Where we start
These six sessions are the spine. They do not tell you what to build at your business, because nobody can do that from outside it. They give you the four moves a roadmap needs, in order, so that when you write one for a particular industry you are filling in a frame rather than re-deriving it: sort the candidates, pass a verdict on each, gate them against the data you actually hold, and sequence what survives against the way the business makes money.
★ What you walk out with tonight. Your own candidate list sorted into the five, the rung your organisation is actually standing on, and the name of at least one case that is three rungs above it.
A ladder, not a menu
7 min liveThe order matters and it is not a maturity model about your organisation. It is a dependency: each rung consumes the output of the one below it, so an untrusted rung poisons everything above it.
LiveWhy the ladder is a dependency rather than a maturity score4 min▶
Maturity models rank organisations. This does not. Plenty of businesses run a genuinely good "do it" case sitting on a genuinely solid "see it" foundation for one narrow slice of their operation, while being nowhere on the ladder for everything else. The ladder applies per question, not per company.
What makes it a dependency is that each rung consumes the rung below as an input. A churn score is a prediction built on an explanation of why people leave, which is built on a retention figure everybody agrees on. Break the bottom and the top still runs. That is the whole problem: it does not error, it just becomes confidently wrong, and the confidence scales with how sophisticated the top rung is.
The practical use is diagnostic. When somebody wants a "do it" case and the "see it" rung underneath is contested, you have not found an ambitious roadmap. You have found a reporting problem wearing an AI costume, and it is cheaper to say so in week one.
Self-studyThe three rungs people actually skip, and what it looks like afterwards4 min▶
Skipping "see it" to get to "predict it". The commonest. A churn model gets built while two teams still define an active customer differently. The model trains fine, scores fine, and the first time somebody investigates a flagged account the conversation becomes an argument about whether that account was ever active. Everything after that is spent relitigating the definition.
Skipping "explain it" to get to "decide it". A ranked call list arrives without anybody having established why people leave. It ranks something, the team works it, and nobody can tell whether the calls that worked worked because of the ranking. There is no way to improve it, because there was never a mechanism to improve.
Skipping "decide it" to get to "do it". The rarest and the most expensive. Automatic action on a score nobody has yet decided how to act on manually. The tell is that no human has ever done this task by hand often enough to know what the edge cases are, so the edge cases are discovered in production, by customers.
In all three the failure is the same shape: the thing ships, nothing errors, and the cost arrives later as an argument nobody can settle with data.
Whose categories these are, and whose they are not
4 min liveThere is a well-known published set of categories for this, it has three rather than five, and the difference is worth understanding rather than glossing.
The five above are this playbook's own frame. They are not attributed to anybody and there is no paper behind them. They are a working sort order, chosen because the dependency between rungs is the thing that actually predicts trouble. Anybody presenting a five-category scheme like this as established doctrine is overselling it, including these pages.
LiveThe published three, in their own words3 min▶
The best-known published taxonomy comes from a 2018 management-review article by Davenport and Ronanki. Its three categories, in the article's own terms, are:
Process Automation - the "automation of digital and physical tasks... using robotic process automation", which the article describes as the cheapest and the least smart of the three.
Cognitive Insight - "using algorithms to detect patterns in vast columns of data", which it calls "analytics on steroids".
Cognitive Engagement - engaging "employees and customers with natural language processing chatbots, intelligent agents, and machine learning".
Two things worth noting. Those are the actual category names, and secondary retellings routinely flatten them into generic phrases like "automating processes, gaining insight, engaging customers" - which sound like the same thing and are not quotable as the source. And the three are a taxonomy of technique, sorted roughly by sophistication, where the five above are a taxonomy of question, sorted by dependency. Both are useful; they answer different things.
Self-studyWhat could not be verified, and is therefore not claimed2 min▶
The 2018 article is widely cited as resting on a survey of 250 executives and a study of 152 projects. Those figures repeat consistently across citation indexes, and the article's own text could not be opened during this build to confirm the method. So they are not stated as fact anywhere in these pages, and you should treat them the same way until you have read the original yourself.
There is a second well-known framework in this area, from a 2015 article on automation and augmentation, usually summarised as five strategies beginning with the word "step". The five names are well attested. Their actual definitions could not be read in any primary or full-text source during this build, so they are not described here, and a framework you can only cite by its list of names is not a framework you can apply.
