AX Field Notes

All casesNo. 061

Brandeis Gives Its IT Help Desk an AI First Line

A three-month documentation fix came first; at six weeks, chat contained 36%, voice 32%

Case no.
061
Primary source
Trade press article and interview + university ITS pages
Source published
2026-10-23

1. Why this case

Brandeis University in Massachusetts put an AI voice and chat assistant in front of its IT help desk in summer 2026. Callers to the help desk line now reach the assistant first, and the IT website offers it as a chat, at any hour. It answers routine technical questions from the university’s own documentation. When it cannot, it collects the person’s details and hands the request to a human, either as a ticket or, during business hours, as a live call transfer.

The case is small, but it is unusually specific about three things that most help desk AI stories leave out:

  1. Where the effort went. The build took more than eight months, and Steve Knowles, executive director of ITS Client Services and Engagement, calls the three-month knowledge base revamp “the most important and underestimated component of the entire project,” in EdTech Magazine’s words.
  2. What the assistant is not allowed to touch. The team limited it to Tier 0 and Tier 1 topics that need no connection to systems holding student records protected under FERPA, the U.S. student privacy law.
  3. An early number against a stated goal. Six weeks into deployment, chat containment was 36% and voice containment 32%, against a goal of 30% to 50% within the first six months.

The university also says plainly that the motive was not staffing. The account comes from an EdTech Magazine article (2026-10-23) quoting Knowles, a UC Today interview (2026-10-05) with Eli Jacobson, whose team built and tested the assistant, and Brandeis ITS’s own web pages. The figures are the university’s own statements, not independently audited results.


2. Problem

A staffed desk with office hours. Brandeis ITS publishes its Help Desk hours as Monday to Thursday 9 a.m. to 5 p.m. and Friday 9 a.m. to 2 p.m., closed on weekends and holidays, with a note that “Hours may vary when classes are not in session.” Outside those hours, a person can submit a ticket online, leave a voicemail, or send an email, and “Requests are reviewed and addressed on the next business day.”

Waiting makes the problem bigger. Knowles describes the help desk’s job as partly emotional: “A big part of what client service centers do is anxiety remediation. If someone has an issue that they sit on for an extended time, the anxiety builds,” he told EdTech Magazine. “Then, you are no longer just helping to fix a technology issue, you are also calming the anxious and sometimes angry emotions that arise when your technology isn’t doing what it is supposed to.”

The motive was access, not cost. Knowles is explicit: “Our main focus was simply delivering a modern campus experience, giving everyone instant, round-the-clock access to answers on their own schedule,” he says. “It wasn’t a reactive fix for ticket surges or staffing challenges, but the operational benefits have been a huge win.”

The documentation did not agree with itself. When the team prepared the knowledge base for the AI, it found “conflicting information scattered across departmental websites, public-facing pages and internal runbooks.”

The common requests sit next to sensitive data. EdTech Magazine notes that an assistant handling password resets, multifactor authentication (MFA) troubleshooting, and account access questions is “in contact with student authentication workflows and data protected under the Family Educational Rights and Privacy Act.”

Confirmed facts: published Help Desk hours Mon–Thu 9–5, Fri 9–2, closed weekends and holidays, may vary when classes are not in session; after-hours requests reviewed the next business day; motive stated as round-the-clock access (“It wasn’t a reactive fix for ticket surges or staffing challenges”); conflicting information found across departmental websites, public pages, and internal runbooks; common requests touch authentication workflows and FERPA-protected data.


3. Solution

3.1 Documentation first

The build “took more than eight months” and “was anchored by a three-month knowledge base revamp,” according to EdTech Magazine. “Support documentation is everything,” Knowles says. “These agents thrive on the data you feed them.” His advice to other IT directors is to prioritize the knowledge base.

Brandeis ITS describes the result as a closed set of sources: “The AI assistant answers strictly from a curated library of verified Brandeis IT documentation. The assistant cannot search the open web or generate unverified information.” Knowles also ties the assistant to future documentation decisions: “If any future knowledge base governance group decides to pivot to a new knowledge base model, we will work with them to ensure it functions correctly with the AI assistants.”

3.2 One assistant, two channels

Brandeis ITS says the assistant is “powered by Zoom.” UC Today names the product as Zoom Virtual Agent and reports that “Over the summer, Jacobson’s team built and tested voice and chat versions” of it “to provide 24-hour support for tier one queries.”

  • Chat: a “How can I help?” button on ITS web pages. The person types a question, and the assistant returns “instructions, troubleshooting steps, or quick links.”
  • Voice: the assistant greets callers to the help desk number automatically and talks them through initial troubleshooting steps.

The ITS page lists 20 topic areas, including account login and password resets, MFA (Duo), email, Wi-Fi, VPN, the Moodle learning management system, Workday, software licensing, hardware, printing, classroom technology, and phishing reports. The assistants answer IT questions only.

