What problem does it solve?
Even with good self service, the contacts that reach a human are the hard ones: complex products, exceptions, complaints, distressed customers. Agents juggle several systems while the customer waits, search the knowledge base mid call, and then spend minutes writing notes after every contact (three to five minutes per call at Definity before it automated summaries). New agents take months to become proficient, and attrition means there are always new agents.
Agent assist attacks the time around the conversation (searching, typing, summarizing) and the knowledge gap of less experienced staff, without handing the customer to a machine. That makes it one of the lower risk ways to bring generative AI into regulated customer service, provided the suggestions are grounded in approved content and the recording is handled correctly.
- Industry estimates cited by Brynjolfsson, Li and Raymond suggest that 60% of contact centre agents leave each year, costing firms $10,000 to $20,000 per agent.Generative AI at Work (NBER Working Paper 31161) (2023)
- In the support operation studied by Brynjolfsson, Li and Raymond, agents without AI assistance needed more than six months of tenure to perform as well as assisted agents with two months.Generative AI at Work (NBER Working Paper 31161) (2023)
How does it work?
- Transcribe live. Streaming speech recognition turns both sides of the call into text in real time; on chat the text is already there.
- Understand the moment. The system detects the customer's intent and key details (product, account type, the problem) as the conversation develops.
- Surface knowledge and next steps. Relevant procedure snippets, eligibility rules and next best actions appear in the agent desktop, retrieved from approved content.
- Draft, do not send. On chat and email it drafts replies the agent edits; on voice it suggests wording for disclosures and explanations.
- Wrap up automatically. After the contact it writes a structured summary, fills CRM fields and service request forms, and the agent confirms them.
- Pause on sensitive data. Card numbers and authentication answers are not transcribed or stored, using pause and resume or redaction.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Mainstream
- Channels
- Agent desktop, Phone and voice, Web chat
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Productivity gain | Too few to pool | 15% to 15% | 2 | 2 vendor |
| Accuracy | Too few to pool | about 100% | 1 | 1 organization |
| Handling time reduction | Too few to pool | 20% | 1 | 1 vendor |
| Time saved per task | Too few to pool | 3.5 minutes | 1 | 1 organization |
Value drivers: Employee productivity, Lower cost to serve, Customer experience, Compliance quality.
Indicative value
A contact centre with 500 agents
USD 700,000 to USD 3.6 million
Agent capacity released per year
How this is calculated
Formula: agents * agentCost * onContactShare * timeReduction. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Agents using the assistant agents, agents | 500 | 500 | The reference organization. |
| Fully loaded cost per agent agentCost, USD per agent per year | 40,000 | 60,000 | Editorial assumption. Replace with your own cost, including outsourced seats. |
| Share of paid time spent on contacts and wrap up onContactShare, fraction of paid time | 0.7 | 0.8 | Editorial assumption for occupancy. Replace with your workforce management data. |
| Reduction in handling time, including wrap up timeReduction, fraction of handling time | 0.05 | 0.15 | Conservative against the evidence on this page (Definity says automated summaries saved 3.5 minutes per call; Google Cloud reports a 20% cut in call handle time for Definity's whole program, which also automated caller authentication; the NBER field study measured 14% more issues resolved per hour). |
What it leaves out: Released capacity, realized only if staffing or service levels are adjusted. It leaves out platform and transcription costs, and quality effects such as fewer errors, better compliance and faster onboarding of new agents.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
DBS Bank
Singapore · Banking · 2024
DBS built CSO Assistant in house for its 500 customer service officers in Singapore, who handle queries from more than 250,000 customers a month. It combines a language model tuned to local languages and parlance with telephony and speech recognition: it transcribes the call in real time, searches the knowledge base live, then writes the call summary and prefills service request fields. Pilots began in October 2023; full rollout in Singapore was planned before the end of 2024, followed by Taiwan and Hong Kong. The 20% cut in call handling time in the release is an expectation, not a result.
- Accuracy: about 100%, pilot, October 2023 to July 2024
"Based on data collected since pilots began in October 2023, CSO Assistant has demonstrated transcription and solutioning accuracy of nearly 100%, and when fully deployed, is expected to reduce call handling time by up to 20%."
Claimed by: organization
Definity
Canada · Insurance · 2026
Definity, the parent of several Canadian property and casualty insurers, worked with Deloitte to use Google's AI in its contact centre: summarizing calls, automating caller authentication, analyzing customer sentiment and giving team members real time recommendations. Calls are transcribed, passed through data loss prevention and summarized by language models, with the summaries stored in Salesforce. Definity says automated summaries cut three and a half minutes from each call within about a month; Google Cloud reports shorter call handling times and higher productivity overall.
- Handling time reduction: 20%
"Definity, with help from Google Cloud partner Deloitte, leverages Google’s AI capabilities to summarize calls, automate caller authentication, analyze customer sentiment, and provide real-time recommendations to contact center team members — reducing call handle times by 20% and boosting productivity by 15%."
Claimed by: vendor - Productivity gain: 15%
"Definity, with help from Google Cloud partner Deloitte, leverages Google’s AI capabilities to summarize calls, automate caller authentication, analyze customer sentiment, and provide real-time recommendations to contact center team members — reducing call handle times by 20% and boosting productivity by 15%."
Claimed by: vendor - Time saved per task: 3.5 minutes, per call, within about a month of automating call summaries
"Within about a month, we cut the time agents spend on each call by three and a half minutes."
Claimed by: organization
SEB
Sweden · Banking · 2025
SEB, a Nordic corporate bank, worked with Bain & Company to build an AI agent on Google Cloud for its wealth management division. The agent suggests responses during conversations with customers and generates call summaries afterwards. Google Cloud reports a 15% efficiency gain.
