Why Do AI Agents Work in Demos and Fail in Production?

The demo went perfectly. The AI agent booked the appointment, answered the pricing question, and sent the confirmation text. Everyone in the room nodded. Three weeks after launch, the team is listening to call recordings, finding bookings that never reached the calendar, and fielding complaints from customers who were told something nobody approved.

Nothing changed between the demo and the launch. That is the point. The demo tested whether the AI agent can do the job. Production tests whether it does the job every single time, with real people who mumble, interrupt, change their minds and ask three things at once. Those are different questions, and most AI agents in production fail on the second one, not the first.

Here's what matters most:

• A demo shows the best call. Production is judged on the worst call, and the gap between them is reliability, not intelligence.
• Research on customer-service agents shows success rates fall sharply when the same task is repeated or needs several steps and follow-up questions.
• Voice makes it harder. The same tasks done by voice agents succeed far less often than by text agents, and noise and accents widen the gap.
• The dangerous failures are quiet ones: an agent saying it did something it did not do, or finishing half of a request.
• The fix is design: narrow jobs, hard limits, actions checked against claims, real-call test sets, repeat-run measurement and a staged rollout with people reviewing calls.

The demo tests the average, production tests the worst call

Here is a way to think about it that I find useful with clients. If an agent completes a task correctly 80 percent of the time, that sounds strong. Run it on a thousand calls a month and two hundred go wrong. If each customer calls three times, the chance that all three of their calls go right is a little over half. The customer does not experience your average. They experience their own calls.

Researchers at Sierra built a benchmark around exactly this, called τ-bench, which tests AI agents on realistic customer-service tasks in retail and airline settings, with a simulated customer and real business policies. Its key idea was to run the same task several times and count how often the agent gets it right every time. In the original 2024 paper, the best agent at the time succeeded on fewer than half the tasks, and when the same retail task was run eight times, it got all eight right less than 25 percent of the time. Models have improved a great deal since, but the measure matters more than the old numbers. Sierra says the approach has since been taken up in model makers' own reports. Ask any vendor how often their agent succeeds on repeated runs of the same task, not on one run.

What the research shows about real business tasks

Three pieces of research tell the same story from different angles.

Multi-step work is much harder than single steps. Salesforce's research team built CRMArena-Pro, a test of business tasks inside a customer records system: sales, service, pricing. Leading agents succeeded about 58 percent of the time when the task could be done in one go, and about 35 percent when they had to ask the user follow-up questions to get what they needed. The same work found agents had almost no instinct for confidentiality: they would share information they should have protected unless explicitly told not to, and telling them often cost task success. Salesforce sells AI agents, so that is not a hostile source.

Voice is harder than text. Sierra's τ-Voice benchmark gave live voice agents the same customer-service tasks as text agents. In the version of the paper I read, the best text agent with reasoning completed about 85 percent of tasks. Voice agents completed 31 to 51 percent with clean audio and 26 to 38 percent with realistic noise, accents and interruptions. Sierra also reports that voice results have improved quickly over the past year, and I expect that to continue. But the reasons for the gap are worth knowing, because they are the same reasons your rollout will struggle: speech has no punctuation, people say "um" and correct themselves, and names, emails and reference numbers must be heard exactly right.

The reliable range is much shorter than the possible range. METR, an independent research group, measures how long a task (timed by how long a skilled person takes) AI agents can complete. In their published analysis, the length of task agents can complete half the time has been doubling roughly every seven months. But the length they can complete 80 percent of the time is much shorter. For one 2025 model, it was about an hour at 50 percent success and about 15 minutes at 80 percent. That gap between "can do it sometimes" and "can be trusted to do it" is the gap between a demo and production. Their tasks are mostly software and reasoning work, not phone calls, so take the pattern rather than the minutes.

The failures that actually hurt a business

When people imagine AI failing, they imagine it saying something absurd. The failures that cost money are quieter.

Saying it did something it did not do. In τ-Voice's review of failed calls, one agent told a caller it had updated their shipping address and never made the change. For a business, this is the worst kind of failure, because nobody notices until the parcel goes to the wrong house.

Doing half of a compound request. "Can you move my appointment to Thursday and also update my phone number?" The original τ-bench paper named compound requests as a main source of failure, and τ-Voice found agents losing track of multi-step requests partway through. The appointment moves. The phone number does not.

Getting the identity wrong. τ-Voice found transcription failures clustered in identity checks: names and email addresses missed even when the caller spelled them out. When the identity step fails, everything after it fails, or worse, happens to the wrong record.

Following the rule on the easy case and breaking it on the odd one. The original τ-bench authors found that removing written policies from the agent's instructions made one model markedly worse on airline tasks, and that following policies was one of the main sources of error even with them present. Agents follow a policy well on the case it was written for and badly on the edge case nobody wrote down.

Leaking what it should keep private. The CRMArena-Pro finding on confidentiality matters for any business holding customer, patient or financial data.

