Every talent show has the contestant with the gorgeous voice and absolutely nothing to say. Pitch perfect, breath control for days, and by the second chorus you have quietly picked up your phone. Technically flawless. Emotionally, a car park.

That is roughly where the AI voice agent industry found itself this month. On July 23, DevRev — the Palo Alto company behind the agentic AI product Computer — announced Voice AI across its Customer Agent, extending the same organisational memory that already powers its chat and email agents into live phone calls. And tucked inside the announcement are three numbers that every business owner thinking about automating their phones should read twice.

Three Numbers That Explain the Disappointment

First: a Gartner survey published in February 2026 found 91% of customer service and support leaders are under executive pressure to implement AI. Ninety-one percent. That is not a technology trend, that is a stampede, and stampedes are famously bad at reading the ground in front of them.

Second: Gartner reported in March 2026 that only 20% of organisations have actually reduced agent headcount because of AI. Everyone is buying. One in five is seeing the thing they bought it for.

Third, and this is the one that should make you sit up: Forrester predicts roughly one-third of brands rolling out AI in self-service will fail in 2026 — largely because the data foundations underneath the AI are not ready. Not the models. The plumbing.

DevRev's own summary of the gap is the best sentence written about this industry all year: enterprises are buying conversation and discovering they need resolution.

Conversation Is Cheap Now. Resolution Isn't.

Manoj Agarwal, DevRev's co-founder and president, put the engineering problem plainly. "The interesting engineering problem isn't making an agent sound more natural," he said. "It's giving an agent enough context to make the right decision while the customer is still on the line." In an enterprise, he noted, the answer rarely lives in one place — it lives across tickets, documentation, customer records, orders, engineering systems and business applications. If the agent cannot reason across that, it eventually hands the call to someone who can.

Which is a polite way of describing the thing we have all experienced: a very pleasant robot taking ninety seconds to transfer you to a human, who then asks for your account number again.

DevRev's pitch is that its Voice AI is grounded in a shared organisational memory synced from the customer's existing systems, permissions intact, so the agent can work out what happened, explain it, give the current status, kick off the corrective workflow and document what it did — inside one call. When it does need a person, it hands over with the full conversation context attached, so the customer isn't starting from scratch.

The Ninety-Day Test

There's a second detail worth stealing. DevRev says Computer studies an organisation's business data and ships a working first version of Voice AI within hours, then improves it with every conversation, recommending weekly refinements based on real performance — so that the agent on day ninety is measurably better than the one launched on day zero.

Set aside the marketing gloss and that is a genuinely useful benchmark to hold any vendor to. Not "how good is the demo," but "how much better is this in three months, and how would we know?" A voice agent that performs identically on day ninety and day zero is not a product. It's a very expensive answering machine with a nicer accent.

DevRev also appointed Jennifer Heape to lead Voice AI product; she previously co-founded and led product strategy at Vixen Labs. Her framing of the problem is the customer's, not the engineer's: a call can sound polished and still be a poor experience if the agent cannot reach the information needed to solve the problem. Customers, she pointed out, do not care which system holds the answer. They care whether they have to repeat themselves, wait for a transfer, or ring back.

What a Small Business Should Actually Take From This

DevRev is selling to enterprises with SIP telephony and eight-figure support operations. You are not that. But the failure mode Forrester is warning about scales down perfectly, and it is arguably worse for you, because a big company losing a call loses a ticket and a small company losing a call loses the whole customer.

So before you evaluate a single voice on a single demo, do the boring bit. Write down the five things people actually ring you about. Then, for each one, ask what information the agent would need to genuinely finish the job — your calendar, your price list, your job status, your supplier lead times — and whether that information exists somewhere the agent can reach it, or only in your head and a notebook in the van.

Anything on that list that is only in your head is not an AI problem yet. It's a write-it-down problem. Fix those first, pick the one call type where the agent can go end to end without a transfer, and let it prove itself there. The most convincing voice in the world can't look up a booking that was never written down.