AI voice agents for Chicago companies

How an AI voice agent works

An AI voice agent is software that takes part in a phone call: it listens, decides what to say or do, and speaks back in real time. "AI phone agent" names the same technology on a phone line. This page covers the technology and the rules. For the front-desk job built on it, see AI receptionist.

Most agents run four jobs in a loop: speech-to-text turns the caller's audio into words, a language model chooses a reply or a tool call, text-to-speech turns the reply into audio, and a turn-taking layer decides when the caller has finished and when to stop talking if interrupted. ElevenLabs lists exactly these parts, and Deepgram describes the same loop over one WebSocket connection.^[30],[24]

There are three ways to assemble that loop, and OpenAI's voice guide names all of them.

• Chained pipeline. Your application runs speech-to-text, the agent workflow and text-to-speech as separate stages, so you can store the transcript, run policy checks and swap any component.^[19]
• Speech-to-speech. One model hears the audio and answers in speech. OpenAI says skipping the text steps lowers latency and keeps the caller's tone and inflection available to the model.^[19],[18]
• Full duplex. A voice model listens and speaks at the same time and hands reasoning and tools to a separate backend, so callers can keep talking while it works.^{[19][19],[18]}

This trades control against speed; it is not a quality ranking. OpenAI advises choosing the audio architecture first, then designing tools and guardrails as for a text agent.^[19]

Latency, turn-taking and barge-in

People set the standard. In a study of question and answer pairs in ten languages, the mean gap between one speaker finishing and the other starting was 208 milliseconds, and in every language the most common response came within 200 ms. ITU-T G.114, the telecom delay recommendation, says keeping end-to-end delay under 150 ms gives essentially transparent interactivity for most applications, and that 400 ms should not be exceeded for general network planning.^[38],[37]

A voice agent adds recognition, model and speech time on top. Deepgram says its models are optimized for streaming transcription latency of 300 ms or less, and it separates transcript latency from end-of-turn latency, the time from the caller going quiet to an end-of-turn event. ElevenLabs lists about 75 ms for one text-to-speech model and about 150 ms for a real-time recognition model, excluding application and network latency. OpenAI gives no figures; it says to measure how long callers wait for a useful spoken answer, comparing the median with the 95th percentile.^{[27],[35],[19]}

Turn-taking is the harder half. Deepgram's agents start preparing a reply when recognition is moderately confident the caller has stopped, but release audio only once the turn is confirmed; if the caller was only pausing, the draft is discarded. Tool calls are the catch: by default they go out at once, so Deepgram advises holding anything that cannot be undone, such as ending a call, charging a card or sending a message, until the turn is confirmed.^[26]

Barge-in is the caller talking over the agent. OpenAI's Realtime API detects the speech and cancels the response in progress. On WebRTC and SIP (Session Initiation Protocol, the protocol for phone calls over the internet) it also trims the audio that was never played, so the model knows where it was cut off. On a WebSocket your code must stop playback and report how much audio the caller actually heard. Twilio's ConversationRelay has an interruptible setting (none, dtmf, speech or any) and an option that filters short acknowledgments such as "uh-huh".^{[18],[22],[20]}

Phone lines, transfers and tool connections

A voice agent reaches the phone network in one of a few ways, and the vendors document each.

• Telephony relay. Twilio's ConversationRelay connects the call to your server over a WebSocket. Your server gets the caller's words as text and sends text back to be spoken. You supply the model and the logic.^[22]
• SIP trunk to a model API. OpenAI documents pointing a SIP trunk at its endpoint over TLS; a webhook fires for each incoming call and you accept or reject it with an API call. Media runs over SRTP.^[20]
• SIP trunk or PBX to an agent platform. ElevenLabs connects an existing phone system for inbound and outbound calls, recommends TLS, required media encryption and digest authentication, and carries G.711 audio at 8 kHz or G.722 at 16 kHz.^{[31][22],[20],[31]}

Phone audio is narrow: Deepgram says upsampling 8 kHz audio gives no benefit. Echo cancellation stops the agent answering its own voice; noise suppression can reduce false interruptions but often lowers transcription accuracy, so A/B test on your own recordings.^[28]

