B. Institutional Use Cases: Commissioning AI for Group and Department Tasks

Purpose: Six cases from the perspective of a research group (DISSINET) and a university department (Religious Studies at MUNI), where AI is not a personal assistant but a commissioned workforce for institutional tasks. The argument is different from the personal cases: not "this helped me work better" but "this made a project start and finish that would otherwise never have happened."

The periphery task problem

Every research group and department has a stable backlog of tasks that are:

These tasks accumulate. The departmental website stays outdated for three years. The thesis archive exists as a spreadsheet nobody looks at. The five-year synthesis gets written once for the grant report and never again. The curriculum guide lives as a vague intention.

AI changes the calculation. A task that would take a researcher 40 hours takes 4 hours of commissioned work plus review. The threshold for "worth doing" drops dramatically. Projects that would not start, start. Projects that would not finish, finish.

This is the gain. But it comes with risks worth naming directly — covered at the end of each case and synthesised at the close.

UR — Religious Studies Department

Case 1: Website audit against purpose

The task. The departmental website has accumulated content across years of different editors, platform migrations, and shifting priorities. Nobody has a complete picture of what is there, what is broken, what is redundant, or how well the site serves its actual purpose. The audit that "should" happen stays in the backlog because it would take someone several days of tedious reading and cataloguing.

The commission.

Here are the scraped contents of the department website in markdown (120 pages).
Here is the programmatic document describing what this website is supposed to do
and who it is for.
Here are the visit statistics for the last 12 months.

Do four things:
1. Evaluate how well the content serves the stated purpose, section by section.
2. Find dead content: pages with no visits, outdated information, orphaned sections.
3. Find structural problems: broken links, inconsistent navigation paths, missing cross-references.
4. Find cooperation candidates: student projects, research outputs, or faculty profiles
   that could be linked but are currently isolated.

Output: a structured report with a priority tier for each finding (fix now / fix later / consider).

The gain. The audit that would take a person three days of careful reading takes an afternoon — one session to run the analysis, one to review the output and check the most important findings against the actual site. The structured report with priority tiers is immediately actionable: it goes to the department meeting as an agenda item, not as a vague "we should look at the website."

The quality gain is also real. Claude reads the entire 120 pages with the same attention — it does not get bored by page 80, it does not mentally downgrade the student section because it is less interesting than the research section. The consistency is structural.

What you can lose.

False actionability. The report exists and has priorities — so it looks like a decision. But it is a report about a website that a real person with institutional knowledge has not yet fully reviewed. A finding that says "this cooperation partner page is underlinked" may be correct technically, but the political reason it is underlinked may be something Claude cannot know. Acting on the report without that review can produce technically correct but institutionally wrong changes.

Maintenance debt. The audit produces a list of things to fix. Fixing them takes time. If nobody is assigned ownership of the website before the audit, the report creates a to-do list without a to-doer. The audit is done; the work is not.

Case 2: Curriculum navigation guide

The task. Students — especially at the bachelor's level — navigate the curriculum badly. They pick courses without understanding dependencies, miss electives that would serve their interests, and get to year three having never seen the big picture. A guide that maps the curriculum, explains the dependencies, and gives each course a human-readable description in context would help. It would also serve as a self-study guide and an advising aid. Nobody writes it because it would take weeks to compile and keep current.

The commission.

Here are the course syllabi for all 34 courses in the Religious Studies curriculum
(converted to markdown from the university information system).
Here is the curriculum regulation document with requirements by level.

Build an interactive navigation guide:
1. A visual map of course dependencies — which courses build on which.
2. For each course: a 3-sentence human description (not the official abstract) of
   what you actually learn, what it connects to, and what kind of student it suits.
3. A "where am I" guide for three typical student profiles:

   - first-year BA student who wants to specialise in comparative religion

   - second-year student who realised too late they need more methodology

   - MA student considering a thesis in sociology of religion
4. A self-study extension list: for each core course, 2–3 things to read or try
   outside class that deepen the material.

