inputs. The AI answering questions in ChatGPT, Claude and Gemini is a language model trained on text. The AI recommending products, flagging fraud and setting your ad bids is a model trained on transactions and behavior. The AI in your phone's camera finding faces is a model trained on images. None of them understands anything in the way you do. They learned the statistics of their domain well enough to produce useful output, and for a shocking number of tasks that turns out to be enough.
Examples you already use: search ranking, the spam filter, the arrival time in a navigation app, voice assistants, translation, every recommendation feed on every platform, autocomplete, fraud detection on your card, radiology, and the chat assistants that write, summarize and answer. What AI does well is anything with a large pile of examples and a clear score. What it does badly is anything requiring facts it hasn't seen, judgment outside its training, or someone to be accountable. The reason a business should care has nothing to do with intelligence and everything to do with cost: a task that took a person twenty minutes now takes a model two seconds, and the economics of a lot of work change when that's true.
AI is pattern-matching at a scale that looks like thinking. Price it as labor, not as magic.
AI works by training a model on data and then using the trained model to make predictions about new data. Training means showing it an enormous number of examples and nudging its internal parameters, often billions of numbers, until its outputs get closer to the right answers. Using it, inference, means feeding it a new input and reading what comes out. The model is the pile of numbers training produced. That's all it is.
For the language models behind the chat assistants, the training task is embarrassingly simple: predict the next piece of text. The model reads a staggering quantity of writing and learns to guess what comes next, token by token. Doing that well turns out to require learning grammar, facts, reasoning patterns and style, which is why a model trained only to continue text can answer questions, write code and summarize a contract. A second stage, using human feedback on which responses are better, turns the raw text-continuer into an assistant that follows instructions and declines the obviously bad requests. When you type a question, the model generates its reply one token at a time, each one picked from a probability distribution over possible next tokens. That's why the same question gets different answers, and it's why the model will state a falsehood with perfect confidence: it's producing probable text, not retrieving verified facts. There is no fact-checker inside. There is a very good guesser.
The models are neural networks: layers of simple math units whose connections hold the learned parameters, arranged in architectures suited to the data (transformers for language, convolutional networks for images). They run on GPUs and similar chips, both for the months of training and for the fraction of a second each answer takes. The practical consequence for anyone using AI in a business: the model is a fixed artifact. It knows what it was trained on, up to a cutoff date. Anything newer or private has to be handed to it in the prompt or fetched by a retrieval system at the moment of the question. It does not know your numbers. It was never told.
A language model is a next-word guesser trained on most of the internet. It is confident by construction, and confidence is not evidence.
Machine learning is the branch of AI where the system learns to do a task from data instead of from rules a programmer wrote. Instead of coding "if the email contains these words, it's spam," you hand the system ten thousand emails labeled spam or not and it finds the patterns itself. Machine learning is AI; it's the part of AI that produced essentially every practical result of the last twenty years, and when someone says AI in 2026 they mean a machine learning system. "Learning machine" is the same thing with the words swapped.
Three broad kinds. Supervised learning: the training data carries the right answers (labeled images, past sales with outcomes), and the model learns to predict them. Unsupervised: no answers; the model finds structure, clusters, anomalies on its own. Reinforcement learning: the model learns by acting in an environment and getting rewards, which is how game-playing systems were built and part of how chat assistants get refined.
Deep learning is machine learning with neural networks stacked many layers deep, hence the name. Deep networks learn their own features from raw data, pixels, audio, text, instead of needing a person to specify what to look for, which is why they own image recognition, speech and language. Every large language model is a deep learning model. The price is appetite: deep learning needs enormous data and enormous compute, which is why the frontier of it lives in a handful of labs with a handful of balance sheets.
Learning machine learning starts with the math it stands on, linear algebra, probability and enough calculus to know what a gradient is; then Python with its data libraries (NumPy, pandas, scikit-learn); then a real course, Andrew Ng's Machine Learning Specialization on Coursera for the foundations, fast.ai's Practical Deep Learning for the code-first route, Google's and Kaggle's free material for applied practice. The step everyone skips is working on messy real data for a problem they care about, which is the only place the skill actually forms. The courses teach you the vocabulary. The dataset teaches you the job.
