AI + HumanContact center for support teams

Support phones run best with AI and your agents working together

The AI takes status checks, policy questions and callback booking on the first ring. Your agents keep refunds, repeat callers and upset customers, which beats either one alone.

  • Resolution, not containment
  • Hand-offs that carry context
  • Every call recorded and scored
Support teams: Tuesday 4:50 pm · queue wait past 90 secondsAI agent on the line
AI agentResolution, not containment
OutcomeReship approved · notes drafted for one-click accept · call scored

Where is my order?

Order or case status, read from your order system and said back in seconds.

Callback when the wait is long

Queue past 4 minutes: book a 6 pm slot and text the confirmation.

Missing delivery, second call

A person approves the reship with the full story already in front of them.

Cancel, complaint or distress

Straight to an agent who has the authority and the care.

01 · Why AI + Human

Why support phones will run on AI and agents together

The phone is where hard issues land, and wrap-up takes 13.7% of agent time. In US centers, 8.9% of calls were abandoned in 2024 after a 99 second average wait. A people-only team cannot staff every peak. An AI-only line fails the other way: 83% of consumers repeat themselves after a transfer, and 87% say a way to reach a human is essential.

AI agentsLow cost per call, instant scale, consistent
  • Answers on the first ring, even at 4:50 pm with a long queue
  • Gives the same order and policy answer on every call
  • Books a callback and texts the slot while the queue stays full
On its ownCannot approve a reship, weigh a refund exception or calm a caller on their third call about one problem.
Your peopleJudgment, authority, care, new solutions
  • Approve reships, credits and exceptions on the spot
  • Read a frustrated caller and change the approach
  • Solve odd cases that fit no script, like a lost parcel
On their ownCannot staff every peak or every after-hours call, and lose time on lookups and typing notes.
Faster answers, and a person who already knows the storyOne desk, one record of every call
  • Overflow callers get an answer or a booked callback instead of abandoning
  • The agent opens with the summary, transcript and mood, so nobody repeats
  • Agents accept proposed notes in one click, and every call is QA-scored
Side by side
People only (traditional)
AI only
AI + Human
Answer speed
Callers wait when the queue is full
First ring, any hour
First ring, with a person one request away
Cost per routine call
Highest, one agent per call
Lowest, billed per second
Lowest on lookups, agent time kept for the rest
Consistency
Varies by agent and shift
Same answer every time, right or wrong
Same answers, and QA catches the wrong ones
Judgment calls
Strong: authority and tone
Weak: should not decide credits
Agents decide, with the AI's case notes
Solved with no repeat contact
Strong once reached, but long waits lose callers
Strong on lookups, weak on hard cases
AI finishes lookups, agents finish the rest
How to think about the split
Give the AI lookups, keep judgment

If the answer sits in one system of record, the AI owns it. If it needs authority, money or empathy, a person owns it.

Design the hand-off first

Decide the summary, transcript and mood an agent sees before the first AI call. Then audit 20 hand-offs a week.

Judge on resolution, not containment

A caller who gave up counts as contained. Track solved issues and 7-day repeat contacts for the AI and for agents alike.

The short answer

How should a support team combine AI and human agents on the phone?

Let the AI answer when the queue is full or closed, finish routine calls such as order status and callback booking, and hand everything else to a person with the summary, transcript and mood attached. Judge it on resolution, repeat contacts and hand-off quality, not containment. Start with overflow on one queue, score every call, and widen only when the thresholds you set in week one are met.

Reviewed November 2026 · 17 sources cited below
  • Phone (live agent plus self-service) is about 71% of inbound interactions in US centers (ContactBabel).
  • Wrap-up takes 13.7% of agent time, more than idle time at 12.2% (ContactBabel, 2025 guide).
  • US centers transferred 12.0% of calls in 2022, up from 8.2% in 2012 (ContactBabel).
  • 87% of customers say a way to reach a human is essential when GenAI is used (Gartner, 2026).
02 · The stakes

What a head of support is actually managing

The phone is still where hard issues land. The numbers below are why a good AI plan starts with the queue, the wrap-up and the hand-off, not with a chatbot demo.