Output in structured markdown suitable for a web page.

The gain. This is genuinely a weeks-long task if done by a person — reading 34 syllabi carefully, drafting non-bureaucratic descriptions, mapping dependencies, writing for three different student profiles. Claude can produce a solid first draft in one session. The first draft will need significant editing — the human descriptions will be generic in places, the dependency map will miss tacit knowledge — but editing a draft is a different kind of work than building from scratch. The project moves from "we should do this someday" to "we are doing this."

What you can lose.

Institutional voice without institutional authorship. The guide carries the department's voice — it is published under the department's name, shapes how students understand their education. If it is AI-drafted without the faculty being involved in reviewing it, you have an institutional document that nobody actually wrote. The three-sentence course descriptions, in particular, need to be reviewed by the people who teach those courses — because a technically accurate description of what a course covers can miss what the teacher believes the course is for.

Currency. Curriculum guides go stale. If nobody owns the task of updating it when courses change, a guide that was accurate when published becomes misinformation within two years. The guide being easy to produce does not mean it is easy to maintain.

Case 3: Twenty years of theses as public visualization

The task. The department has 20 years of bachelor's and master's theses — several hundred pieces of student work. The metadata exists in the university repository: titles, authors, years, keywords, supervisors, grades. This is a rich archive of what students have researched, what has changed over time, and what directions the field has taken at this department. A visualization would celebrate student work, help current students find adjacent theses, and show the department's intellectual history. Nobody has done it because it requires sustained data work that sits outside anyone's job description.

The commission.

Here is a CSV export of 23 years of thesis metadata from the university repository:
title, author, year, degree level, keywords, supervisor, defence grade.

Do three things:
1. Clean the keywords (there are obvious duplicates and variations — normalise them).
2. Produce an analysis: keyword frequency over time, supervisor clusters,
   topic trends by decade, relationship between topic and degree level.
3. Draft a narrative text for a public page: "What has Religious Studies at MUNI
   researched over 23 years?" — aimed at prospective students and the general public.
   Keep it warm and specific — this is about real student work, not statistics.

Output: cleaned CSV, analysis in markdown, narrative text in markdown.

The gain. The data processing step — cleaning keywords across 600+ entries, building frequency tables, identifying clusters — is exactly the kind of systematic work Claude does well and humans find tedious enough to avoid. The narrative text requires human editing but Claude produces a solid structure. A project that would sit in the spreadsheet indefinitely gets done.

What you can lose.

Student consent and data sensitivity. The theses are in a public repository, but compiling them into an analysis that surfaces individual names, supervisors, and grade distributions creates a new kind of document. Some students may not expect their work to be analysed this way; some grade data may be sensitive. The visualization needs a consent and privacy review before publication, which is easily skipped when the output looks finished and clean.

Celebration vs. evidence. A public visualization of 23 years of student work can be read as celebration (look what students have done) or as evidence (here is what this department has emphasised, what supervisors dominate, where grades cluster). These are not the same document. If the department publishes what they intend as celebration but the data also shows supervision imbalances, the document can do unintended work.

DISSINET — Research Group

Case 4: Working-progress data snapshots

The task. DISSINET produces large structured datasets from ongoing encoding of historical sources. At any given time, several researchers are encoding in parallel, and the current state of the data — what has been encoded, what remains, what the distributions look like — is visible only to people with access to the repository and the knowledge to query it. Making the current state of the data visible to the whole team, and eventually to the public as a working-progress snapshot, is valuable but never urgent enough to happen systematically.

The commission.

Here are the current worktables from the DISSINET repository (CSV exports):
persons, events, sources, encoders, dates.

Produce a working-progress snapshot:
1. Current state: how many records of each type, coverage across the source corpus,
   encoding progress by source and by encoder.
2. Distribution analysis: key variables (event types, person roles, geographic spread)
   with charts.
3. A one-page public summary: "Where the DISSINET corpus stands as of [date]."
   Written for a scholarly audience who knows what we are doing but not the current details.
4. Three "emerging findings" that can be said honestly at this stage — patterns
   that are already visible in the data without waiting for the full corpus.

