Skip to main content
Sales & strategy

Latency in Sales Calls: Why 500ms of Delay Costs You Orders

Sales calls with low AI latency are completed far more often. See what delay costs in sales, and what the technical standard is in 2026.

Author: Agent360 editorial team
Published: 10 April 2026
Last updated: 7 November 2026
Reading time: 18 min read

Why does a short response time matter so much in AI-based sales calls?

If there's a noticeable delay in the response, the customer loses trust in the technology immediately, even if the answer is entirely correct.

When your AI assistant responds 500 milliseconds too late, you lose the order. Not because the answer is wrong, but because the customer has already decided that the technology doesn't work. In sales calls, latency isn't a technical problem: it's a revenue problem. AI systems with a low response time complete markedly more conversations than systems with a noticeable delay.

TL;DR: the key points about AI latency in sales:

  • Latency above 500ms is physically felt by the customer and breaks the natural flow of the conversation
  • Sub-300ms response times give a markedly higher completion rate
  • Latency is an engineering discipline, not a feature you can add on afterwards

What is AI latency, and why does it matter for sales?

AI latency is the time that passes between a user finishing speaking and the AI system starting to respond. Measured in milliseconds, it is the difference between a conversation that feels natural and one that feels broken.

In human conversation, a pause of 100-200 milliseconds is entirely normal. The human brain processes the language and formulates a reply. It's the rhythm we are evolutionarily programmed to accept. We know it. We expect it.

But beyond 500ms, something fundamentally different happens. The customer starts to read the pause as a fault. "Has the system gone down?" "Is there a connection problem?" "Does this even work?" This cognitive process (called uncertainty anxiety in the psychological literature) kicks in automatically and takes control of the conversation away from your salesperson.

In B2C telesales, where the average call lasts under three minutes, 500ms of latency isn't just irritating. It's fatal.

Last updated: March 2026

The human psychology behind the latency problem

To understand why latency costs orders, we need to understand how people perceive pauses in conversation.

Natural conversational rhythm: 200ms is the limit

See also our guide to AI voice agents.

According to Gartner's research on conversational AI, real-time response for conversational AI systems is defined as under 200ms in order to maintain a natural conversational flow. This figure isn't arbitrary; it reflects the physiological limit at which the human auditory system starts to perceive a pause as unnatural.

Beyond 200ms, the brain starts looking for explanations. Beyond 500ms, it starts drawing conclusions. In a sales context, those conclusions are rarely positive.

The loss of trust is instant

According to Salesforce's State of the Connected Customer report, 88% of customers expect companies to use modern technology properly. When the AI system stutters or pauses for too long, the customer reads it as incompetence: not as a technical problem, but as a signal that the company doesn't have its systems under control.

In a sales call, where trust is the primary currency, this is the fastest route to a "no thanks".

What Google's research shows about user patience

According to Google's research on page speed and user behaviour, 53% of mobile users leave a page if it takes more than 3 seconds to load. Translated to voice AI: the tolerance threshold for delay is lower than most technology leaders assume. And in a conversation structured as a dialogue rather than a page error, the cognitive effect of delay is amplified, because it's personal. It isn't an error on a screen. It's a pause in a human interaction.

What exactly happens above 500ms of latency?

Here is the problem made concrete. In a typical sales call with AI support, whether it's a real-time assist system, an AI meeting booker or a post-call analysis system, the following chain of reactions unfolds at 500ms+ of latency:

Second 0-0.2: The customer finishes their sentence. The expectation is set.

Second 0.2-0.5: The customer waits. This feels normal. But the uncertainty subtly begins.

Second 0.5-1.0: The customer starts to doubt. "Did the system hear me?" The voice is raised, the question repeated, or the call cut short.

Second 1.0+: Trust is broken. The customer is now in defensive mode. The likelihood of a purchase drops markedly.

This isn't theory. It is reproducible behaviour observed across hundreds of sales calls.