13.7%

of US agent time is post-call wrap-up; idle time is 12.2%. ContactBabel US guide, 2025 edition.1

8.9%

of calls abandoned in US centers in 2024, up from 5.4% in 2012. Average speed to answer was 99 seconds. ContactBabel.1

83%

of consumers repeat themselves after an AI-to-human transfer; 96% of leaders think context survives. 2026 vendor CX survey, US, UK, Germany.2

4%

of service organizations reached the AI return they expected; 24% report any positive return. Gartner, 2026.3

03 · The whole journey

Where one support desk, AI and agents together, touches the customer

Support is not only the call that fixes a problem. The same desk shapes how a customer feels at first use, at every issue, at renewal and after they leave. Here is how the AI and your agents split each moment.

Pre-sale questionsSTAGE 1
Setup and first contactSTAGE 2
First-use helpSTAGE 3
Everyday issuesSTAGE 4
Billing and plan changesSTAGE 5
Renewal and at-riskSTAGE 6
Lapsed customersSTAGE 7
The moment
A buyer asks about returns before ordering, the queue is long, and they leave.
A new customer cannot find login or setup steps and assumes it is broken.
First use goes wrong and they call before giving the product a fair try.
Where is my replacement? The customer calls again and explains it all over.
A higher charge or plan swap triggers a call, and the answer is in two systems.
A customer with two bad contacts goes quiet and no one notices until they quit.
A customer left after an unresolved case and no one followed up on why.
AI agent
Answers policy and delivery questions on the first ring and books a callback with the right person.
Walks through account access, runs the identity check and sends the setup link by text.
Gives how-to answers from the knowledge base and flags repeated questions for the team.
Reads order and case status aloud, then hands off with a summary, transcript and mood.
Runs the identity check, explains the charge from the record and takes simple plan changes.
Flags negative mood and repeat contacts on scored calls, and reaches out to book a check-in.
Places an approved follow-up call, asks what went wrong and records the reason.
Your team
Takes the buyer with a complex need, like a bulk order or a compatibility question.
Takes failed checks and setup problems, with the steps already tried on screen.
Handles defects and confusion, and fixes the cause so the customer stays.
Approves reships, credits and exceptions, and calms a caller on a third contact.
Handles disputes and refunds over the limit, and offers a better-fit plan when it helps.
Calls the flagged customer, fixes the open issue and offers a fair option before renewal.
Takes the customer who is willing to talk, fixes the old case and invites them back.
What to track
Pre-sale calls answered; callbacks kept
Setup calls resolved on first contact
Early repeat contacts down
Resolution rate; repeat-yourself rate
Billing calls resolved; plan changes completed
Flagged customers contacted before renewal
Reasons captured; win-back calls booked

Read down a column for one stage, or across a row to see what the AI and your team each do from first contact to win-back.

04 · AI or person

Which support calls should the AI own, start, or never touch?

Sort your top 15 call reasons into three piles. Volume labels below are typical for a consumer-facing tier 1 line; tag your own calls for two weeks before you copy them.

AI handles it end to endAI starts, a person finishes with the contextStraight to a person
Work
Volume
Who takes it
Why
Order, case or ticket status
'Where is my replacement?' The AI reads status from the order or helpdesk system and says it.
high
AI handles it
One system of record, a lookup rather than a judgment, and the same answer every time.
Hours, policy and how-to questions
Warranty length, return window, how to reset the router, what the outage notice says.
high
AI handles it
Answers live in the knowledge base. Scoring every call catches wrong or stale answers quickly.
Callback booking when the queue is long
Wait is 4 minutes at 4:50 pm. The AI books a 6 pm callback slot and texts the confirmation.
medium
AI handles it
Turns an abandoned call into a scheduled one without adding a person to the floor.
Password, access and account changes
Reset request, address change, plan swap. The AI runs the standard identity check first.
high
AI, then a person
Low-risk steps can run. A failed check or a high-risk change goes to a person with what was tried.
Returns, exchanges and refunds
The AI captures order, reason and photo link. An agent approves exceptions or refunds over the limit.
medium
AI, then a person
The rules are clear and the exceptions are not. The person sees the full case and decides.
Billing questions and disputes
'Why is this charge higher?' The AI finds the invoice and explains it. A credit or dispute goes to billing.
medium
AI, then a person
Explaining is a lookup. Deciding a credit needs authority and the right tone.
Technical troubleshooting
The AI runs first steps (power cycle, status page, model check), then hands off with the steps already tried.
medium
AI, then a person
Technical calls resolve first time only 63% of the time (SQM). Never make callers repeat the steps.
Repeat caller on an open case
Third call about the same ticket, and the caller says so before the greeting ends.
low
Straight to a person
Trust is already spent. This caller needs one owner with authority, not another round of questions.
Complaints, cancellations, safety, distress
Cancel-or-keep conversations, injury reports, a caller in distress, anyone asking for a manager.
low
Straight to a person
Judgment and authority decide the outcome. Complaints resolve first time only 47% of the time (SQM).
TakeawayThe AI owns lookups and simple changes. People own judgment, money and emotion. Blend rows are where hand-off design decides whether callers feel helped or processed.
05 · One call, start to finish