Output: markdown summary + chart-ready tables + public text.

The gain. The snapshot that previously required a researcher to spend half a day querying, tabulating, and writing now takes one focused session. Done regularly — once a quarter, once per grant period — it creates a living record of the project's progress that serves both internal orientation and external communication. The team knows where the data stands; funders and collaborators get a real picture, not a placeholder.

What you can lose.

Premature commitment. "Emerging findings" that are visible in partial data can be wrong when the corpus is complete — the first half of the sources may systematically differ from the second half. Publishing patterns from incomplete data, even with honest caveats, creates impressions that are hard to retract. The snapshot is useful internally; public release requires more caution than the clean output might suggest.

Case 5: Five-year synthesis narrative

The task. After five years of research, DISSINET has published papers, produced datasets, built tools, trained students, and established an international network. The synthesis — what has been achieved, how it fits together, what it means for the field — is something that should exist in multiple forms: a funder report, a public statement, an onboarding document for new members, a conference talk outline. It never gets written in full because the people who could write it are busy doing the next thing.

The commission.

Here are the materials from five years of DISSINET:

- Published papers (titles, abstracts, DOIs)

- PhD theses completed under the project

- Dataset documentation

- Tool documentation (InkVisitor, DISSINET database, network viewer)

- Annual reports to the funder (previous four years)

- List of international collaborations and events

Produce:
1. A narrative synthesis (2000 words): what DISSINET has achieved, how the
   parts connect, what the intellectual contribution is. Audience: scholarly,
   non-specialist in the specific period.
2. A 300-word version for the project website.
3. A set of 6–8 "claims we can now make" — statements about what the project
   has established that go beyond individual papers.
4. An onboarding reading list: for a new member joining now, the 5 papers/datasets
   to read first to understand what we know and how we work.

Output in markdown.

The gain. A synthesis that would take weeks of writing — pulling threads from memory, re-reading five years of outputs, drafting and revising — takes a day of commissioned work plus a day of careful editing. The "claims we can now make" output is particularly useful: it forces the question of what the project has actually established, which is valuable for grant renewals, conference presentations, and recruitment.

What you can lose.

The narrative is not the achievement. A well-synthesised account of five years of work can be written from the abstracts and reports alone — and can sound more coherent and cumulative than the actual research trajectory was. Research groups do not develop linearly; there are dead ends, pivots, methodological arguments that were never resolved. A narrative that smooths those over is easier to read and less true. The researchers involved need to verify that the synthesis matches their experience of what happened, not just the publishable outputs.

Centroid pull. Asking Claude to synthesise a research group's achievements over five years is asking Claude to describe the group in terms that fit the statistical centre of how similar projects are described. The distinctive, difficult-to-name thing that DISSINET does — the specific combination of historical close reading and computational network analysis, the particular sources, the methodological position — is exactly what is hardest to represent in a synthesis that draws on abstracts and reports. This is where the B.centroid-periphery problem is most acute: the distinctive things are the things that do not summarise easily.

Case 6: Research opportunity mapping ("normal science")

The task. Every established research project reaches a point where the founding questions have been addressed and new directions are not yet obvious. There is a large volume of data and methods that could support many different studies — but identifying which ones are tractable, which ones are original, and which ones connect the group's existing work to open questions in the field requires systematic thinking that no individual researcher can do alone while also writing papers.

The commission.

Here is the DISSINET corpus description and current data structure.
Here are the main published papers (abstracts).
Here are the methods and tools we have built.

Do two things:
1. Identify 8–10 research questions that could be answered with the existing data
   and methods — studies that could be done without major new data collection.
   For each: what data is needed, what method, what the likely finding would look like,
   and what its contribution would be.
2. Identify 3–4 areas where the existing data intersects with current debates in
   the broader field (based on what you know of recent historiography of medieval
   heresy and religious movements) — places where DISSINET's specific corpus
   could make an external contribution.

