What Durham told us
Fifty nine people spent two days telling us what they need. This is what they said, drawn from every mark they left on the walls.
Companion to the AI Starter Kit workshop debrief
Draft. Confidential, for engagement use only.
Two days in Durham answered two questions that would otherwise have taken months of guesswork, and exposed four that nobody has answered yet.
The answers are good news, and they surprised us. Campuses asked for a starting package they can adapt; nobody asked for a library of options to browse. And the list of what they are short of reads as permission and paperwork: contracts, policies, training, and cover to experiment safely. Technology barely appears on it.
The four open questions are the reason a well built kit could still sit unused. Campus data is not clean enough for half of what is planned. Nobody agrees what ready means. No one has said what success would look like. And nobody's job is to keep any of it current.
All four can be closed. None of them requires a large amount of money. Two of them can be settled in the one hour convening on September 11. That is the most useful thing this workshop produced, and it is the argument for what comes next.
What happened in the room
And why writing on walls turns out to be good evidence
On the first day, participants worked in small groups across five rooms, rotating three times, so that most people talked to most people. Overnight, everything said in those sessions was pulled together into six large themes and written out by hand on flip chart paper. Six stations went up around the room.
On the second morning, everyone walked the room and responded. They wrote on the sheets. They added notes on sticky pads. Where a question had been drawn as a line between two options, they placed a mark along the line to show where they stood. They voted with stars for the things they most wanted.
That exercise produced unusually good evidence. People were reacting to a summary of the previous day instead of to survey questions we wrote, and they wrote in their own words, in front of peers. So what appears on those walls is closer to what a vice president actually believes their campus needs than most instruments would produce.
Every sheet was photographed. All thirty five of them have since been read line by line, including crossed out words, notes squeezed into margins, and the exact position of a sticky note on a drawn scale. Six hundred and two separate marks were recorded, along with one hundred and eleven votes. Everything below comes from those marks, and every point can be traced back to the sheet it came from.
How much was written at a station shows where people spent their time; it is a poor guide to what matters most. The Policy station drew the fewest notes and produced some of the sharpest statements of the two days.
What campuses told us
Two clear answers, each of which arrived at several stations at once
They want a starting package they can adapt
Four different groups, working in different rooms on different topics, were asked a version of the same question. All four gave the same answer, and none of them knew the others were being asked.
The navigation group drew it as a straight choice. Should the kit offer ready made packages by institution type, or should a campus assemble its own from separate parts? Four votes landed along the line between the two options rather than at either end, and the written answers said the same thing. Recommendation bundle w/ easy a la carte subs.
Start with a template and customize to individual context.
One person simply wrote Both.
Navigation station
The technology group was asked whether campuses should buy or build. The answer on the sheet was Buy but customize to individual institute context.
Technology station
The policy group reached the same place from the content side. Frameworks / Templates that a campus can use as a starting point.
And, on course design, Build module maps that allow institutions to customize a course to fit their needs, but have that module map to help guide what is needed.
Policy station
Two people independently named what would make swapping parts workable, in the same two words. Flow chart.
Flow chart assessment.
They want something that tells a campus which package it should be starting from. Navigation station
The bookshelf idea is close, but a shelf is something you browse and campuses said they do not have time to browse. What they described is a recommended package for their kind of institution, a clear way to swap out the parts that do not fit, and something at the front that tells them which package they are. That is a design decision, and this workshop made it answerable without further study.
What they are short of is permission and paperwork
When people voted for what they most wanted, the votes went to contracts, policies, training, and cover to experiment without getting into trouble. AI itself drew almost nothing.
The use case group listed seven things the kit could deliver and invited votes. Ten of the twelve votes cast went to two of them. First, an agreed list of approved tools and the rules for what data can go into them. Second, training designed separately for faculty, staff, students, and administrators, including guidance on whether to buy that training or build it. The remaining two votes went to a block asking for governance structures, role definitions, changes to org charts, and implementation templates. Nothing went to drafting an AI policy from scratch, or to procurement, or to monitoring. Use case station
The navigation group wrote out a list of twenty five things they wanted on the shelf. Most of the list is about dealing with vendors. Criteria for assessing them. The questions to ask them. In one participant's words, the watch outs, pitfalls, landminds when working with vendors.
Contract language that protects transparency and data ownership. How to manage the running cost of these tools. Honest feedback from peers who have already tried something. Navigation station
The community group's most supported items were about cover rather than capability. A sandbox that is easy to use, and security for internal testing + exploration, safe to fail + learn.
The group also marked administrators, not faculty, as the main audience for the kit, and put two votes on the word playbook. Community station
One note names the problem underneath all of it. Many opportunities & incentives. Are there any sticks?