One call, start to finish: the missing delivery

A caller reaches the AI because the queue is long, hits a case the AI should not decide, and lands with an agent who already knows the story. Here is the same call on one record.

0:00

Overflow, disclosure, recording notice

The queue wait passes 60 seconds and the call forwards. The AI says it is an AI, that the call is recorded, and that a person is one request away.

0:25

Verify and look up

The AI asks for name and ZIP, reads the order number back to confirm, and pulls status and prior contacts from your order system and helpdesk.

0:55

Recognize the limit

Delivered, not received, second contact. Your escalation rule says missing deliveries and repeat contacts go to a person. The AI does not try to talk the caller out of it.

1:10

Warm hand-off to the right skill queue

The agent sees a one-line summary, the transcript, the caller's mood (frustrated) and what the AI did. The caller does not repeat anything.

1:20

Agent resolves with AI Assist

The agent approves the reship while a live transcript runs beside the order. No typing of the story, no hunting for the policy.

5:40

Wrap-up and scoring

AI Assist proposes the disposition, callback time and notes, accepted in one click. The call is QA-scored automatically and a negative mood flags it for the supervisor.

06 · Start small

How to start: one queue, overflow only, eight weeks

Do not put the AI at the front door. Put it where callers would otherwise wait or hang up, measure for two weeks first, and write your pass marks before the first AI call.

Start with this

Overflow and after-hours calls on one queue: status and callback booking

Overflow callers would otherwise wait or abandon (8.9% of calls in US centers, 2024). Status and callback are lookups with a system of record, so the AI cannot make their experience worse.

How calls reach the AI

Forward one queue's overflow to a VoiceGenie number: when the longest wait passes 60 seconds, or after the queue closes. Your IVR and ACD stay as they are.

Weeks 1 to 2

Measure first

  • Record and score every call on the queue. Log baseline ASA, abandonment, transfers, FCR and wrap-up time.
  • Tag the top 15 call reasons for two weeks and pick the AI scope from the data.
  • Write escalation rules and the hand-off fields an agent sees: summary, transcript, mood, steps done.
  • Run 40 test calls with your own staff, including angry callers and callers who go off script.
Weeks 3 to 4

Overflow only

  • Forward the queue's overflow and after-hours calls. The AI handles status and callback booking only.
  • Every other reason goes to a person with the summary, or into a callback slot.
  • A supervisor listens live in week one. The team reviews every transferred or failed call each morning.
  • Fix knowledge gaps and tool errors daily. Change one thing at a time.
Weeks 5 to 6

Help the agents too

  • Turn on AI Assist for the agents on that queue: live transcript and proposed disposition and notes.
  • Compare wrap-up time with the week-one baseline. Ask agents which suggestions they edit and why.
  • Calibrate: have QA leads score 30 calls by hand and compare with the automatic scores.
  • Keep the AI scope unchanged so you can tell which change moved which number.
Weeks 7 to 8

Decide and widen

  • Hold the gate review against the pass marks you wrote in week one. Do not move them now.
  • If met, route the status IVR option to the AI in business hours, with 'agent' always reaching a person.
  • Add one more call reason, such as returns or policy questions, and repeat the cycle.
  • If not met, fix one named cause with a date. No decision by week eight is pilot purgatory.

Expand when all of these are true

  • Resolution on in-scope AI calls meets the bar you set in week one (planning range 55 to 75%).
  • 7-day repeat contact on AI-handled calls is no worse than your human baseline for the same reasons.
  • At least 90% of audited hand-offs reach the right queue with the caller repeating nothing.
  • CSAT on AI calls is within 3 points of human calls, and queue abandonment is below baseline.
07 · What to measure

Metrics a support leader can defend

Baselines are national figures, not yours. Capture your own in the first two weeks. Planning ranges are what to aim for with AI and people together, not promises.

