01
The wrong starting question
Most AI programs begin where the model is easiest to insert: draft, summarize, classify, retrieve, score, recommend. That is a reasonable place to intervene. It is often a terrible place to stop the analysis.
The wrong starting question is: Where can we use AI? That question produces a use-case list. Use-case lists are politically easy, locally measurable, and frequently optimized for what already has an owner, a benchmark, and a story that does not require renegotiating decision rights.
Those same properties make them prone to local optimization. Putting AI on an inherited process may improve the process while preserving the assumptions that should have been challenged.
If we were designing this outcome-producing system today, with current AI capability available from the beginning, what work, roles, handoffs, approvals, and structures would we still choose to create?
That probe is useful. It is not proof that greenfield redesign is feasible. Politics, sunk capital, regulation, and customer contracts constrain what can change. The point of the question is to stop treating the inherited workflow as destiny.
Refuse to score an AI idea as a transformation case until the value path is written past hours saved. A tool case can still be worth doing. Call it a tool case.
Tool case vs transformation case
- Tool case— local capability plus local metric; surrounding system unchanged. Can still be valuable. Do not inflate it into "transformation."
- Transformation case — designed conversion path and complementary work-system changes that reach a value pool and a financial or mission outcome.
Confusing the two produces ROI theater and underinvestment in the redesign work that actually converts capacity into value.
02
The value mechanism and capture chain
The value-capture chain: Capability → Capacity → Redirection → Outcome → Capture. Local capability is not value until capacity is redirected and captured as an outcome the business cares about.
Do not start by choosing the AI intervention. Start by choosing the value mechanism.
Use this as the basic reasoning chain:
Examples of complete chains:
faster document synthesis → shorter underwriting cycle → faster decisions → higher conversion → revenue growthbetter anomaly detection → improved credit selection → fewer defaults → higher risk-adjusted returncheaper routine analysis → delegated low-risk authority → senior capacity released → more expert attention on complex cases → higher decision quality
If the story stops at hours saved, better insights, or adoption, the value case is incomplete.
Where released capacity goes if you do not choose
Released capacity is absorbed when AI reduces effort for some work but management does not deliberately convert that capacity into a target outcome.
Before deploying an AI productivity intervention, decide what will happen to the released capacity if the intervention works.
Absorbed capacity is not necessarily wasted. The problem is unintentional absorption: the organization claims one value mechanism while the released capacity flows somewhere else.
03
Binding constraint vs local inefficiency
After you name a value pool, ask:
What currently prevents the organization from realizing more of this value?
The constraint may be scarce expert judgment, customer demand, approval capacity, poor information quality, slow coordination, legacy software, risk tolerance, policy, physical capacity, incentives, organizational structure, lack of authority, or weak feedback.
An AI intervention on a non-constraint may create local productivity without improving the target outcome. Capacity improvement is not throughput improvement.
Constraint test questions
- What currently limits the outcome we claim to want?
- Is the AI acting on that constraint or on a visible local inefficiency?
- What is the capacity of the next one or two stages?
- What happens if volume doubles at the intervention point?
- Where will excess work wait, and which quality thresholds bend under load?
- If the intervention works, where does the bottleneck move?
Do not celebrate a local productivity win until you can say whether the binding constraint moved—and what complementary change is required for the next constraint not to absorb the gain.
04
Modern AI suitability
Technical suitability tells us what machines can do; it does not tell us what the organization should do.
A modern suitability lens should distinguish at least:
- Raw model capability — what the system can do under specified conditions (benchmarks map this).
- Technical suitability — reliability under real task variation; tool and data access; context; error detectability; verification burden.
- Economic suitability — productivity delta after verification and coordination costs; whether the task is a binding constraint; surplus after displacement.
- Organizational suitability — authority requirements; learning and expertise effects; incentives; political feasibility; downstream consequences.
Benchmarks map capability. Transformation requires mapping capability into a work system.
05
The redesign menu—as a bridge
Once outcome and value mechanism are clear, do not reduce the decision to "automate or not." Generate competing redesigns across a short menu—then take the chosen direction into Chapter 3, where the depth lives.
Redesign space — options to generate
- Eliminate — remove work that no longer earns its keep.
- Simplify — cut policy, process, data, approval, or coordination complexity before adding technology.
- Automate — move suitable execution to deterministic software or AI.
- Augment — improve human judgment, speed, quality, or capacity without transferring authority.
- Delegate — move defined decision or execution authority under explicit constraints.
- Reorganize — change sequence, handoffs, specialization, boundaries, or expertise allocation.
- Create — introduce new work, roles, controls, or feedback loops that did not exist.
- Leave alone — preserve work when intervening would cost more than it returns.
The right transformation often combines several. If AI was treated as the answer, eliminate / simplify / leave alone disappear from the option set.
06
Monday
Pick one live AI idea. Write the value path past hours saved before the build plan hardens. If you cannot, call it a tool case or stop.
- Name the outcome and value pool — not the model task. Who is supposed to receive the value?
- Write the chain — capability → work-system change → operational effect → value pool → financial / mission outcome. Mark where it currently stops.
- Name the binding constraint — what prevents more of that value today? Is the intervention on it?
- Decide released-capacity destination — cost, throughput, quality, cycle time, expert attention, growth, slack—choose deliberately.
- Generate at least three redesign directionsacross the bridge menu—including one that is not "add AI."
- List complements — roles, decision rights, incentives, data, interfaces, ownership.
- Name winners, losers, and vetoes — and the evidence that would change the recommendation.
One Monday pass will not finish portfolio strategy. It will change which ideas survive the next funding conversation.