Policy station
What holds these institutions back is being unable to move: missing contracts, missing guardrails, and missing internal agreement about what is allowed. A kit that ships model contract language, a vendor question set, a safe sandbox pattern, and a training ladder will get used quickly. One that ships technical reference designs will be admired and shelved. That is a far cheaper thing to build than most people expected going in.
What nobody has answered
Four gaps, in the order of how much damage each one does if it stays open
Campus data is not clean enough to use half of this
This came up at three separate stations, and in each case somebody stated it as a fact about their own institution. Nobody framed it as a request for help.
The clearest version runs across two sticky notes written by one person. No school currently has student engage data that is clean/valid across all support teams. This is the necessary foundation to deploy AI to support navigating a university.
The same person wrote a version of it on two further sheets, which is the strongest signal of conviction anywhere in the two days. Navigation station
The use case group described the same thing from daily experience. Connecting up the data was called the hardest part. Degree audits are being checked by hand because the course catalog is wrong and out of date. Sharing data across institutions to match records was named as a real obstacle. Data sharing and reporting drew the most votes on that sheet. Use case station
The change group asked directly where data infrastructure fits into the readiness conversation. The answer that carried a vote was Immediately. First.
Change station
The kit as currently described has six shelves and none of them is about data. A campus that takes a navigation tool off the shelf and runs it on records nobody trusts will get a worse result than doing nothing, and will conclude that AI does not work. The fix is cheap. Instead of building a seventh shelf, put a short honest check at the front of the kit that tells a campus whether it is ready for a given tool yet, and points it at what to fix first if it is not. That single step protects every other investment on this list.
Nobody agrees on what ready means
The disagreement runs deeper than which assessment to pick. It is about what the word is measuring in the first place, and the room said so out loud.
One note states the confusion in three words. Readiness is not maturity.
Another asks the question underneath it, whether readiness should be broken down by who is being assessed, faculty, students, staff, or by what is being assessed, such as data or governance. Nobody answered. Change station
The grid people used to organize the discussion was drawn onto the paper in ballpoint during the walk itself, and a second person drew a different grid on the same sheet. Neither existed beforehand, which tells us there was no shared way of thinking about this to react to. Change station
Across two sheets, at least eight different readiness assessments were named by the people who use them, including instruments from consulting firms, from a cloud provider and a higher education association, from a nonprofit intermediary, one forthcoming from a major funder, plus several homegrown surveys of students and faculty. Not one was recommended over another. Change station
When asked which assessment a campus should use, and for which purpose, the sheet received a single blank sticky note and no written answer at all. Change station
One participant proposed turning the question around. Starting with the goal, why are we using AI? to what end? Then, assess how ready we are for that?
Change station
Publishing all eight assessments on a shelf hands the hardest choice back to the campus least equipped to make it, and campuses have already told us they will not work through a list. The better answer is already written on the wall. Ask a campus what it is trying to achieve, then assess it only against that. This turns the readiness shelf from a catalogue into the front door of the kit, which is the same thing campuses asked for in the first finding.
No one has said what success would look like
The question of whether any of this works came up at five of the six stations, each time as a question and never with an answer.
The single most supported note of the whole change station asks for a method rather than a number. Each use case should have an impact theory.
In plain terms, before a tool goes on the shelf, somebody should have to say what they expect it to change and how they would know. Change station
The navigation group drew the question of who decides whether an initiative worked as a line running from the student at one end to the institution at the other, and asked people to mark where they stood. Five marks went down. One near each end, and three bunched in the middle. The written answers were This is a very deep philosophical question. It depends.
and It depends, but that should be known and determined up front.
Navigation station
Asked what success looks like for the Starter Kit itself, the room gave four different answers and voted for none of them. Adoption and contribution. Faster timelines. The number of use cases in the library over time. Student outcomes. Elsewhere, the only two measures written down were how much the resources get used and how much gets contributed back, both of which count activity rather than effect. Change station
Two people asked for the measure to be fixed before anything is deployed. Target outcome measures for students, faculty, system level, defined at the outset.
And, from the policy group, what outcomes are we looking for here? and how to measure?
Change and Policy stations
This is the gap that matters most to anyone funding the work, and it is also the cheapest one to close. Requiring a one page statement of expected impact before anything goes on a shelf costs a contributor an hour. It gives the collaboration something real to evaluate against in a year, and it is a standard that very few comparable efforts hold themselves to. Adopting it would make this the rare initiative that can show what it changed rather than how much it was downloaded.