Handing a call to a person is a telephony feature, not a prompt trick. OpenAI's SIP guide has a refer call that asks your SIP provider to move the call to a tel: or sip: destination. ElevenLabs documents conference, blind and SIP REFER transfers, and only some can read a summary to the person who picks up. Twilio lists a separate $0.1000 charge per refer on its SIP interface.^{[20],[32],[23]}

Outbound calls also reach voicemail. ElevenLabs' voicemail detection lets the agent leave a configured message or end the call. For advertising or telemarketing messages, the Federal Communications Commission (FCC) requires a toll-free number that connects to an automated opt-out.^[33],[2]

Tools are how the agent reads and changes your CRM, calendar or ticket system. In OpenAI's Realtime API the model emits a function call with JSON arguments, your application runs it and returns the result; Deepgram and ElevenLabs document the same pattern. Your application runs the tool, so that is where permissions are enforced.^{[18],[24],[30]}

• Does alone

  Answers from approved material, looks up a record, collects details and reads them back. Reads and low-risk steps.

• Waits for confirmation

  Anything that cannot be undone: a cancellation, a charge, a message sent for the caller.

• Goes to a person

  Requests outside its scope, upset callers, anyone who asks for a person. In our engagements, risky steps ask for approval and every action is logged.

The sample call below shows these pieces together: consent on file, a disclosure, an interruption and a hand-off.

Calling rules: FCC, FTC and Illinois

Direction matters. The Telephone Consumer Protection Act (TCPA) rules below govern calls a company makes. In its 2024 rulemaking notice the FCC wrote that the TCPA's requirements do not extend to technologies used to answer inbound calls. The recording and biometric rules in the next section still apply to inbound calls.^[7]

Outbound AI calls need consent. The FCC's Declaratory Ruling, adopted February 2, 2024 and released February 8, confirms that the TCPA's limits on an "artificial or prerecorded voice" cover current AI technologies that generate human voices. Callers need the called party's prior express consent, absent an emergency or an exemption, and the TCPA allows no carve-out for technology that purports to provide the equivalent of a live agent.^[1]

• Consent. Such calls to wireless numbers need prior express consent. If the call includes or introduces an advertisement or is telemarketing, it needs prior express written consent: a signed agreement, electronic signatures allowed, that names the number and says signing is not a condition of buying. The FTC adds that the seller cannot rely on third parties to obtain permission.^[2],[4]
• Identification and opt-out. The message must state at the start the business responsible, under its registered name, and give a callback number. For advertising or telemarketing calls, an automated voice or keypress opt-out must follow within two seconds, record the number and end the call at once.^[2]
• Revocation. Consent can be revoked by any reasonable method, an automated opt-out on the call counts, and the request must be honored within a reasonable time, not more than ten business days.^[2]
• Do not call and hours. No telephone solicitation to a residential subscriber on the national do-not-call registry, or before 8 a.m. or after 9 p.m. at the called party's location.^[2]
• Caller ID. Telemarketers must transmit caller ID, may not block it, and the number must accept do-not-call requests in business hours. Voice service providers must implement STIR/SHAKEN call authentication on their IP networks, subject to exceptions, so ask your carrier how your outbound numbers are signed.^{[5],[6][2],[4],[5],[6]}

The cost of getting it wrong. The TCPA lets a person sue for actual monetary loss or $500 in damages per violation, whichever is greater, and a court may raise the award up to three times for willful or knowing violations.^[3]

Illinois adds an autodialer statute. The Automatic Telephone Dialers Act bars operating an autodialer in Illinois between 9 p.m. and 9 a.m., requires it to disconnect within 30 seconds after the call ends, and says it may not impede caller ID. Playing a prerecorded message placed by an autodialer without the called party's consent is a violation, with three times actual damages plus fees and $500 in statutory damages per violation, and Attorney General enforcement.^{[10],[11],[9]}

The Act defines a recorded message as a taped communication soliciting the sale of goods or services without live voice interaction. Whether a two-way AI conversation fits that definition, and which exemptions apply, is a question for counsel.^[9]

Telling callers it is an AI. We did not find a final federal rule that requires it. The FCC proposed one in August 2024: callers using AI-generated voice would disclose at the beginning of each call that it uses AI-generated technology. We could not confirm a final order, so check docket 23-362. California has an enacted rule: AB 2905, approved September 20, 2024, requires the announcement on prerecorded dialer calls to tell the person called if the message uses an artificial voice.^[7],[8],[2]

We did not verify an Illinois statute on this point. We recommend saying it at the start of every call anyway; it costs one sentence.