Output: structured markdown, one section per research opportunity.

The gain. The output is a research menu — a set of tractable options that the group can evaluate against their interests, capacity, and strategic priorities. This is useful as an input to a research planning meeting, a starting point for a student looking for a thesis topic, or a framework for a grant proposal. No individual researcher has time to systematically map the possibility space; Claude can produce a first draft that the group then filters and augments.

What you can lose.

Centroid capture at the research design level. This is the highest-risk case in this document. Claude's picture of what research questions are worth asking in medieval religious history is the statistical centre of the published literature it was trained on. The questions it identifies as tractable and original are the ones that look tractable and original from that centroid — which means they are probably the ones most likely to confirm existing frameworks and least likely to challenge them. A research group that adopts this menu without critical scrutiny is outsourcing its intellectual agenda to the training distribution.

The use is to get a starting list and then argue with it: What is not on this list? What does this list assume that we do not assume? What would a researcher outside this tradition say is missing? The output is useful as a provocation, not as a plan.

False tractability. Claude does not know the actual state of the data — the gaps, the encoding inconsistencies, the sources that were harder than expected. A research question that looks tractable from the corpus description may not be tractable in practice. The menu needs review by the people who actually work with the data.

What you gain and what you can lose

What you gain

Projects that start and finish. The website audit, the curriculum guide, the thesis visualization, the five-year synthesis — these are tasks that have been on the institutional backlog for years, not because nobody cared about them, but because they required sustained effort that was never available. AI lowers the threshold. A project that required 40 researcher-hours now requires 4 hours of commission and review. The calculation changes: it is now worth starting. And because starting is no longer the hard part, finishing becomes realistic too.

Scale at quality. AI does not just produce outputs faster — it produces them at a consistency that humans working under time pressure cannot match. Reading 120 website pages with equal attention on every page, cleaning 600 keyword entries with consistent normalisation, extracting the same set of variables from 34 syllabi — these tasks benefit from the consistency of an agent that does not get bored, does not skip the last third because they are tired, does not apply different standards to the courses taught by senior vs. junior colleagues.

The first draft as a forcing function. Having a complete draft — even an imperfect one — changes the nature of the work that remains. Editing is different from writing. The curriculum guide draft forces the faculty to read it and correct it, which is a more productive activity than trying to write the guide from scratch in a committee. The five-year synthesis draft makes the gaps and smoothings visible, which prompts the better conversation. The draft is an argument to have a conversation with.

What you can lose

Ownership without authorship. The outputs of these commissions — website content, curriculum guides, public narratives — carry institutional authority. If the institutional actors have not deeply engaged with them, the authority is not backed by the knowledge it claims. A curriculum guide that was published after light review by one administrator is a different thing from one that every course teacher has read and corrected. The output being clean does not mean the ownership is established.

False closure. A finished output creates the impression that the task is done. The website has been audited. The synthesis has been written. But audit findings that are not acted on are not improvements; a synthesis that no team member can defend under questioning is not knowledge. The risk is that the output becomes a substitute for the work rather than a beginning of it.

Maintenance debt. Periphery tasks done at scale produce artefacts that need maintenance. A curriculum guide that becomes outdated in two years is worse than no guide — it actively misleads. Before commissioning a task, ask: who owns this when it is done? If the answer is nobody, the task may not be ready to be started.

Centroid capture at institutional scale. When a research group asks Claude to identify research opportunities, or a department asks Claude to describe what it is about, the answers reflect the statistical centre of how similar groups and departments are described in the training data. The distinctive things — the unusual methodological combination, the specific intellectual position, the research culture that makes this group different — are exactly the things that do not transfer well through a commissioned synthesis. Every case in this document requires human review for exactly this reason: not to catch errors, but to restore the specificity that AI synthesis tends to smooth out.

Related