Nobody's job is to keep any of it current
The gap shows up twice, once inside the collaboration and once on every campus, and in both places it was written down as an open question with no proposal attached.
On the collaboration side, the navigation group asked how anyone would know what good practice is, and immediately identified what is missing. Need curator of AI Starter Kit.
Then, who does this now? who actually 'does' AI well + can be a trusted source?
Then, how to keep it relevant, answered with three words. Fast, focused, flexible.
Navigation station
On the campus side, the policy group opened with Who owns the policy?
and separated two things that usually get muddled. Governance & policy are not the same thing.
A further note set out what governance would have to settle, including which decisions get made, who needs to be in the room to make them, and what structure lets that happen. Policy station
The same group named the role campuses do not yet have, and the problem underneath it. Who becomes the AI champion for a campus? What do credentials/skills look like for such a person?
The community group put it as a flat requirement. Need an AI champion for each dept.
Policy and Community stations
The technology group raised it as a risk to anything that gets built. Ownership + support post-build.
And one note pushes back against handing it to a single office. Encourage shared ownership between Academic, IT, & Research. No one portfolio can own all of it.
Technology and Policy stations
Almost everything campuses asked for goes out of date within a year: the vendor guidance, the tool comparisons, the contract language. The room asked for that material and warned about its shelf life in the same breath. Without someone whose job is to refresh it, the kit becomes an archive within a year or two of launch. A named person with a refresh schedule is a small, specific line item, and it is exactly the sort of thing that gets cut from a budget because it produces no deliverable of its own. Cutting it would waste the rest.
What they asked us to build
Taken from the lists people wrote themselves, in their own words, grouped for use
The six shelves are categories. This is the content. It comes from the long written lists rather than from individual sticky notes, so it captures considered asks. Everything here was requested by name.
- Standard AI language for contracts and requests for proposal
- Contract terms that protect transparency and data ownership
- A tool that reviews a campus policy and tells it what is missing
- A way to spot duplicate vendor contracts and compare contracts against policy
- A generator for course level AI policy
- Worked examples of governance structures, including who needs to be at the table
- Guidance on what belongs in policy, what belongs in guidance, and what is simply a stated belief
- Criteria for assessing AI vendors, and the questions to ask them
- The known traps, written down by people who have hit them
- Honest peer feedback on tools colleagues have already tried
- Demonstrations of tools campuses built themselves, with a name and contact for each
- Guidance on managing the ongoing running costs of these tools
- A vetted list of organizations that can help, including any that will help for free
- A training ladder in three steps, proposed independently by two different groups, moving from basic AI literacy, to judgement about when to use it, to genuine fluency, with the prerequisites for each step made clear
- Different training for faculty, staff, students, and administrators, with a certificate at the end
- Guidance on whether to buy training or build it internally
- Templates for small innovation grants that get faculty experimenting
- A way of measuring whether the training actually changed anything, by group
- An open reference design for the underlying technology, which one group built out on paper in full detail during the session
- A guided tool that asks a campus sixteen questions about its own situation and produces a technical design that fits
- An up to date picture of what tools institutions are actually running today
- A practical way to test whether an AI tool is producing useful results, including human review and side by side comparison of models
- A simple inventory of what AI a campus already has, which two separate groups asked for
- Course scheduling and registration guidance for students
- Help finding a way through financial aid, including satisfactory academic progress rules
- Career exploration and simulation tools
- Support for advisers managing large caseloads, and a way of reaching students with an unpaid balance before their classes get dropped
- A way of seeing how students are engaging with support services, and how those services are reaching back
- Methods for checking whether an implementation actually worked, and for testing whether a tool's output can be trusted
- Guidance on where the line sits between what people should do and what AI should do
- How to move from a pilot to full scale, and what that does to job descriptions and daily work
- A plain glossary of terms
- A full eleven step sequence for running this kind of change was written out during the session, starting with naming an owner and building a small team, moving through explaining why, choosing use cases, anchoring the work in institutional values, setting policy and governance, implementing, guarding against change fatigue, and finally building it into budgets, performance management, and planning so it does not depend on enthusiasm
- Fifteen named references and frameworks were listed alongside it and can be used directly
- Practical ways to head off change fatigue, raised as a need that will persist well past launch
- Process mapping, so that a campus understands a broken process before automating it
- The smallest team you can get away with at each stage
What did not fit anywhere
The parking lot is worth reading, because several items on it belong in the plan
Seventeen topics were written onto a sheet for things that had no home. Some are genuinely peripheral. These are not.
- How far the kit should stretch. Written in noticeably larger letters than everything around it. This is the same question as the first finding, and it needs a decision rather than a parking space.
