Your AI agent is talking to people and booking nothing — the diagnosis

If your AI agent is talking to people and booking nothing, the model is rarely the fault. Check seven causes in order: fake conversations, no ask, an over-tight qualifying gate, silent booking-tool failures, early hang-ups, voice mechanics and an empty calendar. In τ-bench, top tool-calling agents completed under 50% of tasks.

Sources last checked: 23 September 2026. Every external figure on this page links to the publisher that produced it, and was re-read at that source before publication.

• First cut: recompute the booking rate on conversations that lasted past 20 seconds. Voicemail, IVR menus and instant hang-ups logged as “conversations” make a working agent look dead.
• Cheapest test: count the transcripts in which the agent said a time was booked, then count the calendar events created in the same window. Any gap above 2% is a bug, not a sales problem.
• Most common cause once the denominator is clean: the prompt has no stop condition, so the agent answers questions helpfully and never proposes a specific slot.
• External evidence: in a Marketing Science field experiment (Luo et al., 6,200+ customers), disclosing a chatbot before an outbound sales call cut purchase rates by more than 79.7%.

Why is my AI agent not booking meetings? The seven causes, in order

The table is the talk-to-book ladder: symptom, likely cause, the test that separates it from the others, and the threshold that confirms it. The order is our judgement as the operator behind 50,769+ AI-booked appointments since 2017. It is not a published frequency count, and the thresholds are deliberately blunt working cuts, not research. Run the tests top to bottom.

The general version of this differential, covering speed, decision-maker reach and follow-up depth, is on our page on why an appointment set rate is low. This page covers only failures that come from the agent itself.

Is my AI agent actually talking to anyone? Check what counts as a conversation

An AI agent’s booking rate is only as honest as its definition of a conversation. Many dialler and agent platforms log a “conversation” whenever audio is exchanged. That includes voicemail greetings, IVR menus, gatekeepers and wrong numbers.

The test takes one export. Bucket every logged conversation by duration: under 20 seconds, 20-60 seconds, over 60 seconds. If more than half fall under 20 seconds, you do not have a booking problem yet. You have a classification problem, and possibly a contact-rate problem. Recompute the booking rate on the over-20-second bucket only before you change the prompt.

Why does my AI agent talk well and never ask for a time?

An agent tuned to be helpful will stay helpful indefinitely unless the prompt tells it when the conversation has done its job. The transcripts show accurate answers, polite rapport, “is there anything else I can help with?”, and no proposed slot. Proposing a Tuesday at 2pm is a goal, and the goal has to be written into the instructions as a stop condition. For example: once need, timing and authority are confirmed, offer two specific slots before answering anything further.

This mirrors the human-rep slot test on the set-rate page linked above, run on the agent’s transcripts. Tag 50 transcripts from conversations that passed qualification with one binary field: was a specific day and time proposed? Below 70%, the prompt is the cause. The second pattern to check is the qualifying gate. If more than half of your over-20-second conversations end “not qualified”, re-score 20 of them by hand against the rules your best rep actually uses. Five or more wrongly rejected means the agent applies the checklist more literally than any human would.

My AI agent says “you’re booked” but nothing is in the calendar

A silent calendar write failure is the cause most likely to make an AI agent look like it books nothing when it is closing. The agent says the words, the prospect hangs up happy, and the API call behind the words failed. An expired OAuth token, a wrong calendar ID or a timeout is enough, and the transcript still reads as a success.

The external evidence says to expect this. In τ-bench (Yao et al., 2024), a benchmark of agents using tools in simulated customer-service domains, even state-of-the-art function-calling agents such as GPT-4o succeeded on under 50% of tasks. Their pass^8 score, the chance of succeeding on all eight repeats of the same task, was under 25% in the retail domain. It is not a booking system, but it measures a closely related thing: completing a multi-step tool sequence reliably.

The arithmetic compounds. A booking usually needs three tool calls: read availability, create the event and write the CRM record. At 95% reliability per call, 0.95 × 0.95 × 0.95 = 85.7% of verbal yeses become real meetings. At 90% per call, it is 72.9%. The test is a reconciliation query: count the transcripts containing a confirmation, then count the events created in the same window. A gap above 2% is a defect to fix in the integration, not in the prompt. Cause 7 hides here too. If the availability call returns nothing because the agent is reading the wrong calendar or the wrong time zone, the agent will truthfully report that there are no slots. When the meeting has to land in a named rep’s calendar, the routing rules become part of the test, as covered in our guide to rolling out AI appointment setting across a corporate sales team.

Why do people hang up on my AI agent in the first 15 seconds?