Metric
Typical today
Planning range with AI and people
Answer speed and abandonment1
Average speed to answer for the human queue, plus share of offered calls abandoned before an answer. Measure in the ACD.
ASA 99 s; abandonment 8.9% (2024)
Abandonment under 5%
Containment rate
Share of all calls the AI finished with no transfer. Counts callers who gave up, so never judge on this alone.
No independent voice benchmark
25 to 40% of all calls
Resolution rate (in-scope calls)
Share of in-scope AI calls where the issue was solved and the caller did not contact you again about it within 7 days.
No independent voice benchmark
55 to 75%
Transfer rate and precision4
Share of calls transferred, and the share of AI transfers that reached the right queue first time. Audit 20 a week.
12.0% of calls transferred (2022)
90%+ reach right queue
Repeat-yourself rate2
Share of audited hand-offs where the caller restated what the AI already had. Score by listening, not by survey.
83% of consumers repeat themselves after a transfer
Under 10% of audited calls
First-call resolution5
Issue solved with no further contact on it. Measure by post-call survey and by 7-day repeat. Internal counts run high.
71% by survey; 78% center-reported
Hold or gain 1 to 3 points
After-call work per call1
Seconds from call end to the agent ready again, split by call reason. Pull from the ACD or the wrap-up timer.
Wrap-up is 13.7% of agent time (2025 guide)
One-third less on AI Assist calls
QA coverage and calibration6
Share of calls scored, and how closely automatic scores match a human reviewer on the same sample.
Often about 4 calls per agent per month
100% scored, 30 reviewed by hand

Planning ranges are targets to test in a pilot, not promises. Measure your own baseline for two weeks before you change anything.

08 · The return

What the numbers look like for a 16-person team

Capacity is the gain: calls answered, callers who reach a resolution instead of hanging up, and agent hours moved from typing to talking. Change the inputs to match your own queue.

A worked example. Move the sliders to match yours.

7 minaverage call length
$5,300loaded monthly cost per person
34 a daycalls one person handles
$15 · 50%value of one abandoned call recovered and resolved · share of recovered calls that become one
Estimated monthly gain
$37,116a month
Today, per call
$6.24
With VoiceGenie
$3.26
1,080more calls answered a month
540extra abandoned call recovered and resolveds a month
$8,100value of those outcomes
8,400calls your team still takes, with context

AI at $0.14 a minute with volume credit applied, plus Insights seats for the people who take handed-over calls. Phone line charges are your carrier's and are not included. Illustration, not a quote. Open the full calculator

Where these numbers come from

  • Call length: the average US service call ran 426 seconds in 2022 and about 423 seconds in 2024, so 7 minutes 41.
  • Cost per person: BLS median CSR pay of $21.53 an hour 7 times the 1.43 total-to-wages ratio in BLS compensation data 8 is about $64,000 a year, or $5,300 a month, before software and supervision.
  • Calls per person per day: 8 hours times 49.6% talk time 1, divided by 426 seconds, is about 34.
  • Missed calls: 9% matches the 8.9% abandonment rate in 2024 1. The 50% conversion is a placeholder; measure your own callback and resolution rate.
  • Outcome value: $15 is two avoided repeat contacts at the $7.20 average inbound call cost 9. It leaves out retention value on purpose.

What the dollar figure leaves out

  • New agents ramp faster: an AI assistant lifted novices' resolutions per hour 34% in a 5,179-agent study [s11].
  • Every call is scored, so coaching targets real patterns, not four lucky or unlucky calls.
  • Callers in a long queue get an answer or a booked callback instead of a ringing line.
09 · Rules to get right

Rules a support team has to get right

This is a practical guide, not legal advice. Support lines are mostly inbound, so recording, AI disclosure, card data and sector rules matter more than outbound calling rules. Confirm with counsel.

US states

Tell callers about recording first10

California, Pennsylvania, Florida, Massachusetts and other states need all-party consent to record. Consent must come before recording starts, and a vague 'may be monitored' line may not hold. Say it in the first sentence.

US states

Disclose the AI at the start11

Utah requires disclosure when a consumer clearly asks, and upfront in high-risk exchanges. Other states are adding rules. One opening meets all: say 'I'm an AI assistant' in the first line and offer a person at any time.

Global

Keep card numbers out of the AI path12

Card security codes must not be stored after authorization, including in recordings that can be queried, and card numbers pull the system into PCI scope. Use a secure pay-by-link or a hosted, masked keypad capture.