- A marketplace or inventory of tools. Overlaps directly with the vendor material campuses asked for, and with the AI inventory two separate groups requested.
- Certification designed and run by institutions themselves, for staff, administrators, faculty, and students, including a version taught by students to students. Four separate institutions raised this, the highest agreement recorded on any single item across the two days.
- Risk assessment. Written as policy against compliance against a third thing the writer left blank, which suggests nobody has worked out what that third thing is.
- Working directly with the major cloud and AI providers, with a commitment from them to bring industry change back into curriculum and faculty development.
- A standing community of practice and a group carrying lessons between convenings, which is the ownership gap described in different words.
Four moves
What this points to next. The recommendation itself belongs to the group.
Decide whether the kit ships packages or a catalogue
This is the clearest thing campuses said, and it is a design decision that can be made now, without more research. Building recommended packages by institution type is a different job from filling six shelves, and switching later is expensive.
Decidable in the September 11 convening.
Put a readiness check at the front door
One short honest assessment that asks a campus what it is trying to achieve, tells it whether its data can support that yet, and points it at the right starting package. This closes two gaps at once and stops campuses from taking tools they cannot use.
Needs a small amount of design work and one clear owner.
Require a statement of expected impact before anything goes on a shelf
One page. What this is expected to change, and how anyone would know. It costs a contributor an hour and it gives everyone funding this something to hold the work against in a year.
Decidable in the September 11 convening. Effectively free.
Name and fund a curator
Someone whose actual job is keeping the material current, with a set refresh schedule. This is the smallest line item on the list and the one most likely to be cut, because it produces no deliverable of its own. Without it, the rest goes stale inside a year.
A resourcing conversation among the funders involved.
Two of these four can be settled inside a one hour call. The other two need a short conversation about who pays for what. None of the four requires further study, because the evidence for all of them is already written on the walls of a student center in Durham.
How we know this
This section is deliberately more careful than the rest of the document, because it is where you decide how much to trust the rest
What was read
Thirty five photographed flip chart sheets across six stations. Every sheet was read twice at full resolution, and anything hard to make out was enlarged and read again. Six hundred and two individual marks were recorded, along with one hundred and eleven votes. Each one carries a note on how legible it was and, where the handwriting allowed more than one reading, what the alternative was. Quotations here preserve the writer's own spelling and shorthand. Personal names written on notes have been removed. Institution names are kept only where an institution publicly offered something to the group.
Three things that limit how far this can be pushed
- The six stations were themselves a summary. They were produced overnight from recordings of the first day using AI tools, so the headings and the pre written themes are a machine assisted reading of day one rather than something participants wrote. What people added and voted on the next morning is their reaction to that reading. The four strongest findings here rest on what participants wrote themselves rather than on those pre written themes, which is why we put weight on them.
- Vote counts are not a poll and cannot be compared between stations. One hundred and eleven votes were cast by up to fifty nine people, and only one station carried a written instruction to vote at all. Votes show where energy went on a particular sheet. They cannot tell you what the room as a whole ranked highest.
- Some readings are uncertain. A handful of names, one attribution, and several notes partly hidden by tape or other notes could not be read with full confidence. None of the findings above depends on any single uncertain note. Any specific assessment tool or institution named in the readiness section should be confirmed before it is quoted to anyone outside this group.
What would have to be true for the main argument to be wrong
The argument is that the four open questions matter more to whether this gets used than adding more content does. For that to be wrong, campuses would have to be held back by missing resources rather than by unmade decisions. That is possible, and there is a way to tell. If institutions that already have clean data, a chosen readiness assessment, and a named owner still are not using the kit six months from now, then content is the constraint and this document has the emphasis backwards.
There is also something conspicuously missing from the walls. Money, staffing capacity, and internal politics barely appear anywhere in the two days, beyond a single pencilled note reading cost. It is very unlikely those concerns were absent from the room. Writing on a wall in front of your peers selects for what is comfortable to say out loud, and budget pressure is not that.
What we looked for and did not find
No written summary of the workshop existed before this one. A three panel whiteboard from the session has not yet been transcribed, so anything captured there is not represented here. We worked from the gallery walk material and did not go back to the underlying day one recordings.
The question this does not answer
How many of the fifty nine people intend to actually use the kit at their own institution, and what would stop them. A gallery walk tells you what a room finds interesting. Whether an institution will adopt any of it is a different question. A small number of direct follow up conversations before the September convening would answer it, and the answers could reasonably change the order of the four moves above.
Drawn from the gallery walk record of 31 July 2026. Thirty five sheets, six hundred and two marks.
Draft for internal review. Confidential, for engagement use only.