Chapter 2

Production is where the real story starts

Nearly three in four enterprises have been forced to shut down or roll back a deployed AI communications agent. That number holds across every region and every industry in the study, and it doesn't decline with experience nor investment. In fact, among organizations that describe their guardrails as fully mature, the rollback rate is even higher.

This is happening in production, on customer-facing channels, in real interactions. When something goes wrong, customers are watching. They notice. And when that happens, a technical issue becomes a real trust event.

Read the chapter
Image for Production is where the real story starts

Shipping didn’t solve the problem

In AI customer communications, most companies have already shipped. By the end of the year, nine in ten organizations will have a live AI agent in production across one or more of their channels.

But 74% of those that have already shipped have been forced to roll back or shut those agents down. Among organizations that describe their guardrails as fully mature, the rollback rate is 81%.

All along, the market has been drawing the wrong finish line. We thought getting to production was the hard part. The data shows it isn’t. Instead of solving the problem, organizations are deferring it to a moment when stakes are higher, and customer trust is on the line.

74%

Average rollback rate among organizations that reached production. (Sinch, 2026)

81%

Rollback rate among organizations with fully mature guardrails. (Sinch, 2026)

The AI Paradox - Daniel Morris 2

Why rollback rate increases with mature organizations

Organizations with mature governance instrumentation can see failures that less mature organizations miss entirely. Those reporting no rollbacks aren’t necessarily running cleaner AI programs. In many cases, they simply lack the monitoring to know when something goes wrong.

The most likely explanation, supported by the broader data, is that the programs failing at 81% – those that reported fully mature guardrails – aren’t failing because they’re poorly run. They’re failing because their improved monitoring allows them to detect issues that others can’t spot. The organizations reporting no governance failures are not the benchmark. Instead, they may be the ones with the least visibility into what’s happening.

What’s triggering the rollbacks

Personal identifiable information (PII) or customer data exposure is the leading cause among those that reported a governance failure rollback, cited by 31% of organizations. Hallucination or brand risk is second at 22%. Lack of auditability – the inability to diagnose what went wrong – is third at 16%. 

These aren’t abstract risk categories. PII exposure means a customer’s personal data surfaced in an interaction it shouldn’t have. Hallucination means an AI agent said something confidently wrong to a real customer, on a live channel, under your brand’s name. And when 16% of rollbacks can’t be fully diagnosed because there’s no audit trail, the organization is left with a failure it can’t learn from and no way to prove it’s been fixed.

Sinch data (2026) reveals PII or data leakage (31%) and hallucinations (22%) are the main reasons for rolling back an AI agent.

Confidence is not protection

90% of enterprise decision-makers describe themselves as confident in their AI agent readiness. But the moment you stack those confidence scores against operational reality, a gap opens. 

Of those that rate their confidence as „somewhat“ or „very confident“, 75% have experienced at least one governance rollback. Confidence doesn’t correlate with fewer failures. In fact, both deployment and rollback rates are broadly consistent regardless of how ready organizations feel.

Confidence is a leading indicator of ambition, not a guarantee of governance readiness. The more useful question for any leadership team isn’t „are we confident?“ It’s „what would we see if something went wrong right now, and how fast would we see it?“

Sinch research (2026) shows 90% of business decision-makers describe themselves as confident in their AI readiness.

Image for What to do with this

What to do with this

74% of organizations have had to shut down a live agent. 

That rollback rate is a reason to be precise about what you’re building on and what you’ll be able to see when something goes wrong. And the uncomfortable implication of the 81% figure isn’t that mature programs are failing more – it’s that they’re able to see what less governed organizations are missing. 

One question worth reflecting on: If your AI communications agent failed right now, would you know before your customers did? 

The monitoring gap is where organizations are most exposed, and where a technical issue suddenly becomes a customer trust issue.

Pilot purgatory is over Chapter 1 Pilot purgatory is over The real cost of governance failures Chapter 3 The real cost of governance failures