Voiceprints and recordings in Illinois

Voiceprints. The Illinois Biometric Information Privacy Act (740 ILCS 14) lists a voiceprint among its biometric identifiers. A private entity may not collect one unless it first tells the person in writing that it is being collected, why and for how long, and receives a written release; an electronic signature counts. It must keep a public retention schedule, destroy the data when the purpose is met or within 3 years of the last interaction, whichever comes first, and not sell or profit from it.^[12],[13]

Remedies are $1,000 per negligent violation or $5,000 per intentional or reckless violation, or actual damages if greater, plus attorneys' fees. Since an amendment effective August 2, 2024, repeated collection of the same identifier from the same person by the same method is one violation with at most one recovery.^[14]

What to ask each vendor. Does any feature create or store a voiceprint, such as speaker identification, voice authentication or voice cloning? Where is that data kept, for how long, and who can reach it? Get the answers in writing before launch.

Recording. Illinois's eavesdropping statute (720 ILCS 5/14) makes it an offense to knowingly and intentionally use a recording device, in a surreptitious manner, to record a private conversation you are part of unless all other parties consent. "Surreptitious" means obtained or made by stealth or deception. A first offense is a Class 4 felony. One way to document consent is a short notice at the start of the call that the caller is speaking with an AI assistant and that the call is recorded, as in the sample call above. We have not verified court decisions applying the statute to AI voice recordings, so ask counsel.^{[15],[16],[17]}

Platform or custom build: what each approach gives you

Vendor names appear because we cite their public pages. They are not endorsements, and prices change.

Sources: ^{[18],[20],[21],[19],[24],[29],[30],[31],[34],[22],[23]}

Do not compare the per-minute figures head to head: GPT-Live's excludes backend usage, ElevenLabs' excludes the model and your carrier, and Twilio's sits on top of its voice minutes. A real estimate adds carrier minutes, numbers, model usage and recording, which Twilio lists at $0.0025 per minute.^{[21],[34],[23]}

Platform or custom? OpenAI points to the chained pipeline when you want control over each stage. Deepgram names its cost: glue code for turn-taking, interruptions and audio streaming, with each hop adding latency and failure modes. A platform trades some control for less to build and run. If the data is sensitive, ask each vendor where audio and transcripts are stored; self-hosted models are possible when privacy requires it.^[19],[24]

How to test and measure call quality

Test the conversation and the finished task. OpenAI's guide gives a booking example: listen to the confirmation, then check that the right appointment was saved. Save audio, events, tool results and application state for each run, and keep the caller, configuration, tools and transport fixed when comparing changes. Build up in stages: synthetic speech, then recordings of real callers, then a simulated caller for multi-turn conversations with interruptions, plus human listening for pacing. ElevenLabs also documents automated tests, A/B experiments and conversation analysis, so a platform may cover part of this.^[19],[30]

Sources: ^{[36],[19],[25],[22],[18],[2],[16]}

A pilot on a small share of real calls, with a person ready to take over and this scorecard in hand, is the right gate before a wider rollout.

When this is not a good fit

• No consent record for an outbound list. The FCC ruling requires prior express consent for AI-voice calls absent an emergency or an exemption. If you cannot show it for each number, do not automate the calls.^[1]
• Calls that need professional judgment. Medical, legal and financial questions should reach a qualified person. An agent can book that conversation; it should not replace it.
• Voiceprint authentication without legal review. It triggers the written notice, release and retention duties of the Illinois Biometric Information Privacy Act.^[13]
• Noisy audio you cannot improve. Microsoft rates call-center audio, often 8 kHz and noisy, as medium quality (under 30% word error rate) with deletion errors that can be high. If your test calls score worse, fix the audio or vocabulary first.^[36]
• Nobody to take a transfer. If no person is reachable during your call hours, a transfer ends in a dead end. Staff the handoff first.
• Very low call volume. Setup, testing and monitoring are fixed effort. With only a few calls a week, voicemail or a person is simpler.
• Front-desk answering or support queues. Those jobs are covered on the AI receptionist and AI customer service pages.

How an engagement works and what to prepare

1. Request

   You describe the calls: inbound or outbound, rough volume, the systems involved and the rules you already follow.