Machine learning is learning from examples instead of rules. Deep learning is the same with more layers and a bigger power bill.
Generative AI is AI that makes new content, text, images, code, audio, video, instead of classifying or predicting something about content that already exists. ChatGPT writing your email, Midjourney drawing from a sentence, Copilot finishing your code, a model cloning a voice: all generative. The difference is in the output. A model that reads a review and labels it positive isn't generative. The one that writes the review is.
So a generative AI task is any task whose result is new content: drafting, rewriting, summarizing, translating, writing code, producing an image or a video or a track, generating synthetic data. Fraud detection, demand forecasting, spam filtering and recommendations are machine learning and not generative; they produce a label, a number or a ranking. The certification question that asks which task is generative is looking for the one that creates.
Generative models work by learning the structure of their medium well enough to produce plausible new examples. Language models go token by token; image models usually start from noise and refine toward something that matches the prompt. The outputs are fluent and often excellent, and they are unverified. A generative model is optimizing for plausibility, not truth, so every factual claim it produces needs checking by something that isn't it. In a business that's the entire design problem, and it's not a hard one to state: generative AI is phenomenal at the drafting, transforming and summarizing work that fills most knowledge jobs, and it needs a review gate between its output and anything that reaches a customer, a regulator, or a decision. The companies that built the gate are shipping. The companies that trusted the output are in the news.
Generative AI makes things. It does not check things. Build the checker or become the cautionary tale.
An AI agent is a system that uses a model to chase a goal across many steps, choosing actions, using tools, reading the results, deciding what to do next, instead of giving one answer to one prompt. A chat assistant answers a question. An agent gets "find out why conversion dropped last week and draft the report," and it runs the queries, reads the data, tests a hypothesis, writes the document, and only comes to you if it gets stuck. Agentic AI is the term for systems built this way. "Agentic" means having agency: acting toward a goal rather than reacting to an input.
What makes an agent is a loop. The model gets the goal and the current state; it picks an action (call a tool, run a query, write a file, ask a question); the action runs and the result goes back to the model; repeat until the goal is reached or a limit is hit. Around the loop sit the tools it's allowed to use, some memory of what it's done, and, in any system that touches real data or real customers, the rules about what it may do without a human saying yes.
Building an agent in 2026: pick a model (the frontier models from Anthropic, OpenAI and Google all support tool use natively), define the tools as functions with a schema, write the loop or use a framework that provides it, give it a clear goal and clearer limits, and then test it against the ways it fails. And it will fail in ways a script never would: misreading a result, taking a confident wrong turn, doing something it was permitted to do that no sane person would have done. The loop is maybe forty lines. The engineering that matters is the gates: what the agent must verify before it acts, what it may never do, what a person reviews before anything leaves the building. The pattern that holds up in production is an agent that runs the analysis every morning, checks its own numbers against the source data before it believes them, and routes anything it can't ground to a human instead of to the report. That's the shape of the agent fleets the marketing systems service builds, and the grounding gate is the part that took longest to get right.
The loop is easy. The gates are the product. An agent without a grounding check is an intern with root access.
Using AI, for most people, means a chat assistant (ChatGPT, Claude, Gemini, Copilot) through its site or app: say what you want in plain language, read the result, push back, refine. The skill is in the asking. Give it the context it needs, the format you want, an example of good, the constraints; ask for one thing at a time; and check anything that's a fact. Past chat, AI shows up as features inside software you already use and, for businesses, as APIs that put a model inside your own systems.