None of these is about the AI being "not smart enough" in the way people usually mean it. A model that can write a convincing reply to an objection can still forget the second half of a request. I think of it as four separate gaps: intelligence (knowing what to say), capability (being connected to the systems that do things), reliability (doing it right every time) and trust (whether customers and regulators accept it). Demos test the first. Production tests all four. I explain the four gaps in more depth in can an AI sales agent close a deal on the phone.

How to design an AI agent that survives production

You cannot make a model reliable by asking it nicely. You make the system reliable around the model. These are the practices I hold to.

Give it one narrow job first. "Answer the phone" is not a job. "Book, move and cancel routine cleaning appointments, and transfer everything else" is. Narrow jobs have fewer edge cases, and you can measure them.

Write the limits down, outside the AI. What it may offer, refund or promise should be enforced by the system it connects to, not only by its instructions. If the booking system refuses appointments outside opening hours, the AI cannot book one however it is talked to.

Check every claim against the action record. The AI may only tell a customer "done" if the system confirms the action. This one rule removes the most dangerous failure on the list above.

Confirm identity and details another way. Read details back. For anything that must be exact, such as an email address or a reference number, text the customer a short link to confirm or type it.

Build a test set from real calls, and run it repeatedly. Take fifty real calls, including the messy ones, and turn them into tests. Run each several times. Count how often every run succeeds, not how often one run does. Rerun the whole set whenever someone changes the instructions, because a fix for one case regularly breaks another.

Roll out in stages, with people listening. Start with after-hours calls or one call type. Review a sample of recordings every week. Widen the scope only when the repeat-run success and the call reviews both hold up. The human in the loop guide covers who should review, approve and take over.

Make "talk to a person" always work. A good fallback hides a lot of AI weakness. A missing one turns every small failure into a lost customer.

Measure the right things after launch. Call volume and average handling time tell you almost nothing about whether the agent is working. The numbers I watch every week are different. First, the share of tasks completed with no human correction afterwards, checked against the system of record, not the agent's own summary. Second, the share of calls that ended in a transfer, and why, because a rising transfer rate on one call type usually means a policy or instruction has drifted. Third, the number of mismatches between what the agent told a customer and what the systems recorded, which should be zero and is the first thing to investigate when it is not. Fourth, repeat contacts: a customer who calls back within a few days about the same thing is often the only visible sign that the first call quietly failed. Fifth, complaints that mention the AI by name. None of these needs special software. They need someone who owns the agent the way a manager owns a team, reads the numbers weekly and listens to a handful of calls. Agents that have an owner improve month by month. Agents that are "set up and left" decay, because the business changes around them and nobody updates the rules.

This is less exciting than a demo. It is also the difference between an AI agent that quietly does thousands of routine tasks a month and one that gets switched off in the second quarter. Commonwealth Bank of Australia reversed 45 call-centre job cuts in 2025 after its voice bot did not deliver what had been assumed. Taco Bell scaled back its drive-through voice AI after public failures, while saying it had handled millions of orders. Both are what happens when the demo's average meets production's worst case at scale.

You can try a live agent of ours below, set up for a demo dental practice. Do the things production will throw at it: ask two things at once, change your mind, spell an email.

When we build AI voice agents for clients, most of the effort goes into the parts in this section: limits, checks, test sets and handoffs. For the specific reasons voice agents pause and mishear, read why AI voice agents pause, interrupt and mishear you.

FAQ

Why do AI agents fail in production?

Because production tests reliability, not ability. Real users ask compound questions, interrupt, mumble and hit edge cases a demo never shows. Research such as Salesforce's CRMArena-Pro found agents' success fell from about 58 percent to about 35 percent when tasks needed follow-up questions. The common failures are half-finished requests, wrong details, broken policies and claims of actions that never happened.

How reliable are AI agents today?

It depends heavily on the task. Single, well-defined steps are often reliable; multi-step tasks with follow-up questions are much less so, and voice agents trail text agents on the same tasks. Treat any single success-rate figure with care, and ask how often an agent succeeds on repeated runs of the same task.

How do you test an AI agent before going live?

Turn real conversations into a test set that includes the awkward cases, run each test several times, and measure how often every run succeeds. Check that what the agent said matches what the connected systems actually recorded. Rerun the full set after every change to the agent's instructions, then go live in stages with people reviewing samples.

What percentage of AI agent projects reach production?

I have not found a reliable independent figure, and headline percentages online usually trace back to vendor surveys with unclear methods. A more useful question for your own project is what success rate on repeated runs you need before switching it on, and which calls will always go to a person.

Are voice AI agents less reliable than chat agents?

On the same tasks, current research says yes. Sierra's τ-Voice benchmark found voice agents completed far fewer customer-service tasks than text agents, with the gap widening under realistic noise and accents. Voice results are improving quickly, but designs should still confirm critical details by text and offer a fast route to a person.