2. Short call and written scope

   We talk it through, then write a scope: the goal, the tools the agent may use, the steps that need approval and the success measure. Nothing is built before that.

3. Pilot

   A limited set of calls with a person ready to take over. The agent asks for approval on risky steps and logs every action.

4. Rollout and review

   Each agent gets its own service account with least-privilege access, runs on accounts you own and can be switched off. The scorecard is reviewed against real calls.

To make the scope concrete, bring or gather:

• Real calls or transcripts you may use, including the hard ones, for the test set
• The reasons people call and what the agent may do for each
• The systems it must reach (CRM, calendar, ticketing) and who owns their admin accounts
• Your phone provider and numbers, and whether they can route to a SIP trunk or a relay
• For outbound calls: where each number came from, the consent wording and date, and your do-not-call process
• Who takes transferred calls, at what hours, and the recording and AI statements counsel has reviewed

Questions answered

Sources

1. FCC 24-17, Declaratory Ruling, CG Docket No. 23-362 (adopted Feb 2, 2024; released Feb 8, 2024) (Federal Communications Commission), accessed 2026-10-12. AI-generated voices are an artificial or prerecorded voice under the TCPA; consent; no carve-out for live-agent equivalentshttps://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf
2. 47 CFR 64.1200, Delivery restrictions (eCFR, text version dated 2026-10-01) (Electronic Code of Federal Regulations (eCFR)), accessed 2026-10-12. Consent, written consent, identification, opt-out within two seconds, voicemail callback number, revocation, do-not-call and calling hourshttps://www.ecfr.gov/current/title-47/part-64/section-64.1200
3. 47 U.S.C. 227, Restrictions on use of telephone equipment (subsection b(3) and b(4)) (Legal Information Institute, Cornell Law School (reproduces the U.S. Code)), accessed 2026-10-12. Private right of action: $500 per violation or actual loss, up to 3 times for willful or knowing violations; FCC forfeitureshttps://www.law.cornell.edu/uscode/text/47/227
4. Complying with the Telemarketing Sales Rule (FTC business guidance) (Federal Trade Commission), accessed 2026-10-12. Prerecorded telemarketing calls need prior signed written agreement; seller cannot rely on third-party permissionhttps://www.ftc.gov/business-guidance/resources/complying-telemarketing-sales-rule
5. 47 CFR 64.1601, Delivery requirements and privacy restrictions (caller ID for telemarketers) (Electronic Code of Federal Regulations (eCFR)), accessed 2026-10-12. Telemarketers must transmit caller ID, may not block it, and the number must permit do-not-call requests in business hourshttps://www.ecfr.gov/current/title-47/part-64/section-64.1601
6. 47 CFR 64.6301, Caller ID authentication (STIR/SHAKEN) (Electronic Code of Federal Regulations (eCFR)), accessed 2026-10-12. Voice service providers must implement STIR/SHAKEN in their IP networks; they make attestation-level decisionshttps://www.ecfr.gov/current/title-47/part-64/section-64.6301
7. FCC 24-84, Notice of Proposed Rulemaking and Notice of Inquiry, CG Docket No. 23-362 (adopted Aug 7, 2024) (Federal Communications Commission), accessed 2026-10-12. Proposed AI call disclosure and consent-form wording; FCC statement that TCPA requirements do not extend to answering inbound callshttps://docs.fcc.gov/public/attachments/FCC-24-84A1.pdf
8. California AB 2905 (Chapter 316, Statutes of 2024), amending Public Utilities Code 2874 (California Legislative Information), accessed 2026-10-12. Example of an enacted state rule: announcement must inform the person called if the prerecorded message uses an artificial voicehttps://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240AB2905
9. 815 ILCS 305/5, Automatic Telephone Dialers Act, Definitions (Illinois General Assembly), accessed 2026-10-12. Definitions of autodialer and recorded messagehttps://www.ilga.gov/Documents/legislation/ilcs/documents/081503050K5.htm