Using AI in marketing works where the task is high-volume, text- or data-heavy, and has a review step built in. What's reliable in 2026: drafting and varying ad copy, product descriptions and email at volume for a person to pick from; summarizing reviews, support tickets and research into themes; tagging and classifying content, queries and customers at scale; first-draft analysis of performance data, with every figure checked against the source; turning one brief into a dozen creative concepts to test; and automating the mechanical half of reporting, pulling the numbers, building the deck, flagging what moved. What doesn't work is letting a model publish or decide without a gate: the statistic it invented for the blog post, the support reply that promised a refund your policy doesn't allow, the "insight" that was a date-range artifact. The rule I run on: AI produces drafts and candidates, people produce decisions, and software enforces the line between them. Not a policy document. Software.
"Creating an AI" for almost any business does not mean training a model. That costs tens of millions and is done by a handful of labs. It means building on an existing model through its API: your data by retrieval, your tools by function calling, your rules by instruction, your quality bar by testing. A company's "own AI" is nearly always a frontier model plus the company's data and guardrails, and that is the right way to build one. The companies announcing they've "built their own AI" have, in nearly every case, built a prompt.
AI makes drafts. People make decisions. If software isn't enforcing that line, the line doesn't exist.
Learning AI depends on what you want to be able to do, and most people pick the wrong target. To use it well, which is what nearly everyone actually needs, the fastest route is daily practice with a frontier assistant on your real work, plus a short course on prompting and on where the models fail. A few weeks of that is worth more than any certificate. To build with it, learn Python, then the vendors' API docs (Anthropic, OpenAI and Google all publish excellent quickstarts), then retrieval, tool use and evaluation, by building something small you'll actually use. To understand it at the level of the models, take the machine learning path: math, Python, a foundations course, deep learning, then the papers.
What's free and genuinely good: the model vendors' documentation and cookbooks; Andrew Ng's courses and DeepLearning.AI's short applied courses; fast.ai; Google's Machine Learning Crash Course; Kaggle's Learn modules; and Andrej Karpathy's lectures on building a language model from scratch, which are the clearest explanation of what's inside one that exists in any medium. Learning artificial intelligence as a field is a computer science degree with a specialization. Learning it as a professional skill is the first two routes above, and the first one starts today, with the tool you already have open in another tab.
Learn to use it by using it on real work. Learn to build with it by building something you need. Certificates are for the people who sell certificates.
AI is dangerous in specific, documented ways, and it is not going to take over the world in the way the headline means. The real risks in 2026 are concrete: models that state falsehoods fluently and get believed; systems automating decisions about credit, hiring, policing and benefits with biases learned from the past; convincing fake text, images, audio and video used for fraud, harassment and political manipulation; the displacement of writing, support, design, coding and analysis work faster than people retrain; a few companies holding most of the capability; enormous energy and water use; and the ordinary failure of software handed too much authority too early. All of that is happening now, and all of it is the subject of regulation, lawsuits and real technical work.
Is AI good or bad? Wrong question. It's a capability, and both are true at once. It has already improved medical imaging, drug discovery, translation, accessibility for blind and deaf people, scientific research and the productivity of a great deal of ordinary work. It has also powered fraud at industrial scale, buried the internet in content nobody wrote, and been deployed in places it wasn't ready for by people who didn't check. Why AI is bad in any given case almost always comes down to one thing: used without checks, for a purpose where the errors have victims.
AI ethics is the study of how these systems should be built and used: fairness, transparency, accountability, privacy, consent for the data they learn from, honesty about what they can and can't do. AI governance is the practical version: the laws (the EU AI Act, in force in stages from 2025, is the broadest), the standards, and the internal policies that decide which uses a company permits, who's accountable when it fails, how outputs get reviewed, how incidents get handled. For a business deploying AI, governance isn't abstract. It's the written answer to "what is this system allowed to do, who checks it, and what happens when it's wrong," and a company that can't answer that in one page shouldn't be deploying.
The "take over the world" question is about the longer-term concern that systems more capable than people might pursue goals we didn't intend. Serious researchers take it seriously, current systems can't do it, and the work on it is called alignment. The risks above are the ones that will reach your business this year.