What real sales calls tell us about latency

Real-time conversation analysis is only effective if it actually happens in real time. The consequences of latency typically look like this:

  • Conversations with low latency (under 300ms) are completed markedly more often than conversations with 500ms+
  • Conversations with latency above 500ms are more often cut short early: the customer hangs up before the sales point has been reached
  • Order size is typically lower in conversations with high latency, because the salesperson loses momentum and the closing technique breaks down

See also our guide to AI in the sales department.

Sub-300ms response time isn't a selling point. It's a minimum requirement.

The latency consequence table: what each interval costs you

Latency intervals and their typical consequences in sales calls (illustrative)

Latency interval User experience Impact on the conversation Completion rate
< 200ms Completely natural. The customer notices no delay. Optimal flow. Sales momentum intact. Baseline (100%)
200–300ms Acceptable. Marginally noticeable, but tolerated without irritation. Minimal impact. The conversation continues naturally. ~97% of baseline
300–500ms Noticeable delay. The customer starts to adapt. Rhythm is broken. The salesperson has to actively restore the flow. ~82% of baseline
500ms–1 sec. Frustrating. The customer reads it as a system fault. Trust drops markedly. A cut-short call is highly likely. ~66% of baseline
> 1 second Unacceptable. The customer assumes a technical breakdown. The conversation is effectively lost. The customer has mentally left. <40% of baseline

Why latency is an engineering discipline, not a feature

Here is the critical point that many companies overlook when evaluating AI systems for sales:

Latency can't be fixed with a software update or an extra layer of caching. It is determined by the architecture.

The three technical causes of high latency

1. Model inference time Large language models (LLMs) take time to generate responses. A model with 70 billion parameters on shared cloud infrastructure can have inference times of 400-800ms alone, before network delay and audio processing are added.

2. Audio pipeline latency Speech-to-text conversion, AI processing and text-to-speech synthesis are three separate operations. If they are architecturally designed to run sequentially rather than in parallel, the latency adds up. It's not unusual to see systems with 1.2-1.8 seconds of total latency because the pipeline is built as a waterfall model.

3. Network and infrastructure latency Cloud services hosted in the US (US-East) typically have 80-120ms more latency for European users than EU-hosted alternatives. For a Danish sales call this isn't trivial; it's a structural disadvantage built into the infrastructure from day one.

What sub-300ms requires architecturally

Delivering consistent sub-300ms latency in voice AI for sales calls requires:

  • Streaming inference instead of batch generation: the system starts speaking while it is still generating the rest of the answer
  • A parallel audio pipeline: speech-to-text and LLM inference overlap instead of running sequentially
  • EU-local inference: hosting the model in the EU markedly reduces the network round trip
  • Optimised models: distilled models (7B-13B parameters) with acceptable accuracy trade-offs are fundamentally faster than frontier models for standard sales use cases

These choices are made on the architecture drawing board. Not in a settings panel.

Real-time assistance is binary

AI assistance in a sales call is either real-time or irrelevant.

There is no middle category. A system that delivers insights to the salesperson 800ms after the customer has spoken delivers them after the fact, in a context that has already moved on. The salesperson can't use the insight. The customer experiences confusion. The result is the opposite of what was intended.

This isn't an observation about usability. It's an observation about fundamental sales timing.

In a typical transactional B2C sales call there are 3-4 decisive moments where a precise AI input can change the direction of the conversation:

  1. When the customer's main objection arises
  2. When the salesperson presents the main offer
  3. When the customer signals doubt (a shift in tone, a pause, a "hmm")
  4. When the salesperson tries to close

These moments don't last a second. They last half of one. An AI system with 500ms+ of latency never reaches them. It's like trying to catch a ball three seconds after it has landed.