- Productivity gain: 15%
"The agent, built with Google Cloud, enhances end-customer conversations with suggested responses and generates call summaries, helping to increase efficiency by 15%."
Claimed by: vendor
SIGNAL IDUNA
Germany · Insurance · 2025
SIGNAL IDUNA, a German insurer, built Co SI with Google Cloud, BCG and Deloitte: a knowledge assistant that helps customer service agents answer complex health insurance questions. Google Cloud reports that for less experienced agents, information searches are 30% faster and inquiries that previously needed further escalation dropped from 27% to 3%.
No outcome disclosed.
Oportun
United States · Banking · 2024
Oportun, a US consumer lender, replaced manual, sample based QA with Cresta's AI quality management across all calls, combined with real time guidance for agents. Coaching now focuses on the behaviours that drive performance, visible across every call, instead of a small sample reviewed weeks later. No quantified outcome is published.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Approved knowledge articles and procedures with owners
- Call recordings or transcripts to tune intents and test summaries
- The CRM fields and disposition codes the summary must fill
Systems to integrate
- Telephony or contact centre platform with access to the audio stream
- Agent desktop and CRM
- Knowledge base
- Card data redaction or pause and resume for payment calls
Complexity: Medium
Summaries after the call are straightforward. Real time guidance needs low latency streaming transcription, integration with the telephony platform and agent desktop, and a knowledge base that is clean enough to surface the right snippet in seconds.
- 1
Start with after call summaries
Automated wrap up is the fastest, safest win: the agent reviews and confirms every summary, and time saved is easy to measure.
- 2
Clean the knowledge before surfacing it
Real time suggestions are only as good as the articles behind them. Fix duplicates and outdated procedures for the top call reasons first.
- 3
Add real time guidance for a few intents
Pick high volume intents with clear procedures and add knowledge surfacing and disclosure prompts. Watch whether agents use or ignore suggestions.
- 4
Handle sensitive data by design
Make sure card numbers, authentication answers and special category data are redacted or never captured, and that transcripts are stored in region with defined retention.
- 5
Measure with a control group
Roll out by team and compare handling time, resolution and satisfaction with teams that do not yet have it, on the same contact mix.
- 6
Agree how the data will not be used
Tell agents what is recorded and agree that assist data is not used for individual performance scoring unless that is assessed and disclosed separately.
Guardrails
- Suggestions only from approved knowledge, with the source visible to the agent
- The agent confirms every summary and every drafted reply before it is saved or sent
- Card data and authentication answers redacted or excluded from transcription
- No inference of agents' emotions
- Transcripts stored in region with a defined retention period
KPIs to instrument
- Average handling time and wrap up time, against a control group
- Summary accuracy on a weekly sample
- Suggestion acceptance rate per intent
- First contact resolution and customer satisfaction
- Time to proficiency for new agents
Human in the loop
The agent decides what is said and done on every contact and confirms summaries before they are saved. Team leaders review a sample of summaries and suggestions for accuracy, and knowledge owners fix content behind wrong suggestions.
Common failure modes
- Summaries nobody checks
- Agents approve summaries without reading them and errors enter the CRM. Sample and score them, and make edits easy.
- Suggestion noise
- Too many prompts distract agents mid call. Limit suggestions to what is relevant and measure acceptance.
- Card data in transcripts
- A payment call is transcribed and stored with the card number. Build redaction or pause and resume in from the start.
- Assist becomes surveillance
- Transcripts are reused to score individuals without disclosure or assessment. That changes the risk class and erodes trust.
What are the risks and rules?
EU AI Act
Depends on design
As a pure assist tool for agents it is minimal risk; the customer does not interact with the AI. It becomes high risk under Annex III point 4(b) if its data is used to monitor and evaluate individual agents' performance, and inferring agents' emotions at work is prohibited under Article 5(1)(f).
Guidance
- Article 5, prohibited AI practices (European Union, Europe). Point 1(f) prohibits AI systems that infer emotions of natural persons in the workplace, except for medical or safety reasons, which rules out emotion scoring of agents.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 4(b) covers systems that monitor and evaluate the performance and behaviour of workers.
- Artificial Intelligence (AI) Model Risk Management, information paper (Monetary Authority of Singapore, Asia Pacific). Example of supervisory good practice for testing and monitoring generative AI applications at banks.
Controls to put in place
- Data protection impact assessment covering call transcription and retention
- PCI DSS scoping of the transcription and storage path
- Written limits on the use of assist data for individual performance management
- Inventory entry with an accountable owner in customer operations
- Regular accuracy sampling of summaries and suggestions
Frequently asked questions
- How much handling time does agent assist save?
- Results vary with what is measured. Definity says automated call summaries cut three and a half minutes from each call, which Google Cloud's customer story puts at 33% of average handle time; Google Cloud's customer list reports a 20% cut for Definity's whole program, which also automated caller authentication, and a 15% efficiency gain at SEB. The NBER field study measured 14% more issues resolved per hour, and DBS expects up to 20% for its CSO Assistant once fully deployed, which is a forecast, not yet a result.
- Who benefits most?
- Less experienced agents. The NBER field study of 5,179 support agents found 14% more issues resolved per hour on average, with a 34% improvement for novice and low skilled workers and little effect on the most experienced.
- Is agent assist high risk under the EU AI Act?
- Not as an assist tool. It becomes high risk if the same data is used to evaluate individual agents (Annex III point 4(b)), and inferring agents' emotions at work is prohibited. Keep those uses separate and assessed.
How to cite this page
Blits.ai AI Use Case Library, "Real time AI assist for contact centre agents", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/live-agent-assist. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published