From information architecture to AI
Domain legibility as organizational infrastructure
What revising a banking navigation model revealed about how work is organized.
Christopher Ferrell
AAA 2026 · National Association for the Practice of Anthropology
1 / 11
Open with the work, not an AI failure statistic: “I was trying to learn whether a proposed structure made sense to people who knew banking far better than I did. The most useful moments were not the correct answers. They were the moments when participants gave me a reason to change the structure.” This presentation follows the submitted abstract’s argument through a small number of concrete revisions. The empirical case concerns information architecture and domain knowledge, not an evaluated AI deployment. Suggested pace: about 45 seconds.
The conditions of inquiry
The usual research route was constrained.
Security
A required browser extension obstructed testing with employees and many clients.
Availability
Frontline bankers could not readily step away for a remote session.
Credibility
Senior practitioners needed questions worth their limited time.
Access required more than recruitment. It required preparation.
2 / 11
Situate the work within teller-product modernization at a large financial technology provider. These are the practical constraints behind the abstract’s account of an adaptive rather than preplanned research program. At the time, UserTesting required a browser extension that internal and client security protocols flagged as unsafe. Do not imply all research or all usability methods were impossible. The extension requirement was subsequently removed. Frontline availability was a separate barrier. More senior personnel could participate, but I needed to understand enough of the domain to make their participation worthwhile. These particulars come from my retrospective account, not from the tree-test data. About 75 seconds.
Learning enough to ask
Preparation made a provisional model possible.
Study the domain
Legacy documentation.
Prudential and compliance requirements.
Propose a structure
Internal card sorting.
A navigation model.
Carefully constructed tasks.
Invite correction
Three rounds of tree testing.
A final moderated client study.
The model was something to think with, not something participants had to endorse.
3 / 11
I had been in the role about six months when I drafted the study and launched recruitment. Studying the legacy teller documentation alone was insufficient: I also needed the broader regulatory context. Product leaders helped identify prospective participants, and the invitation gave them a reason to participate. The larger research sequence unfolded over seven and a half months: regulatory and legacy-system analysis, internal card sorting, two unmoderated tree tests with role-screened panel participants, then a moderated client round. Twelve moderated sessions were conducted and ten retained after technical exclusions, according to the project record. Do not suggest every participant was a frontline teller or that each study used the same recruitment channel. About 90 seconds.
Tree testing as a boundary probe
Not only “Where would you look?”
Also “Why would it belong there?”
The visible task
Find where to configure a banking rule, permission, or document.
The analytic question
Which distinction makes that location meaningful to a practitioner?
A hesitation could reveal a problem in my categories, not a deficit in the participant.
4 / 11
Briefly define tree testing for listeners outside UX: participants receive a task and navigate a text hierarchy without the visual interface. Here the tasks concern configuration and governance, not simply performing a teller transaction. In moderated sessions I could ask about reasoning, clarify, and revisit choices. Those conversations are analytically different from independent first-path performance. The abstract calls this a boundary probe: an inquiry into whether distinctions remain meaningful across practitioners and systems. Shared structure across legacy platforms was a working theoretical commitment, not a fact established in advance. The next slides are schematic reconstructions of documented revisions, not screenshots or verbatim participant transcripts. About 75 seconds.
Revision · What does the label invoke?
“Disclosure” carried an obligation.
Before · Forms label
Disclosure Presentation
Participants repeatedly associated disclosures with compliance.
After · Forms label
Document presentation
Name the presentation of the document without absorbing the obligation to disclose.
The distinction was between presenting an artifact and governing a requirement.
Schematic label comparison. Revision recorded after the moderated study.
5 / 11
The task asked about disclosures presented to customers. Across several sessions, the term disclosure drew participants toward Compliance. Some reconsidered independently; others discussed the distinction. One expert explicitly separated disclosures as a compliance concern from receipts and customer-facing documents as a forms concern. I changed Disclosure Presentation to Document presentation in the Forms area. This was not evidence that practitioners misunderstood disclosures. The original label crossed a meaningful boundary. The tracker records the Forms-label revision as post-study; do not imply it was subsequently validated by the reported final-round score. Source: Session Tracker, Revisions, “Neutralize ‘disclosure’ in Forms”; supporting observations across multiple sessions. All wording on the slide is paraphrase except the documented labels. About 100 seconds.
Revision · Where does authority sit?
Dual control was a permissions question.
Tested path
People & AccessRisk ControlsDual Control
Revised path
Roles and accessRole definitions / capabilitiesDual control / high-risk approvals
Risk and cash were plausible associations. Neither clearly named the authority to approve.
Remove the misleading grouping. Place dual control with role capabilities.
6 / 11
Explain dual control in plain language: certain actions require two people’s approval. One participant reasoned through risk, cash, and roles. That reasoning made operational sense because high-risk teller actions often involve money leaving the building. Another expert directly recommended removing Risk Controls as the intermediate grouping for a permissions change. The final path places dual control under role definitions and capabilities. Separately, a participant suggested Roles rather than People, while other feedback showed that People could evoke customers rather than employees. This slide combines two documented revisions in the final path; do not attribute both to one participant or a single moment. Source: Session Tracker, Revisions rows “Remove ‘Risk Controls’” and “People → Roles,” with supporting observations. This is a hierarchy change as well as a naming change. About 110 seconds.