Latency in practice: three scenarios from Danish telesales

Scenario 1: Insurance sales (B2C, high volume)

A telesales rep at a Danish insurance company makes 80 calls a day. The AI system is configured to give real-time suggestions for objection handling. With 600ms of latency, this happens:

  • The customer says: "That sounds expensive."
  • The AI processes for 600ms.
  • During that time the rep tries to improvise and picks a suboptimal response.
  • The AI's suggestion arrives. The rep sees it, but the moment has already been missed.

The result: the AI system exists but adds no value. The rep loses trust in the system and starts ignoring it. The IT investment is wasted.

Scenario 2: B2B SaaS (AI meeting booker)

An AI meeting booker (such as JesperAI) calls a lead to book a demo. The lead is positive but asks a specific question about integration with HubSpot.

With 500ms+ of latency:

  • The customer asks the question.
  • The pause arises. The customer thinks the line has gone dead.
  • The customer says: "Hello? Is anyone there?"
  • The conversation falls out of flow. Trust is broken.
  • The booking rate drops markedly for this type of call.

With sub-300ms latency:

  • The customer asks the question.
  • The AI answers naturally within the expected window.
  • The conversation continues. The HubSpot integration is confirmed. The meeting is booked.

Scenario 3: Energy company (high-volume outbound, B2C)

A Danish energy company makes 500+ outbound calls a day with AI assist. Latency problems averaging 700ms cost them:

  • Markedly fewer completed conversations
  • A direct loss: if 100 conversations result in 12 sales at low latency, they result in only ~8 sales at high latency
  • At an average order value of DKK 2,400 a year (typical for energy agreements), that equates to DKK 9,600 lost per 100 calls
  • Over a week with 500 calls a day: DKK 240,000 in lost revenue every week, from a single technical parameter

How to evaluate AI systems on latency

Many companies buy AI systems on the basis of feature lists. Latency is rarely mentioned, and when it is, the figures are typically "laboratory figures" measured under ideal conditions.

Here are the questions you should ask any AI vendor before signing:

The five latency questions

1. What is the P95 latency under normal load? P50 (the median) tells you what the system does when it's happy. P95 tells you what it does under pressure. A P95 above 500ms is a red flag.

2. Where are the inference servers located? EU hosting versus US hosting can mean an 80-150ms difference in network latency alone. For a Danish company with a Danish customer base, this is relevant.

3. Streaming or batch generation? Batch generation = the system waits to respond until the whole answer has been generated. Streaming = the system starts speaking while it is generating. Streaming is fundamentally faster for conversational use.

4. Can you show real-world data from the production environment? Lab performance and production performance diverge. Demand documentation from a comparable production environment.

5. What happens to latency at 10x load? Scaling is the fastest way to expose architectural weaknesses. A system that performs well at 10 simultaneous conversations but struggles at 100 isn't production-ready for enterprise use.

Conversation intelligence: from reactive to proactive AI

To understand why latency is so critical, it helps to understand what conversation intelligence is actually designed to do.

Conversation intelligence isn't just a transcription service. It's a real-time system that identifies:

  • The customer's sentiment (positive, negative, neutral)
  • Specific objections and their category
  • Momentum points in the conversation
  • Compliance risks (in regulated industries)

All of this is only useful, however, if it's delivered before the critical window has closed. A sentiment analysis that arrives 800ms after the customer's statement isn't conversation intelligence. It's conversation history, and that is something fundamentally different.

This distinction (between intelligence and history) is the core of why latency isn't a comfort problem. It's a product-competence problem.

Latency and the interplay between salesperson and AI

Human and AI work best side by side during the conversation. The salesperson handles the relationship, the empathy and the complex negotiation nuances. The AI handles information processing, objection categorisation and compliance checks.

For this interplay to work, the AI has to be invisible. Not invisible in the sense that the salesperson doesn't know it's there, but invisible in the sense that its presence doesn't disturb the natural rhythm of the conversation.

That is exactly where latency destroys the model. An AI system with a 600ms response time isn't an invisible assistant. It's a disruptive factor that actively works against the salesperson's performance.