US

A bot that misleads can still create liability13

The CFPB's 2023 report warns that chatbots giving wrong answers, or trapping customers in loops with no route to a person, risk breaching consumer finance law. Lenders and banks should treat support AI as regulated.

US

Outbound callbacks use an AI voice14

The FCC treats AI-generated voices as artificial under the TCPA, so consent applies. A callback the customer requested is the lowest-risk kind; keep it informational and confirm the consent wording with counsel.

Four questions to ask any voice AI vendor, including us

  1. Is our call audio or transcript used to train any model, yours or a third party's? Put the answer in the contract.
  2. Can we export every recording, transcript, QA score and hand-off record, and set retention by queue?
  3. Show me what happens when the human queue is closed or an integration fails mid-call. Where does the caller land?
  4. Can I see resolution, hand-off and repeat-contact data by call reason, not one blended average?
10 · Where it plugs in

Systems your AI and agents need to share

VoiceGenie connects to HubSpot, Salesforce, Zoho and Google directly, and to the rest through the open API, webhooks or Zapier. These are the systems support teams typically wire in.

Helpdesk and ticketing

  • ZEZendesk
  • FRFreshdesk
  • SSSalesforce Service Cloud
  • HSHubSpot Service Hub
  • SEServiceNow
  • INIntercom

Phone system, ACD and IVR

  • ACAmazon Connect
  • TWTwilio
  • CICisco
  • AVAvaya
  • RIRingCentral
  • Any SIP trunk or PBX

Orders, billing and CRM

  • SHShopify
  • NENetSuite
  • STStripe
  • SASalesforce
  • HUHubSpot
  • ZOZoho

Knowledge and reporting

  • COConfluence
  • GUGuru
  • NONotion
  • SHSharePoint
  • LOLooker
  • PBPower BI
Keep your numbers and carrierTwilio, Plivo, Telnyx, Vonage or any SIP trunk or PBX. Put the AI behind one menu option, on overflow or after hours.
CRM and calendarsHubSpot, Salesforce, Zoho and Google connect directly.
Everything elseThe systems above connect through the open API, post-call webhooks or Zapier.
11 · One platform

One platform, so you set it up once and both sides get better

You do not need one vendor for the dialer, another for the AI agent, a third for the agent desk and a fourth for the tools that help agents. VoiceGenie is all four on one record. You connect your systems and load your knowledge once. Every input then improves the AI and your agents together, so set-up takes less of your time.

The usual way

Four tools, four set-ups

Dialerown contract · own integration
own training · own reports
Voice AI agentown contract · own integration
own training · own reports
Agent deskown contract · own integration
own training · own reports
Agent assist and QAown contract · own integration
own training · own reports
With VoiceGenie

One platform, set up once

AI agents
AI-native dialer
Human Agent Desk
AI Assist
Shared by the AI and your people
  • Your phone numbers and carrier
  • CRM and calendar connections
  • Knowledge and call rules
  • Recordings, QA and analytics
One link to helpdesk and orders

Link your helpdesk and order system once, through the open API, webhooks or Zapier. The AI and the agent read the same facts.

One knowledge base

Load policies once for the AI to answer from. A QA finding on a wrong answer becomes one fix to that article.

One recording and QA view

Every AI and agent call is recorded, transcribed and scored in one analytics view. A finding becomes a coaching note and a fix.

12 · What goes wrong

Five ways AI support projects go wrong

Reporting containment as success
Containment counts callers who gave up. Report resolution: solved, and no repeat contact within 7 days. Rising containment with rising repeats means the project is failing.
Putting the AI at the front door on day one
Start with overflow and after-hours on one queue. Move to the IVR option only after the pass marks are met. 'Agent' must always reach a person.
A hand-off that drops context
Define the payload: summary, transcript, mood, steps done. Audit 20 hand-offs a week. The rule is that nobody repeats what the AI already heard.
Calling four scored calls per agent 'QA'
Score every call automatically, flag by mood and topic, and have humans review the flagged calls plus a calibration sample. Coaching then targets patterns.
Cutting headcount before the results are in
Spend the first gains on overflow, wrap-up and training. Reduce through attrition once thresholds hold. Some leaders who cut early are now rehiring.
13 · Questions

Questions people ask before they start

What is the difference between containment rate and resolution rate?

