What you will do

The notebook builds the network for a fictional online travel agency from config/config_travel.xml, then runs the same graph as a management consulting firm.³

1. Load the travel network and list its five specialists. Each one describes what it is good at in a line, and that description is all the broker sees when it decides whom to consult.¹
2. Ask for the general sentiment of recent guest reviews. In the saved run the broker consults only the guest experience specialist, which returns the reply written for it in the file, and the answer comes back with no approval needed.³
3. Ask whether to match a competitor’s lower rates. The synthesis step scores its own answer against the file’s rule, which sends any change to prices or published content to a person, so the answer stops at the human gate.² Approve it and the proposal comes back word for word,¹ or decline it and the run reports the action as cancelled.⁴
4. Load config/config_agency.xml and ask where a post-merger integration can find synergies. The same graph now runs a consulting firm with five different specialists, and its gate fires on layoffs, restructuring and budget moves over $1M,² as it does in the saved run, whose answer proposes a 10% headcount reduction.³
5. Copy either config to start a network of your own. It needs the broker’s two prompts, a list of scenarios with keywords and the agents, each with a description and one canned reply per scenario, and nothing in the package has to change.⁴

Before you start

You need a free Google account and your own OpenAI API key, which you save in Colab’s secrets (the key icon in the left sidebar) under the name OPENAI_API_KEY. Each question makes two gpt-4o calls, one to route it and one to synthesize the answer, which costs a fraction of a cent, and the specialists cost nothing because their answers are written in advance.⁴ The notebook clones the repository and installs its own dependencies, so there is nothing else to set up.⁴ If you would rather spend nothing, the repository’s 19 checks run the real graph against a stub model from a clone, with Python 3.10 or newer and no key.⁴

Articles it puts into practice

What an enterprise AI control plane is turns orchestration and trust into questions to put to any vendor. The broker’s routing and its human gate show both at work in one small network you can read from end to end.

What you will do

The visual guide and its notebook follow one card issuer through four months of purchases, a scenario written for the book, with 318 frauds among 119,556 transactions.²

1. Price each error before fitting a model. A false alarm costs $16 of analyst time and customer annoyance, while a missed fraud costs the amount written off plus $25 of handling.²
2. Measure the model that does nothing. Flagging no transaction at all scores 99.73% accuracy, which is why the lab never reports accuracy again.²
3. Split the data by time, and catch a warehouse column that lifts PR-AUC from 0.30 to 0.98 although it is filed weeks after each purchase, too late to inform the decision.²
4. Compare oversampling, class weights and leaving the data alone. All three rank transactions almost identically, and the first two predict a fraud rate about 36 times the one observed.²
5. Set the threshold from the costs, and then test the review team’s capacity. The cheapest policy flags about five transactions a day for 31% less than flagging nothing, and working every case the team could costs 13 times as much as having no model.²

Before you start

You need a free Google account to run the notebook in Colab. It loads its data from the book’s repository, and Colab already has pandas, scikit-learn and matplotlib, so there is nothing to install, no API key and no cost. Every cell runs as written, and a little pandas helps if you want to change the code.

Articles it puts into practice

88.7% accurate and missing almost every case reads the survey model at five cut-offs and lists the numbers to ask for beside accuracy.

What you will do

A simulated city wants to know whether a parks levy has majority support. Its survey of 1,200 households was written for the book, so the true answer is known and every estimate can be checked against it.²

1. Lay out the design. The four districts are sampled at different rates, so each household carries a weight equal to its district’s population divided by the number sampled there.²
2. Plot support by district before averaging anything. Oldtown, the district least in favor, was over-sampled on purpose and supplies a third of the respondents from under a fifth of the city’s households.²
3. Estimate support twice with DescrStatsW from statsmodels. The raw responses give about 55%, and the weighted estimate is 63%, close to the city’s true 62%.²
4. Give the weighted figure a margin of error that allows for the weights. Unequal weights cut the effective sample from 932 respondents to about 660, and the margin comes to about 4 points either side.²
5. Size the next survey. Holding the margin near 3 points takes roughly 1,000 respondents under simple random sampling and about 1,400 with this design.²

Before you start

A free Google account is all you need to run the notebook in Colab. It reads the survey from the book’s repository and uses pandas, statsmodels and seaborn, which Colab already has, so there is nothing to install and nothing to pay. If margins of error are new to you, chapter 74 of the Visual Handbook explains what sets their size.

Articles it puts into practice

• The same survey says 11.29% or 8.44% applies weights and design-based standard errors to a national survey of 134,897 adults.
• Adoption is outrunning governance reads four recent surveys about AI agents one at a time, because their samples and measures differ.

What you will do

The lab runs on a clinic dataset written for the book, with 200 patients, their blood pressure before and after treatment, their wait times at three clinics and whether they were satisfied.¹

1. Ask the four questions in order. Is the outcome numeric or categorical, how many groups are there, are the groups independent or paired, and do the test’s assumptions hold?¹
2. Read the test off the decision map. Switch to its rank-based twin when the data are skewed or have outliers, and to Fisher’s exact test when a cell’s expected count falls below about 5.¹
3. Route four questions about the clinic through the map. Blood pressure fell by about 9 mmHg within patients (a paired t-test), and the new treatment lowered it more than the standard one did, by 11.6 mmHg against 6.3 (a Welch t-test).¹
4. Report the other two answers with the same care, although neither is significant. Satisfaction shows no clear link to treatment (a p-value of about 0.27), and wait times do not differ by clinic under ANOVA or under its rank-based twin.¹
5. Close each result with its p-value, effect size and confidence interval, as the review chapter does for the five case studies that follow it.²

Before you start

All the notebook needs is a free Google account for Colab, which already has SciPy and statsmodels. The notebook fetches the clinic data from the book’s repository, so it runs free with no setup. It helps to have met the t-test and the chi-square test, and the review chapter turns the whole map into a choose_test() helper with the assumption checks in code.

Articles it puts into practice

Every AI claim is a statistical claim sets out three questions to ask of any AI number. The right test is how an analyst answers two of them, “Compared with what?” and “How sure?”