3.3 Scope limited by data, not only by topic

“At Brandeis, the team intentionally limited the assistant’s initial scope to Tier 0 and Tier 1 support topics that would not require connection to any system containing FERPA-protected data,” EdTech Magazine reports. Tier 0 is self-service and Tier 1 is first-line support. “The only student data the assistant can access is a username, and only if the student volunteers it to generate a ticket.” According to the article, the current deployment has no ERP, customer relationship management, or student information system integrations.

The Brandeis page adds that the system collects “basic contact details (such as your name and Brandeis email or phone number), your specific technical inquiry, and the transcript of your interaction,” that it “complies with Brandeis University Data Governance standards and FERPA regulations,” and that interaction data “is never used to train public AI models.”

Knowles: “Security and compliance are the first part, and arguably the most important part of any service we release.”

3.4 The handoff is half the job

The assistant’s second job is to get the person to a human. Brandeis ITS describes the paths:

  • Chat, not resolved: “The chatbot will gather your details and automatically create a TDX support ticket or provide you with the correct phone number to call (during business hours).”
  • Voice, business hours: the assistant “may ask you some basic questions before transferring you to a member of our team such as your name, preferred contact, phone number and Brandeis Username,” then transfers the call.
  • Voice, after hours: it creates a support ticket, and “an ITS team member will reach out on the next business day.”
  • Any time: “you can ask to speak with a human at any point.”

UC Today reports that connecting the bot to the university’s service management platform “was a massive part of the whole project.” Jacobson: “We’re not just trying to answer everybody’s questions. We are trying to answer people’s questions as best we can, but we’re also using it as a facilitation device for getting people onto the next step, generating tickets and triaging the requests to get them in touch with who can help them as quickly as possible.”

3.5 Phased launch

The rollout came in three steps: a soft launch of the chat assistant to faculty and staff in July, the voice assistant in August, and students during move-in week at the end of August. UC Today reports that “Preparing the campus to accept the bot was a key part of the planning.”

3.6 Daily review after launch

With the assistant “only a few weeks into full production,” Jacobson’s team “reviews the analytics daily, looking at what answers the bot is giving, and gathers user feedback,” UC Today reports. Negative interactions are used “to adjust processes, knowledge bases and the AI guardrails,” through an escalation process that traces each failure to its cause. Jacobson: “Was it that the knowledge base it is pulling the information from was wrong, or is it a training issue for the AI bot, or is it a business process that we need to re-evaluate? All three of those things have come up so far in just the couple of weeks that we’ve been using it in production.”

Confirmed facts: build of more than eight months including a three-month knowledge base revamp; answers only from a curated library of verified Brandeis IT documentation; powered by Zoom (Zoom Virtual Agent per UC Today), voice and chat built and tested over the summer; 20 listed IT topic areas; scope limited to Tier 0/1 topics needing no FERPA-protected system; username only if volunteered for a ticket; no ERP, CRM, or SIS integration; TDX ticket creation from chat and voice; live transfer during business hours, ticket and next-business-day follow-up after hours; July chat for faculty and staff, August voice, students at end-of-August move-in; daily analytics review with failures traced to knowledge base, bot training, or business process.


4. Impact

Figures as reported by EdTech Magazine on 2026-10-23, from Brandeis.

MetricFigureBasis
Chat containment rate36%Six weeks into deployment
Voice containment rate32%Six weeks into deployment
Launch goal30% to 50% containmentWithin the first six months
Calls to the voicebotMatched or slightly exceeded last year’s average call volumeKnowles, as paraphrased by EdTech Magazine
Need for human techniciansDropped by more than halfKnowles, as paraphrased; measure not defined

The six-week figures sit inside the range set for the first six months.

The top deflected ticket categories are “account login and password resets, email access and configuration, wireless network connectivity, VPN access issues and software installation and licensing.”

On the staff, Knowles says: “By letting the voice and chat assistants handle routine questions, our technicians can focus more on interpersonal skills during customer interactions, as well as complex problem-solving and high-touch support.” Brandeis ITS states that the service “does not replace our dedicated and professional IT Staff” and that it is meant to “assist our staff in supporting everyone on campus.” Jacobson, a few weeks into production, was “not ready to call it a success,” according to UC Today.

5. Insight

The first answer moved to the agent, at every hour. Without the assistant, a routine question gets a human answer only during staffed hours, and after hours it waits for the next business day. Now the assistant gives the first answer at any time, and the human desk receives what is left: tickets with contact details already collected, and calls transferred after the assistant has tried. Knowles says technicians can now focus more on complex problems and high-touch support.

The documentation was the project. The model came from a vendor; the three months went into making the university’s own help content consistent. Conflicting instructions across departmental sites, public pages, and internal runbooks would have produced conflicting answers. The commitment to work with any future knowledge base governance group shows the team treats the documentation as part of the assistant, not as a separate library.