How an AI agent opens a call decides whether a booking conversation happens at all. The strongest field evidence comes from Luo, Tong, Fang and Qu in Marketing Science (2019). They randomised more than 6,200 customers of an Asian financial services firm to outbound sales calls from chatbots or humans. Undisclosed, the chatbots sold as well as proficient workers. Disclosed before the conversation, they cut purchase rates by more than 79.7%, and customers were curt and ended calls sooner.

The lesson is not to hide the agent. Disclosure is required or expected in many jurisdictions and on most platforms, and this page is not legal advice. The lesson is that the seconds after the disclosure must carry the reason for the call. Plot when human-answered calls end. If more than 35% end inside 15 seconds, rewrite the opener before touching anything else.

Pauses, talk-over and misheard times on voice agents

Voice agents fail in two ways that text agents do not: timing and hearing. On timing, Stivers et al. (PNAS, 2009) studied conversation across 10 languages. Every language showed a general avoidance of overlapping talk and a minimisation of silence between turns, and the average gaps varied within just 250 ms of the cross-language mean. An agent that pauses for well over a second before every reply sits far outside that band. Callers fill the silence and the agent talks over them. Measure p90 response latency, not the average.

On hearing, Koenecke et al. (PNAS, 2020) tested five commercial speech recognition systems from Amazon, Apple, Google, IBM and Microsoft. They found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. A misheard “Tuesday at two” books the wrong meeting or none. Audit 50 calls where a time was spoken against the time the agent extracted. Our page on what AI voice agents can and cannot do covers the wider capability limits.

A worked diagnosis: 1,000 conversations, 9 meetings

Illustrative figures; substitute your own. An agent logs 1,000 conversations in a month and 9 meetings appear in the calendar, a reported booking rate of 0.9%.

The re-based booking rate is 9 ÷ 430 = 2.1%, not 0.9%. The confirmed losses are 165 qualified people never offered a time and 22 who said yes and got no meeting. A 29% write-through rate is not model flakiness. Three steps each succeeding about two-thirds of the time (0.66³ ≈ 0.29) points to a broken integration. Fix both: offer a slot in 80% of qualified conversations (208 offers), keep the 32.6% acceptance (about 68 yeses) and reach 98% write-through. That gives about 67 meetings. Treat 67 as a ceiling, not a forecast. People the agent did not previously ask may accept less often, so fix one stage per month.

What running this diagnosis costs in hours and skill

The talk-to-book ladder is cheap if the agent is instrumented and close to impossible if it is not. You need five fields per conversation: duration, disposition, a slot-offered flag, tool-call results with error codes, and the calendar event ID. With those, the first pass takes about a day: two hours to tag 50 transcripts, an hour for the reconciliation query, an afternoon for the latency and hang-up curves. Without tool-call logging, cause 4 cannot be tested at all, and adding the logging is an engineering ticket, not an ops task.

The recurring cost is the part teams underestimate. Prompts drift, tokens expire and calendars get reorganised, so the ladder needs re-running monthly and after every prompt or integration change. It needs someone who can read an API error log and judge a sales transcript. In most corporates sales ops owns the agent’s outcomes and IT owns its integrations, so the diagnosis has no natural owner. For how the agent, the integrations and the reporting fit together, see our overview of AI sales agents and their data requirements.

Frequently asked questions

Should my AI agent tell people it is an AI?

Assume yes. Disclosure is required or expected in many jurisdictions, and this is not legal advice. The booking cost of disclosure is real, though. In a field experiment with more than 6,200 customers, Luo et al. (Marketing Science, 2019) found that disclosure before an outbound sales call cut purchase rates by more than 79.7%. The fix is an opener that gives the reason for the call immediately after the disclosure.

How often do AI agents fail at tool calls like booking?

More often than transcripts suggest. In τ-bench, state-of-the-art function-calling agents succeeded on under 50% of simulated customer-service tasks, and pass^8 was under 25% in retail. A booking that chains three calls at 95% reliability each completes 85.7% of the time. Reconcile confirmed bookings against created events every week.

Is it the model or the prompt?

Usually the prompt or the integration. Swap the model only after the ladder has ruled out causes 1 to 4. A new model will not fix a missing stop condition or an expired calendar token.

Why does my voice agent keep getting interrupted?

Check latency first. Across 10 languages, Stivers et al. (PNAS, 2009) found that people avoid overlapping talk and minimise silence between turns, with average gaps varying within 250 ms of the cross-language mean. A long pause before each reply invites the caller to speak again, and the agent then talks over them. Measure p90 latency and fix the slowest step in the chain.

How many conversations do I need before judging a fix?

At a 5% booking rate, 100 conversations gives a 95% margin of about ±4.3 percentage points, which is too wide to see most real changes. At 400 conversations per arm it narrows to about ±2.1 points. Change one stage at a time.