10. 815 ILCS 305/15, Automatic Telephone Dialers Act, Method of operation (Illinois General Assembly), accessed 2026-10-12. No autodialer calls between 9 p.m. and 9 a.m.; disconnect within 30 seconds; no impeding caller IDhttps://www.ilga.gov/Documents/legislation/ilcs/documents/081503050K15.htm
11. 815 ILCS 305/30, Automatic Telephone Dialers Act, Violations (Illinois General Assembly), accessed 2026-10-12. Playing a prerecorded autodialer message without consent is a violation; damages; Attorney General enforcementhttps://www.ilga.gov/Documents/legislation/ilcs/documents/081503050K30.htm
12. 740 ILCS 14/10, Biometric Information Privacy Act, Definitions (Illinois General Assembly), accessed 2026-10-12. "Biometric identifier" includes a voiceprint; "written release" includes an electronic signaturehttps://www.ilga.gov/Documents/legislation/ilcs/documents/074000140K10.htm
13. 740 ILCS 14/15, Biometric Information Privacy Act, Retention; collection; disclosure; destruction (Illinois General Assembly), accessed 2026-10-12. Written retention policy, written notice and release before collection, no sale or profithttps://www.ilga.gov/Documents/legislation/ilcs/documents/074000140K15.htm
14. 740 ILCS 14/20, Biometric Information Privacy Act, Right of action (Illinois General Assembly), accessed 2026-10-12. Liquidated damages of $1,000 (negligent) or $5,000 (intentional or reckless); single-recovery rule since 2024https://www.ilga.gov/Documents/legislation/ilcs/documents/074000140K20.htm
15. 720 ILCS 5/14-1, Eavesdropping, Definitions (Illinois General Assembly), accessed 2026-10-12. Definitions of eavesdropping device, private conversation and surreptitioushttps://www.ilga.gov/Documents/legislation/ilcs/documents/072000050K14-1.htm
16. 720 ILCS 5/14-2, Eavesdropping, Elements of the offense; affirmative defense (Illinois General Assembly), accessed 2026-10-12. Recording a private conversation to which you are a party, in a surreptitious manner, needs consent of all other partieshttps://www.ilga.gov/Documents/legislation/ilcs/documents/072000050K14-2.htm
17. 720 ILCS 5/14-4, Eavesdropping, Sentence (Illinois General Assembly), accessed 2026-10-12. First offense is a Class 4 felonyhttps://www.ilga.gov/Documents/legislation/ilcs/documents/072000050K14-4.htm
18. OpenAI, Realtime conversations guide (VAD, interruption and truncation, function calling) (OpenAI), accessed 2026-10-12. Voice-to-voice design, default voice activity detection, cancelling and truncating on interruption, WebRTC/SIP vs WebSocket, function-call flowhttps://developers.openai.com/api/docs/guides/realtime-conversations
19. OpenAI, Voice agents guide (architecture choices and evaluation) (OpenAI), accessed 2026-10-12. Three architectures (GPT-Live full duplex, Realtime, chained pipeline); evaluation method, latency measurement advicehttps://developers.openai.com/api/docs/guides/voice-agents
20. OpenAI, SIP guide for the Realtime API (accept, reject, refer, hang up) (OpenAI), accessed 2026-10-12. SIP trunk to the API over TLS, webhook per incoming call, SRTP media, refer (transfer) to a tel: or sip: targethttps://developers.openai.com/api/docs/guides/realtime-sip
21. OpenAI API pricing (Realtime and GPT-Live sections) (OpenAI), accessed 2026-10-12. Price units: per 1M tokens (gpt-realtime-2.1 audio $32.00 input, $64.00 output); GPT-Live $0.05 per minute billed per second, backend usage separatehttps://developers.openai.com/api/docs/pricing
22. Twilio, ConversationRelay TwiML noun reference (Twilio), accessed 2026-10-12. ConversationRelay: speech-to-text and text-to-speech over a WebSocket to your own AI app; interruptible and ignoreBackchannel settingshttps://www.twilio.com/docs/voice/twiml/connect/conversationrelay
23. Twilio Voice pricing, United States (Twilio), accessed 2026-10-12. US local calls $0.0140/min outbound and $0.0085/min inbound, local number $1.15/month, refer $0.1000, recording $0.0025/min, ConversationRelay $0.07/minhttps://www.twilio.com/en-us/voice/pricing/us
24. Deepgram, Voice Agent API architecture overview (Deepgram), accessed 2026-10-12. Single-WebSocket speech-to-text, LLM and text-to-speech loop; end-of-turn, barge-in, function calling; telephony providers; the cost of stitching componentshttps://developers.deepgram.com/docs/voice-agent-architecture