Containment is the share of calls the AI finished without a transfer. Resolution is the share where the issue was actually solved and the caller did not contact you again about it. Containment counts hang-ups as wins, so judge the AI on resolution and 7-day repeat contacts, and use containment only as a volume figure.

Which support calls should the AI handle first?

Start with lookups that have one system of record: order or case status, policy and hours questions, and callback booking. They are high volume, easy to check, and low risk if the AI is wrong. Keep repeat callers, complaints, cancellations, safety and distress calls with people, and let the AI start the mixed ones.

Can I add AI to my support line without replacing my ACD or IVR?

Yes. VoiceGenie works with your own carrier and numbers, so you can forward one queue's overflow, an after-hours line, or a single IVR option to it. Your phone system stays in place. Callers still reach a person by asking, and the person sees the AI's summary and transcript when the call arrives.

How much does it cost?

The AI is $0.14 per minute, billed per second, phone line not included. A seven-minute call finished entirely by the AI is about $0.98. Human seats are per person per month on annual billing: Desk $29, Insights $59, Copilot $99. There is free trial credit and no minimum commitment. Enterprise pricing is private.

How long does set-up take?

Plan on two weeks of measuring and building before the first AI call, then a controlled overflow test for two weeks. The work is mostly yours: baselines, escalation rules, knowledge base cleanup and test calls. Eight weeks from first measurement to a go or no-go decision is a sensible plan for one queue.

Will callers hate talking to an AI?

Some will. A Gartner survey of 5,728 customers (fielded December 2023) found 64% would prefer companies not use AI in service. But Gartner's 2026 survey found 87% say a way to reach a human is essential, and a 2026 vendor survey found trust in AI at 55% with a clear human path versus 26% without. Keep a person one request away.

Do I have to tell callers it is an AI and that the call is recorded?

Do both in the first sentence. Recording needs consent before it starts in all-party states, and some states, such as Utah, require AI disclosure in set cases. One universal opening meets all of them: who is calling, that it is an AI assistant, that the call is recorded, and that a person is available on request.

Can AI really score every call for QA?

Every call can be recorded, transcribed and QA-scored automatically, against your scorecard, with AI and human calls in one view. Humans still matter: have QA leads review flagged calls and a calibration sample, such as 30 calls a month, to confirm the automatic scores match their judgment before you rely on them for coaching.

Try it on one call type

Hear it handle your calls before you change anything.

Talk to our AI agent in your browser, or start a trial and point one line at it. Your team keeps every call that needs them.

No credit card · No minimum commitment · Keep your carrier

Sources

  1. ContactBabel, The US Contact Center Decision-Makers' Guide 2025 (17th edition) (2025).
  2. CMSWire (reporting a Five9 survey), Customers: The AI Handoff Is Broken, Contact Center Leaders Won't Hear It (2026).
  3. CMSWire (reporting a Gartner survey), Just 4% of Customer Service Leaders Hit Their Expected Automation Returns (2026).
  4. ContactBabel, The US Contact Center Decision-Makers' Guide 2023 (15th edition) (2023).
  5. SQM Group, What Is a Good First Call Resolution Rate? (2022).
  6. Call Centre Helper, Industry standards: service level, AHT, occupancy, shrinkage, QA (n.d.).
  7. US Bureau of Labor Statistics, Customer Service Representatives: Occupational Outlook Handbook (2025).
  8. US Bureau of Labor Statistics, Employer Costs for Employee Compensation, Q2 2026 (2026).
  9. ContactBabel, The US Contact Center Decision-Makers' Guide 2026 (2026).
  10. Epstein Becker & Green, Eavesdropping by Algorithm: How Courts Are Applying Decades-Old Privacy Laws to AI (2026).
  11. Davis Polk, Utah scales back its generative AI consumer protection law (2025).
  12. PCI Security Standards Council, Protecting Telephone-based Payment Card Data (v2.0) (2011).
  13. Consumer Financial Protection Bureau, Chatbots in consumer finance (2023).
  14. Cooley, FCC: AI-generated robocalls are illegal under the TCPA (2024).
  15. Gartner, Gartner Survey Finds 87% of Customers Say Companies Using GenAI Must Provide Access to a Human Agent (2026).
  16. Gartner, Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI for Customer Service (2024).
  17. National Bureau of Economic Research, Generative AI at Work (Brynjolfsson, Li, Raymond) (2023).