AI Along Your Value Streams: A Practical Map

When leaders ask how to use AI, the question usually arrives as a tool question, meaning which assistant, which platform, which vendor. That framing skips the more useful step, because the place to start is the work itself. What is meant is the chain of activities that turns an input into a result a customer values. This chain is a value stream.

A value stream runs end to end, so Procure-to-Pay runs from a purchase requirement to a paid invoice and Issue-to-Resolution runs from a customer raising a problem to that problem being solved. Once the stream is explicit, the question changes from "which tool" to "which step and how far should AI take it". That is a question you can answer with evidence. The examples in this series are kept deliberately simple. Real value streams are usually far more complex, with many more steps, systems and hand-offs. Mapping them is what makes that complexity workable, which is why the whole stream is always taken end to end. Value reaches the customer only when the whole chain delivers, so a faster single step helps nobody if the work then waits at the next hand-off. Looking at the whole stream therefore means looking hard at the hand-offs between steps, because automating one step on its own often just moves the bottleneck along. Looking at the whole stream is what makes it possible to judge whether a change really shortens delivery or only looks impressive in one place.

For each step the series distinguishes four levels of how deeply AI is involved, four levels that form the backbone of everything that follows.

Classical: people do the work by hand. AI is not involved.

Assistive: a person does the work and AI suggests. The human keeps every decision.

Agentic: AI performs a bounded step on its own, and a human reviews and approves the result before it counts.

Autonomous: AI runs the step without case-by-case approval. A human monitors the system and sets the guardrails.

The pattern across the levels matters more than any single level, because with each one the human role moves further from doing the work towards specifying, reviewing and governing it. The human keeps control, only now at the gate rather than at the task. Each step up trades more of the routine work for more control and integration effort. For a business case that trade matters more than the label on the level.

For most organisations, especially cautious or regulated ones, the realistic target is agentic with a human gate. The AI does the bounded work and a person still approves it before it counts. Full autonomy stays rare and situational.

Each step up has its own price of entry and asks for groundwork in data, connections and controls long before the AI takes anything over. Moving from classical to assistive is mostly a matter of access. It takes digitized processes and data, a capable AI model plus people trained to use its output and to judge it. Moving to agentic asks for a foundation. That means solid data in integrated systems or a data warehouse, API access so the AI can act rather than only suggest plus a bounded process with a clear success criterion and a review gate. Moving to autonomous is the highest bar. It asks for reliable IT infrastructure with monitoring, a mature and largely automated process plus strong data governance and guardrails with a human fallback. The infrastructure builds up and carries forward, while the control points are added level by level. The concrete threshold still depends on the stream, so the maturity view marks what each step up requires. The real cases then show what an organisation actually had in place before it moved.

A second price is easy to miss, because deeper integration asks for more and more of your own knowledge inside the model before the AI becomes useful in your company. Proprietary data and context become part of the cost, alongside the licence fees. How much that matters depends on how you run the model. On public consumer tools your inputs can feed the provider's training, while providers promise the opposite for their interfaces. OpenAI states that data sent to its API has not been used to train or improve its models since March 2023 unless a customer opts in. Zero data retention and a private or on-premise deployment go further still. For a regulated organisation that containment choice belongs in the plan from the start, right next to the value case.

An enterprise can be described with ten value streams, which together cover almost every part of a company.

They fall into three groups, where the core value chain creates and delivers the product, the planning steers it and the supporting functions keep it running. The table below names each stream and what it does.

This map turns the series into a directory, because a company can point to any part of its operating model and read there how AI changes it.

Every stream in the series is described the same way, a consistency that is deliberate because it lets you compare one stream against another and trust that the same questions were asked. Two sets of dimensions do the work.

The first set of dimensions describes the structure of the stream and lays out, for each of the four maturity levels, the steps in the flow and how they change. You can see which steps merge, which shift and which appear. On top of that sits a profile of the whole stream. It names its primary goal, the typical performance indicators, the prerequisites for the agentic and autonomous levels and how the work splits between people and AI. A short assessment of strengths, weaknesses, opportunities and threats rounds it out.

Idea-to-Market is a worked example, because the figure below lays the stream out across the four levels and makes the merging and shifting of steps visible at each one.

The second set of dimensions describes each real-world case, where every example carries the same fields and stays comparable with every other case in the series. Those are the reported metric, the point where a human reviews or approves, the control gates, how much the result depends on context, the risks and lessons and the source. Only cases that could be verified against a credible source are included. Every case carries a reality-check. It says where the number does not transfer and what breaks outside ideal conditions.

A framework is only worth as much as the evidence behind it, so the evidence stays in plain sight. As the series works through the streams, the real cases add up into a library that sits behind every post. Every case is recorded the same way. It gives a quick picture of how the AI sits inside the work, what it changed, the number that was reported and, just as important, the point where that number stops holding.

Think of the map below as the table of contents, where each case sits under its value stream and the collection grows with every post in the series.

Take Amazon Q Code Transformation, behind whose tile sits one case, laid out the way all of them are. It names the claim, the human who signs it off, the limit and the source it rests on, so the case can be checked.

New cases are added over time, each tied to the level of AI integration it shows and each with its positive and negative sides. Together they build a comprehensive picture and give real examples of what AI could mean for your own company, while keeping one thing in view. Not every company and not every industry can absorb an entry at full risk.

The rest of the series takes the streams one at a time, with each post walking a single value stream through the four levels and showing the real cases with their limits. So you can see how far AI can go in that stream, from a careful first step to a fully agentic setup. Next up is Idea-to-Market.