25. Deepgram, Voice Agent latency report (Deepgram), accessed 2026-10-12. Per-turn latency breakdown across speech-to-text, model and text-to-speechhttps://developers.deepgram.com/docs/voice-agent-latency-report
26. Deepgram, Speculative replies and turn confirmation (Deepgram), accessed 2026-10-12. Reply prepared before the turn is confirmed; audio held until confirmation; irreversible tool calls should waithttps://developers.deepgram.com/docs/voice-agent-speculative-replies
27. Deepgram, Measuring STT latency (Deepgram), accessed 2026-10-12. Transcript latency vs end-of-turn latency; vendor figure of 300 ms or less for streaming transcriptionhttps://developers.deepgram.com/docs/measuring-streaming-latency
28. Deepgram, Audio preprocessing and barge-in guide (Deepgram), accessed 2026-10-12. Echo cancellation and noise suppression trade-offs; 16 kHz sweet spot; upsampling 8 kHz gives no benefit; A/B test on your own audiohttps://developers.deepgram.com/guides/deep-dives/audio-preprocessing-barge-in
29. Deepgram pricing (Deepgram), accessed 2026-10-12. Voice Agent API $0.075/min Standard pay-as-you-go (WebSocket connection time); text-to-speech priced per 1,000 characters; streaming speech-to-text per minutehttps://deepgram.com/pricing
30. ElevenLabs, ElevenAgents overview (ElevenLabs), accessed 2026-10-12. Agent components (speech recognition, LLM, text-to-speech, turn-taking model); telephony options; tools, knowledge base, testing and experimentshttps://elevenlabs.io/docs/eleven-agents/overview
31. ElevenLabs, SIP trunking for agents (ElevenLabs), accessed 2026-10-12. PBX or SIP system to agents for inbound and outbound calls; TLS, SRTP, digest authentication; G.711 8 kHz or G.722 16 kHzhttps://elevenlabs.io/docs/eleven-agents/phone-numbers/sip-trunking
32. ElevenLabs, Transfer to number system tool (ElevenLabs), accessed 2026-10-12. Three transfer types: conference, blind, SIP REFER; warm transfer message only on somehttps://elevenlabs.io/docs/eleven-agents/customization/tools/system-tools/transfer-to-number
33. ElevenLabs, Voicemail detection system tool (ElevenLabs), accessed 2026-10-12. Agent can detect voicemail and leave a configured message or end the callhttps://elevenlabs.io/docs/eleven-agents/customization/tools/system-tools/voicemail-detection
34. ElevenLabs, ElevenAgents pricing (ElevenLabs), accessed 2026-10-12. Billed per call minute: $0.08 per additional minute; LLM billed separately; carrier billed by your providerhttps://elevenlabs.io/pricing/agents
35. ElevenLabs, Models overview (speech-to-text and text-to-speech models) (ElevenLabs), accessed 2026-10-12. Component latency figures (about 75 ms for a text-to-speech model, about 150 ms for a real-time speech recognition model) marked as excluding application and network latencyhttps://elevenlabs.io/docs/overview/models
36. Microsoft Learn, Test accuracy of a custom speech model (word error rate) (Microsoft), accessed 2026-10-12. Definition of word error rate, amount of test audio, vendor rule of thumb for WER bands, call-center audio profilehttps://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data
37. ITU-T Recommendation G.114, One-way transmission time (International Telecommunication Union), accessed 2026-10-12. About 150 ms mouth-to-ear delay gives essentially transparent interactivity; 400 ms is the network planning limithttps://www.itu.int/rec/T-REC-G.114
38. Stivers et al., Universals and cultural variation in turn-taking in conversation (PNAS, 2009) (Proceedings of the National Academy of Sciences (via PubMed Central)), accessed 2026-10-12. Mean response offset of 208 ms in question and answer pairs across 10 languages; most responses within 200 mshttps://pmc.ncbi.nlm.nih.gov/articles/PMC2705608/

How we checked this page

Last verified 2026-10-12 · Media Express LLC

General information, not legal, financial or security advice. Rules and vendor terms change; check the cited source before you rely on a figure.

Related

• AI receptionist for Chicago businessesAI receptionist
• AI customer service for Chicago companiesAI customer service
• AI development company for Chicago teamsAI development
• AI consulting for Chicago companiesAI consulting