Scope was set by data access. By choosing topics that need no connection to student records, the team could launch without integrating the student information system or ERP. The trade-off is that the assistant explains how to fix a problem but does not look anything up for the person.

Handoff design carried as much weight as answers. Jacobson calls the bot “a facilitation device,” and the service management integration was “a massive part of the whole project.” At 32% to 36% containment, the handoff path is not an edge case, and the ticket the assistant writes is part of the service for many requests.

Failures are sorted by cause. Each bad interaction is traced to the knowledge base, the bot’s training, or a business process. That keeps a wrong answer from being blamed on the AI by default and sends the fix to the people who own each part.

Limits. The numbers are early: six weeks for containment, a few weeks of full production for the UC Today interview. Containment is not defined, and the article does not say how “need for human technicians” was measured. The figures are the university’s statements, not audited results, and Brandeis says the project was not aimed at staffing. Nothing is published on cost, satisfaction, or what happened to the technicians’ hours.

One-line takeaway: A university help desk can hand its first line to an AI agent if it first spends the time making its documentation agree, limits the agent to topics that need no access to records, and builds the handoff to humans as carefully as the answers.


6. Implementation hypotheses (not in the source)

6.1 A curated article set as the answer boundary

Basis: The assistant “answers strictly from a curated library of verified Brandeis IT documentation.” The revamp found “conflicting information scattered across departmental websites, public-facing pages and internal runbooks.” UC Today lists a wrong knowledge base as one of three failure causes.

The revamp most likely produced one canonical article per common issue, mapped to the 20 topic areas, with an owner for each. Where a departmental page and an internal runbook disagreed, someone had to choose the correct steps, and the losing versions were probably removed from the assistant’s source set rather than left for it to reconcile. Internal runbooks would likely have been split: steps a user can do went into public articles, while steps that need staff privileges stayed out. Voice probably needed its own short, sequential versions of the same articles, since a caller cannot click the “quick links” the chat channel returns. When the daily review finds a wrong answer caused by content, the fix would go to the article owner, and the change then reaches both channels at once.

6.2 Handoff routing by staffed hours

Basis: In business hours the voice assistant transfers calls after asking for name, preferred contact, phone, and Brandeis Username; after hours it creates a ticket for next-business-day follow-up. Chat creates a TDX ticket or gives the phone number. Connecting to the service management platform was “a massive part of the whole project.” The Tech Support page says hours “may vary when classes are not in session” and that some identity checks “can not be done over a standard telephone call or email.”

The assistant probably checks a business-hours calendar before choosing between a live transfer and a ticket, and that calendar would need to follow the academic calendar, not a fixed weekly schedule. The ticket it writes likely carries the collected contact fields, the topic area, and the conversation transcript, since the page says transcripts are collected to generate tickets; that lets a technician start where the assistant stopped. Each ticket likely also records which channel created it, which the team would need to count escalations per channel. Because the assistant connects to no identity or records system and an identity check cannot be done by phone, the password and account requests it deflects are most likely handled by pointing people to self-service reset steps, with anything that needs identity verification passed to staff.

6.3 Containment counting and the daily failure queue

Basis: Chat containment 36% and voice 32% at six weeks against a 30% to 50% six-month goal. The team “reviews the analytics daily” and traces each failure to the knowledge base, bot training, or business process.

Containment in contact-center tools is usually counted as the share of bot sessions that end without a ticket or transfer, computed per channel, which would explain why chat and voice are reported separately. That count also includes people who simply gave up, so containment is not the same as resolution, which is probably why the team gathers user feedback alongside the analytics. The daily review would likely pull sessions that ended in a ticket, a transfer, or negative feedback, tag each with one of the three causes, and route it: content fixes to the article owner, intent and guardrail changes to whoever configures the virtual agent, and process questions to the help desk lead.


7. Sources

  1. EdTech Magazine (Akeya Dickson) — How AI Agents Are Helping Campus IT Teams Manage Student Support During Peak Demand
    https://edtechmagazine.com/higher/article/2026/10/how-ai-agents-are-helping-campus-it-teams-manage-student-support-during-peak-demand-perfcon
    2026-10-23
  2. UC Today (Marcus Law) — How Brandeis University Moved Its Campus Comms to Zoom
    https://uctoday.com/how-brandeis-university-moved-its-campus-comms-to-zoom
    2026-10-05
  3. Brandeis University ITS — AI-Powered Help Desk Voice and Chatbot
    https://www.brandeis.edu/its/projects/chatbot-voice-assistant1/
    Publication date: not shown on page (accessed 2026-11-17)
  4. Brandeis University ITS — Tech Support
    https://www.brandeis.edu/its/tech-support/index.html
    Publication date: not shown on page (accessed 2026-11-17)