AI's real danger is an unchecked system in a job where mistakes hurt people. Governance is one page: what it may do, who checks it, what happens when it's wrong.
ChatGPT is OpenAI's chat assistant, the product that turned large language models from a research curiosity into something your parents use when it launched in November 2022. You type, it answers in plain language; it writes, summarizes, explains, translates, reads uploaded files, generates images, browses the web and runs code. "Chat GPT," "chat gbt" and "AI chat bot" nearly always mean ChatGPT, though every major lab now has an equivalent and several are better at specific things.
Is ChatGPT free? Yes, with a free account at chatgpt.com or in the app on iOS and Android, with limits on how many messages reach the most capable models and on file uploads and image generation. ChatGPT Plus is $20 a month in 2026 and raises the limits and adds the newest models; Pro is $200 a month for heavy professional use; Team and Enterprise are for companies. Getting ChatGPT is creating an account; there's nothing to install on the web. Using it well is the same skill as any assistant: specific asks, context, examples, and checking the facts it states with such confidence.
Is anything better than ChatGPT? Depends what you're doing, and in 2026 the honest answer is that the leading assistants are close and swap places with each release. Claude from Anthropic is widely preferred for writing, long documents, analysis and code. Gemini from Google is wired through Google's products and strong on images and search-grounded answers. Microsoft Copilot wraps OpenAI's models inside Office and Windows. Perplexity is built around answering with cited web sources. The best AI app for you is the one that fits your work and your tools, and most professionals who use these seriously keep two. For AI news, go to the labs themselves (OpenAI, Anthropic, Google DeepMind) for what's real, and The Verge, Ars Technica and MIT Technology Review for coverage that isn't a reposted press release.
ChatGPT is free, Plus is $20, and whether something's "better" depends on the job. Keep two and check both.
AI detectors cannot reliably tell whether text was written by AI, and no tool that claims certainty deserves it. The detectors, GPTZero, Originality.ai, Copyleaks, Turnitin's AI indicator and the rest, measure statistical properties of writing: how predictable each word is given the ones before it (perplexity), how much that predictability varies across sentences (burstiness), and patterns a classifier picked up from human and AI samples. Human writing tends to be less predictable and more uneven; AI writing, especially from older models at default settings, tends to be smooth and uniform. That's the whole signal, and it's a weak one.
Here's how weak. OpenAI launched its own AI text classifier in January 2023 and withdrew it in July 2023 for low accuracy. OpenAI. Independent testing has found again and again that detectors flag non-native English writers, people who write plainly, and students as AI at rates that should end the product category, while light editing, a paraphrase pass, or a prompt asking for varied style sails AI text through unflagged. Vendors publish accuracy in the high nineties measured on their own test sets; in the wild, false positives of several percent are normal, and for a school or a publisher that means accusing innocent people every week. The most accurate AI detector is not accurate enough to act on alone, and most universities have quietly stopped treating the score as evidence.
How to tell if something was written by AI without a tool is a matter of tells, not proof. Uniform paragraph length and sentence rhythm. A weakness for tidy triplets and balanced contrasts. Hedged, agreeable, generic phrasing with no specific detail, no named example, no opinion. Confident claims that are wrong when you check. On a page, no author, no date, nothing that could only have been written by someone who was in the room. The strongest single tell is the absence of anything specific. "De-AI-ing" text, rewriting the output to pass as human, mostly works by adding exactly those things, which is also what makes the writing worth reading.
Other media are a different question. AI images still fail on hands, text, reflections and physics, though each generation fails less; some now carry C2PA content credentials, a signed record of how they were made, which you can check where it exists. Whether a chatbot is an AI, "are you AI," is answered by the chatbot, which the major assistants are instructed to do honestly. And an "SEO checker," which search engines like to show next to "AI checker," is an unrelated tool that audits a page for search problems; that's on the SEO reference.
AI detectors are a coin flip with a confidence score. The reliable tell is the one no tool measures: whether anything on the page could only have been written by someone who was there.