According to research from Forrester on AI in sales organisations, the main reason AI sales tools fail to be adopted isn't missing functionality; it's the user experience. Salespeople who experience the AI as slow and disruptive ignore it. The AI investment is wasted.

Sub-300ms latency is the precondition for that interplay being possible at all.

Technical infrastructure: what sub-300ms requires in practice

Let me make the architectural requirements concrete for a company evaluating AI infrastructure for sales.

Hardware level

  • GPU-accelerated inference: CPU-based inference is 3-5x slower than GPU. For voice AI in sales calls, GPU isn't a luxury choice; it's a requirement.
  • Dedicated vs. shared capacity: Shared cloud infrastructure (typically offered by SaaS vendors) shares compute capacity with other customers. Under peak load, latency rises. Dedicated capacity eliminates this problem.

Model level

  • Distilled versus frontier model: GPT-4o, Claude 3 Opus and comparable frontier models are extraordinarily capable, but they aren't designed for sub-300ms response times in conversational applications. Distilled models (7B-13B parameters), fine-tuned for specific sales use cases, deliver 80-90% of the required capability with 4-6x lower latency.
  • Streaming tokens: Modern inference frameworks support token streaming: the system starts generating audio while it is still generating text. This alone can reduce perceived latency by 40-60%.

Network level

  • CDN distribution: For companies with many simultaneous conversations, a globally distributed inference network can markedly reduce geographic latency.
  • WebRTC vs. SIP: Modern voice communication over WebRTC has markedly lower latency than the traditional SIP protocol. For AI systems that integrate with telephony infrastructure, the choice of protocol is latency-critical.

Latency as a competitive parameter in 2026

In 2026, AI functionality in sales tools is more or less a commodity. Most systems can transcribe calls, identify objections and generate follow-up drafts.

What differentiates isn't what the system does; it's when it does it.

This is exactly the shift we see in the market: infrastructure (not features) is the primary competitive factor for AI-driven sales platforms in 2026.

Latency is the most concrete manifestation of this. You can't market your way out of a slow system. You can't compensate with a better UI. Either the system works in real time, or it doesn't work in a sales context.

What is the latency problem costing you? A worked example

Let's make it concrete for a mid-sized Danish sales department.

Assumptions:

  • 20 salespeople/telesales reps
  • 50 outbound calls per rep per day
  • 1,000 calls a day in total
  • Average order value: DKK 3,000
  • Baseline close rate: 8% (80 orders a day)

Scenario A: sub-300ms latency

  • 80 orders × DKK 3,000 = DKK 240,000 in daily revenue

Scenario B: 500ms+ latency

  • Assume the completion rate falls by a third: 1,000 calls × 66% = 660 completed conversations
  • The effective close rate is reduced: 660 × 8% = 52.8 orders
  • 52.8 orders × DKK 3,000 = DKK 158,400 in daily revenue

Daily loss: DKK 81,600 Monthly loss: ~DKK 1,630,000 Annual loss: ~DKK 19,600,000

An extra 200ms of latency is far from free. For a mid-sized sales department it costs close to DKK 20 million a year.

Seen in this light, investing in the right AI infrastructure isn't an IT expense. It's revenue protection.

FAQ

Q: What is acceptable AI latency in a sales call? Under 300ms is the acceptable range for AI assist in sales calls. Under 200ms is optimal and isn't noticed by the customer. Above 500ms it breaks the natural flow of the conversation and reduces trust in the system. The technical benchmark for conversational AI is under 200ms to maintain a natural dialogue.

Q: Can high latency be compensated for with better AI answers? No. The quality of the answer is irrelevant if it arrives too late. In a sales call, timing and context are inseparable. A perfect objection-handling answer that arrives 800ms after the objection is useless; the context has already moved on. Latency and answer quality are parallel requirements, not substitutes for each other.

