How AI-driven real-time assistance reduces new-agent ramp time

In a study of 5,172 support agents, new hires with AI assistance matched six-month veterans after two months; to get that effect, pair real-time assist (Level AI, Cresta, Observe.AI, Balto) with QA-linked coaching and cohort measurement.

Published Oct 20, 2026 · Updated Oct 20, 2026 · Author: Contact Center AI Index

Primary research hub: Agent assist · Entities: Level AI, Cresta, Observe.AI, Balto, NiCE, Erik Brynjolfsson

AI-driven real-time assistance shortens ramp time mainly by giving new agents the knowledge and wording that veterans carry in their heads. The best independent evidence comes from a study of 5,172 support agents by Brynjolfsson, Li and Raymond, in which agents with two months of tenure and AI access performed as well as agents with more than six months of tenure who did not have it. To get that effect, use a real-time assist tool such as Level AI, Cresta, Observe.AI or Balto, connect it to QA-linked coaching, and measure each hire cohort against a baseline.

What is ramp time, and how should you measure it?

Ramp time (often called speed to proficiency) is how long a new agent takes to perform at the level of an established agent. It is not the length of classroom training. A practical definition:

• Pick two or three proficiency metrics that matter in your operation, such as resolutions per hour, first contact resolution (FCR), the share of issues solved without a callback, and QA score.
• Set the bar at the median of agents with six or more months of tenure on the same queue.
• Track by cohort: the week each hire class crosses the bar on all chosen metrics.

Average handle time (AHT), the total time per contact including after-call work, should be a guardrail rather than the target. Fast new agents who do not resolve issues create repeat contacts. In a Level AI case study, Empyrean reports that AHT initially rose after rollout as agents asked more questions and resolved more on the first call.

What does the research say about AI assistance and ramp time?

The most rigorous evidence is "Generative AI at Work" by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025 and available as NBER Working Paper 31161. It studied the staggered rollout of a generative AI assistant to chat-based customer support agents at a Fortune 500 business-software firm. Key findings from the current version of the paper:

• Average productivity rose about 15%, measured as chats resolved per hour.
• Gains were concentrated in less-experienced and lower-skilled agents, who resolved roughly 30% more issues per hour and improved on speed and quality.
• Ramp compressed: agents with access from their first month reached about 2.5 resolutions per hour after two months, a level agents without access reached only 8 to 10 months into the job.
• Learning persisted: during software outages, agents still performed above their pre-AI baseline, especially those who had followed recommendations closely.
• Attrition fell, driven by retention of newer workers, and customers were less likely to ask for a supervisor.
• Top agents gained little, with small declines on some quality measures.

Treat these as directional. The study covers chat support at a single firm, and the AI was trained on that firm's own top-performer conversations. A voice operation with weaker knowledge content may see smaller gains.

Which parts of ramp does real-time assistance actually shorten?

New agents are slow for four reasons, and AI addresses each differently.

The first two rows are where new hires lose the most time, and they are the least controversial uses: an answer with a citation or a checklist step is easy to verify. Documentation matters too: new agents often spend longer on wrap-up, which our guide to automating after-call work covers. The "what to say" row is where the research shows both the largest novice gains and the slight quality decline for experts.

Which software supports each part of a ramp program?

• Level AI fits teams that want assist, QA and coaching on one data set. Its Agent Assist page describes Phi, a knowledge bot that returns cited answers in the agent desktop, dynamic checklists for identity verification and disclosures, automated summaries, and a supervisor view of live calls. The page claims faster onboarding but publishes no ramp figures. Level AI's coaching page says personalized coaching plans built from 100% of interactions help new hires build foundational skills. Empyrean, which staffs up to 400 agents including 250 temporary CSRs in its busy season, expects assist to be invaluable for brand-new hires on the floor, per the case study.
• Cresta fits programs that want practice before live calls. Its Training Simulator, launched July 2026, builds simulated customers from real conversations, grades practice against the same quality criteria used on live calls, and feeds results into managers' coaching plans. Cresta says this reduces ramp time but published no figures.
• Observe.AI states the most aggressive ramp claim: its Companion Agent page says new reps perform like veterans after two weeks and lists a 12% faster speed to proficiency, with no methodology or customer named.
• Balto fits voice-heavy floors. On its homepage, one customer reports reducing training time from about six months to one month, and One Mnet Health reports onboarding halved from four days to two. Both are vendor-published customer claims.
• NiCE Copilot for Agents fits NiCE CXone shops; NiCE says new agents become experts sooner, without figures.

For a comparison of these tools beyond ramp, see the best real-time agent assist software in 2026.

How should you roll AI assistance out to a new-hire class?

1. Establish the baseline first. Record the last two or three hire cohorts' weeks to proficiency on your chosen metrics.
2. Fix the knowledge base. Remove outdated articles and add the 30 to 50 questions new hires escalate most. An assist tool will surface wrong content just as confidently.
3. Shorten classroom time deliberately. If assist covers lookups, move some product memorization to the floor, but keep systems, policy and empathy training.
4. Practice before go-live. Use simulated calls or supervised shadowing for the hardest scenarios: angry customers, cancellations, regulated disclosures.
5. Turn on assist for the whole cohort from the first live day, with knowledge and checklist prompts first and sales nudges later.
6. Score 100% of new-hire conversations with automated QA in the first 60 days so coaching targets real gaps, not a handful of sampled calls. See the best AI coaching tools.
7. Measure against the baseline cohort, on proficiency, FCR, QA and 90-day attrition.
8. Test without assist at day 60 on a sample of calls or simulations to confirm learning, not just dependence.

What can go wrong?

• Measuring only AHT. Assist can lengthen early calls as agents resolve more; judge ramp on resolution and quality too.
• Low adherence. In the Brynjolfsson study, gains were larger for agents who followed recommendations, and adherence grew over time. Track whether new hires actually use suggestions.
• Over-reliance by top agents. The same study found small quality declines for the most skilled agents, so let tenured agents mute low-value prompts.
• Stale content. Assist amplifies your knowledge base, good or bad.
• Attributing everything to the tool. Hiring quality, schedule and supervisor ratio also move ramp. Use a control cohort where you can.

Limits of this guide

• The Brynjolfsson, Li and Raymond study is the only independent, peer-reviewed evidence cited here. It covers chat agents at one company, and its numbers differ slightly between the 2023 working paper (14% average, 34% for novices) and the current version (about 15% and roughly 30%).
• Every vendor ramp claim cited (Observe.AI, Balto, Cresta, NiCE, Level AI) is vendor-published, and none includes a methodology we could check.
• Ramp benchmarks vary widely by industry, so we do not quote an industry average.
• Product names and features change frequently; Cresta's Training Simulator, for example, launched in July 2026.
• We did not test any of these tools.