Understand the Work

1. Information — gathering, filtering, comparing, reconciling.
2. Judgment — noticing, interpreting, predicting, prioritizing.
3. Expertise — building context, cues, pattern recognition, situational knowledge.
4. Coordination — trust, negotiation, signaling, informal handoffs.
5. Exception handling — detecting surprises, recovering, escalating.
6. Feedback / learning — learning whether an action was good and improving future judgment.
7. Authority / control — recommending, deciding, approving, executing, overriding, escalating.
8. Memory / provenance — preserving why something happened and what evidence supported it.

• Unit of capability — the smallest thing the AI can do.
• Unit of intervention — where you introduce the change.
• Unit of analysis — the smallest boundary wide enough to capture consequential effects.
• Unit of evaluation — where success and harm should actually be measured.
• Unit of accountability — where authority and responsibility for the resulting system reside.

Not every AI deployment is transformation.

• Tool substitution — same work system; one activity becomes easier or faster.
• Workflow redesign — steps, handoffs, sequencing, or information flow change.
• Work-system redesign — roles, coordination, authority, expertise, feedback, incentives, and technology are reconfigured around the outcome.
• Operating-model / boundary change — structure, centralized vs distributed capability, outsourcing, customer participation, or business boundaries change.

The farther the effects propagate, the more the intervention resembles transformation rather than tool adoption. Widening the analysis does not mean every project becomes a giant program. The point is to discover whether a local intervention is genuinely local.

AI does not remove constraints; it often relocates them.

A task can become dramatically faster without increasing end-to-end system throughput if another part of the system remains capacity-constrained. Theory of Constraints, Little's Law, and queueing theory establish the operations mechanism: improve a non-constraint and you may create inventory, not value; push a busy downstream resource toward high utilization and waits rise nonlinearly.

People also adapt to overload: shortened review, rubber-stamping, changed escalation thresholds, prioritizing easy cases. Bottleneck analysis must connect operations to measurement, incentives, authority, and quality.

Ask: What currently limits end-to-end throughput? Is the AI acting on that constraint or a non-constraint? What is the capacity of the next one or two stages? What happens if volume doubles? Where will excess work wait? Which quality thresholds bend under load?

AI changes the expertise-production function. It can compress apprenticeship—or hollow it out.

Removing participation in a task may alter not only today's workflow but the mechanism by which future expertise develops. Some apparently tedious work also builds the context and pattern recognition needed to judge AI output later. Protect the learning mechanism, not necessarily the legacy task. Do not claim that all manual work builds expertise; ask whether this work does.

Ask: What does a novice become by using the system? Compare expert and novice with and without AI on decision quality, ability to challenge recommendations, learning over time, and robustness when the AI is wrong or unavailable—not only on speed or adoption.

AI can change who effectively controls a decision even when the org chart, policy, and stated accountability have not changed.

Capability is not authority. Formal authority and effective authority are different things.

Do not treat "the decision" as one indivisible act. Decompose information acquisition, attention, framing, candidate generation, recommendation, selection, approval, execution, override, escalation, and modification. Authority can shift at one stage while remaining unchanged at another.

A human approval step is meaningful only when the human has the evidence, time, competence, independence, practical ability, and operating capacity to challenge the system. Check those conditions explicitly. A review design that works for fifty cases a day may collapse at fifty thousand. Authority shift is not inherently a design failure. The goal is deliberate allocation under real consequences—not maximal human control, and not ceremonial "human in the loop."

AI can make information dramatically faster to produce without making the surrounding work system faster at absorbing, interpreting, deciding, coordinating, or acting on that information.

When information-production capacity outruns decision capacity, useful insights become queue objects. Handoffs degrade. Feedback gets contaminated by workload pressure, activity metrics, and queue-clearing behavior.

AI changes coordination not merely by changing what information exists, but by changing who produces it, who filters it, when it appears, which boundaries it crosses, how much trust accompanies it, how scarce attention is allocated, and who has enough authority to change the surrounding work when reality disagrees. The work may be tightly coupled even when the organization that owns it is fragmented.

• Prefer observation of real work episodes before abstract process interviews.
• Reconstruct one recent hard case or recovery from beginning to end.
• Use interviews to explain observed work, not as the sole source of Work-as-Done.
• Inspect artifacts: screens, spreadsheets, messages, notes, unofficial tools.
• Ask what alternatives were rejected—and what those rejections make disappear from future evidence.
• Triangulate observation, reconstruction, logs, expert/novice contrast, and downstream outcomes.

Watch a real episode. Notice screens opened first, tools switched between, informal artifacts, skips, hesitations, interruptions, and what moves a case up or down. Use short contextual probes—"What caught your eye there?"—without turning shadowing into a continuous interview. Then reconstruct the decision with the participant using the concrete episode just observed.

• Sequence — What did you do first? What happened next?
• Cue — What caught your attention? What did you notice here?
• Contrast — Why this case rather than that one?
• Counterfactual — What would have changed your mind?
• Expertise — What would a newer person miss here?
• Exception — When would this rule not apply?
• Artifact — Show me where you saw that. What tool, note, spreadsheet, or message were you looking at?

• If the proposed solution begins with a model capability rather than a work outcome, restart at Box 1.
• If you cannot explain how the work behaves on a hard case, return to Box 2.
• If you cannot say where the hidden functions will go, do not automate yet.
• If the redesign changes who may recommend, decide, approve, execute, override, or escalate, treat that as an explicit authority-design decision—not a prompt instruction.

Current state: twelve analysts prepare four packages a week each—48 packages/week analyst capacity. Four senior credit officers can thoughtfully review about 13 packages a week each—52 packages/week review capacity. The system is roughly balanced.

After AI: analyst preparation doubles to 96 packages/week. Senior review remains 52. Excess upstream capacity is 44 packages/week. Local productivity doubles; end-to-end decision throughput need not.

Possible system consequences: queue growth, review delays, shallow review, changed approval thresholds, rubber-stamping, triage changes, pressure to modify senior performance expectations, increased risk or rework—or capacity redirected into higher-value analysis instead of more package volume.

The transformation question is not "How do we make analysts produce 96 packages?" It is "How should the underwriting system change now that analysis capacity is cheaper and faster?"

This is not an argument that all bureaucracy is useful, that all manual work builds expertise, that task automation is inherently misguided, or that every AI productivity gain fails without wholesale redesign.

It is an argument that impoverished work models are a common root of brittle AI transformations, and that practitioners need a discovery and decision method before model selection becomes destiny.

Frank Sellhausen works at the intersection of AI and data products, enterprise systems, regulation and risk, and operating practice— including Army training and bank examination seats that taught the same Work-as-Done discipline this chapter uses. This Field Guide treats AI transformation as organizational and systems design: working theses, labeled evidence, and methods meant to survive contact with real work systems.

The larger arc continues past understanding—decide what should change, redesign the work, design the human–AI system, prove outcomes, govern decision rights, and operate as conditions change. This piece stops at understanding, because that is where most programs skip the work that determines whether the rest can succeed.