Q: How does latency affect AI meeting bookers like JesperAI? For AI meeting bookers like JesperAI, latency is even more critical than for AI assist systems. An AI meeting booker conducts the conversation autonomously; there is no human salesperson who can compensate for a pause. Delay above 400ms in an autonomous AI conversation typically results in the customer interrupting or questioning whether the system works.

Q: What is the difference between P50 and P95 latency? P50 latency is the median: half of the calls are faster, half are slower. P95 latency is the level that 95% of calls fall under. P95 is the relevant figure for sales applications because it tells you what happens in worst-case scenarios under load. A system with a P50 of 200ms and a P95 of 900ms delivers a user experience that is unacceptable 5% of the time; in a call centre with 1,000 calls a day, that's 50 conversations a day that break down.

Q: Can I test latency before I buy an AI system? Yes, and you should always do so. Ask the vendor for access to a sandbox environment, and measure latency under simulated load (not just one user). Use tools such as WebRTC latency testers or simple timestamping scripts in the API integration. Demand P95 latency data from a production environment with comparable load. Numbers are easier to lie about than a live system under test.

Q: What does it cost to build AI infrastructure with sub-300ms latency? It depends fundamentally on the architecture choices. GPU-accelerated inference infrastructure in the EU with dedicated capacity and a streaming pipeline typically costs 3-5x more to build and host than a standard cloud SaaS solution. But as the worked example in this article shows, a mid-sized sales department with latency problems is potentially exposed to DKK 15-20 million in lost revenue a year. The infrastructure investment pays for itself in the first quarter.

Latency and GDPR: an unexpected connection

There is an aspect of the latency discussion that is rarely mentioned: GDPR compliance and latency are architecturally linked.

Compliant AI in sales calls requires:

  1. Consent first: Consent must be obtained before recording begins
  2. EU hosting: All data (including audio streams and transcriptions) must stay in the EU
  3. Data minimisation: Only necessary data is stored

Point 2 is relevant to latency. A system that sends audio data to US-based inference servers to save on infrastructure costs breaches GDPR. And as a side effect, it has 80-150ms of extra latency.

GDPR-compliant AI infrastructure with EU hosting is therefore not just a legal necessity for Danish companies. In itself, it is a latency advantage. The two requirements point in the same direction.

Summary: latency as an infrastructure decision

Let's pull together the key points.

Latency isn't a feature. It is an architecturally determined property of an AI system. It can't be patched. It can't be added on afterwards. It is the result of hundreds of technical decisions made at the design stage.

500ms is the limit for functional selling. Beyond this interval, customers start to read pauses as faults. Trust falls. The completion rate falls. Revenue falls.

The difference is measurable. Sub-300ms systems complete markedly more conversations than 500ms+ systems. This isn't a marginal improvement. It's a fundamentally different business.

The investment pays for itself. A mid-sized sales department can have DKK 15-20 million in latency-related revenue loss a year. The right AI infrastructure isn't an IT cost; it's revenue protection.

If your AI vendor can't give concrete answers to questions about P95 latency, inference location and streaming architecture, that's a signal. The technology may be impressive in a demo, but a demo isn't 1,000 simultaneous sales calls under Monday's peak load.

Build on infrastructure designed for what you actually need.

Poor response times in customer service can cost a considerable share of revenue.

Next steps

Want to hear how AI can be used in your sales calls? Book a review with Agent360.

For companies considering implementing AI support in sales calls from scratch, we recommend starting by understanding what a modern commercial operating system for sales actually requires technically, and then working backwards to the infrastructure specifications.

#latency #voice ai #technical #salesai #realtime

Should your conversations write themselves up afterwards?

Minutes, follow-up and a CRM note without anyone sitting down to do it. We look at what you use today.

How we built it · Sales & follow-up

Sales cockpit with AI

A call list for salespeople. The call is transcribed in Danish, AI suggests the outcome (e.g. “Call again”) and drafts the email based on what the customer said.

  1. The salesperson calls
  2. AI writes transcript and email
  3. The salesperson sends
See the call and the follow-up →