What the revisions made explicit
A shared word is not a shared governing rule.
Disclosure
Present a document.
Govern an obligation.
Dual control
Encounter a risk.
Grant approval authority.
Foreign exchange
Handle physical currency.
Set an exchange rate.
Ask what is being governed, not only which words seem related.
7 / 11
Use foreign exchange as a brief supporting example, not a third full case. The tracker records Foreign Currency Exchange becoming Foreign exchange rates in task and label wording. Yet at least one participant still considered the physical-currency domain after the word currency had been removed. That is why this is not merely a story about keywords. The underlying distinction concerned physical handling versus pricing. Exact timing of the FX wording change is unresolved in the reconstructed record; do not assign it a precise session date. Across these episodes, the same term can connect legitimate concerns without identifying the governing location of a particular configuration. This is the anthropological pivot: categorical meaning is situated in professional practice, not exhausted by a dictionary definition. About 75 seconds.
The organizational artifact
From a menu to an account of the domain.
System settings
Roles and access
Transactions and pricing
Currency, cash and reconciliation
Compliance and risk
Forms and communication
The contribution was not six buckets. It was a more explicit account of boundaries and governing relationships.
Six-domain overview, labels abbreviated for presentation. Thirteen legacy navigation categories were consolidated.
8 / 11
Return to the abstract’s six-domain model and thirteen-category consolidation, but do not recite the whole taxonomy. The domain list is a presentation-level summary, not a reproduction of every final node. A taxonomy says what groups with what; the stronger ambition is to specify what governs what, and where authority resides. The empirical evidence here concerns practitioner interpretation and revisions to the navigation representation. The formal account of governing relationships is an analytic product of that work, not something independently tested as executable AI behavior. Avoid equating agreement among sampled participants with a universal banking ontology. The model remains open to revision. About 70 seconds.
Reading the case with James C. Scott
Legibility for whom?
At the expense of what?
The danger
A formal model can make work administratively visible while erasing practical distinctions.
The methodological obligation
Let practitioners challenge the categories before those categories organize their work.
Legibility is not automatically faithful to the work it represents.
James C. Scott, Seeing Like a State (1998). Interpretive framing, not a quotation.
9 / 11
Scott is a caution, not a stamp of approval for formalization. Seeing Like a State examines the consequences of administrative simplification and the loss or suppression of practical, situated knowledge. This banking project is not equivalent to the state projects he studied. The connection is a question to bring to the model: what becomes visible, whose distinctions count, and what remains outside it? Senior practitioners are knowledgeable but do not stand in for every frontline experience. Participant correction is a way to make simplification answerable to practice, not proof that all exclusions or power asymmetries have been resolved. Do not expand into the labor manuscript here. Keep Scott tied to the actual methodological choices. About 90 seconds.
The implication for AI
Automation inherits the distinctions we give it.
An agent that confuses a document with an obligation, or a risky action with permission to approve it, inherits an organizational problem.
What this case demonstrates
Practitioner feedback changed the language and structure used to represent work.
What this case argues
That work belongs inside implementation, not outside it as preliminary cleanup.
AI implication, not a measured AI outcome.
10 / 11
Return explicitly to the abstract: it argues that foundational domain research constitutes implementation rather than simply preparing for it. Define the scope of that claim. Making governing distinctions explicit is part of implementing organizational meaning; it is not the whole of building, deploying, evaluating, or securing an AI system. No agent was evaluated in these studies. The case does not establish a general causal explanation for enterprise AI failure. It does show why tools cannot be expected to repair an unexamined model of work automatically. This qualified framing preserves the abstract’s substantive argument without claiming more than the evidence supports. About 75 seconds.
Applied anthropology’s contribution
Before a system can act on a model of work, practitioners need a way to correct that model.
The revisions were not cosmetic.
They changed how the work could be understood.
Christopher Ferrell · AAA / NAPA 2026
11 / 11
Close without a new framework: “The value of this research was not that people could find the right answer in my tree. It was that their reasoning changed the tree, and made the account of the work more explicit. That is an applied anthropological contribution to implementation.” Leave the final statement on screen for discussion. Suggested closing: 30 seconds. Total suggested speaking time is approximately 15 minutes, with flexibility to shorten the preparation and artifact slides.
DRAFT REVIEW NOTES: Evidence sources are the submitted AAA Abstract Final, Domain Logic document, anonymized Tree Test Session Notes, Session Tracker workbook, both Tree Test Outline workbooks, and the presenter’s account of access constraints. The revisions used here are documented in the tracker. Before public delivery, verify final displayed labels and institutional naming permissions. No confidential chart screenshots, performance scores, or unverified direct participant quotations are included. Scott citation: James C. Scott. 1998. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press. Speaker notes are part of the downloadable HTML; review them before distributing the file.