Sorting is fast; arguing about the sort is the point
4 min liveTwelve candidates take about ten minutes to sort. The value is not the tidy list at the end, it is the two or three items where the room cannot agree which rung they are on.
| A candidate phrased as it arrives | Rung | Why it lands there |
|---|---|---|
| "A dashboard showing retention by site" | See it | It reports a number. Whether that number is agreed is the whole question, and usually it is not |
| "Find out why the Leeds site is losing members" | Explain it | It asks for a mechanism, which means it needs a trusted figure first and a reason field that is actually filled in |
| "Flag members likely to cancel" | Predict it | A score per member. Sits on top of an explanation, whether or not anybody produced one |
| "Tell the team who to call today" | Decide it | A ranked, costed action list. Needs the score to be good enough that acting on the order beats acting at random |
| "Automatically send the win-back offer" | Do it | No person in the loop. Needs somebody to have already decided, by hand, enough times to know the edge cases |
| "An AI assistant for the front desk" | Ask again | Not a rung. It names a technology and a location, and no question at all. Session 2 is largely about what to do with these |
Self-studyThe disagreements worth having, and the one that is a red flag3 min▶
Useful disagreement: two people put the same candidate on different rungs because they have different ideas of what it would do. That is a specification problem you have just found for free, and the fix is to split the candidate into the two things they each meant.
Useful disagreement: somebody insists a case is "see it" and somebody else insists the underlying figure is disputed. Both are right, and you have found the real first project.
Red flag: a candidate that nobody can place at all, because it is phrased as a technology rather than a question. "An AI assistant", "a data lake", "a copilot for the ops team". These are not ambitious cases, they are unfinished ones, and sorting them is impossible until somebody says what question they answer. Do not assign them a rung to be polite.
Your list, on the ladder
★ 22 min · bring a real oneFive passes. You should end with a sorted list, a rung you are honestly standing on, and at least one candidate you now know is three rungs up.
Write each candidate in the words it arrived in, including the ones phrased as technologies. Do not tidy them into questions yet; the phrasing is evidence about who wants it and why.
Sort into the five rungs, quickly, and note every case where two people would disagree. Those disagreements are the useful output, not the sort.
Park anything unplaceable in a separate list headed "no question yet". Resist assigning it a rung. That list is usually longer than anybody expects and it is a genuine finding.
Mark the highest rung you genuinely trust today for the area this list covers, with evidence rather than aspiration. "Two teams define it differently" means you are not past rung one.
Measure the gap between that rung and the highest rung anybody is asking for. If it is more than one, you have found the thing to say out loud before the next planning round.
If the gap is zero or negative. It happens, and it is not automatically good news. A list where everything sits at or below the rung you already trust can mean the organisation is being appropriately incremental, or it can mean nobody is asking for anything that would change a decision. Session 3 has a bench that tells those two apart, because it scores what each case is actually worth as well as whether it is possible.
Before session 2
30 min- Take your "no question yet" list back to whoever asked for each item and get one sentence saying what question it answers. Some will dissolve on contact, which is the cheapest result available to you.
- Test your trusted-rung claim. Ask two people in different teams to define your most-used metric, separately, in writing. Compare. This takes ten minutes and settles the argument permanently.
- Pick your three highest-rung candidates and write, for each, what it would consume from the rung below. Session 2 passes a verdict on each of them.
- Bring the list. Sessions 2 through 5 each operate on it, and session 6 turns it into the one-page artefact a vertical roadmap can import.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · Why is the five-rung order described as a dependency rather than a maturity model?
The ladder applies per question, not per company: a business can run a solid rung-five case on one narrow slice while being nowhere for everything else. A and D are the maturity-model reading, and the danger of that reading is it suggests you can skip ahead if you are "advanced enough".
2 · Somebody asks for an automatic win-back offer, and meanwhile two teams define an active member differently. What have you found?
The diagnostic value of the ladder is exactly this. C is the tempting answer and the expensive one: the automation will ship, nothing will error, and every later conversation about whether it worked becomes an argument about the definition instead.
3 · A candidate on the list reads "an AI assistant for the ops team" and nobody can agree which rung it belongs on. What is the right move?
Assigning a rung to be polite destroys the information. An unplaceable candidate is not an ambitious one, it is an unfinished one, and that list is usually longer than anybody expects. Some items dissolve the moment somebody has to write the question down.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 1 cheat sheet · pin this
You now have a sorted list and a "no question yet" pile. Sorting tells you what kind of thing each candidate is; it says nothing about what you should do about it. Session 2 passes a verdict on each one, out of four: automate it, augment the person, build something new, or keep it human. The verdict is a separate axis from the category, conflating the two is the commonest error in this work, and there is a good published argument that the whole field is systematically biased toward one of the four verdicts for reasons that have nothing to do with which is right.