Summarize this page with popular models
Summarize this page with popular models
After the dumpster fire, here's how to get it right the next time around.
4 minute read
AI deployments start with promises and optimism. Many of them end with a postmortem. As I wrote in “Welcome to the Seven-figure Dumpster Fire,” letting AI delusion outstrip AI reality has produced more than one burning pile of nonsense (and dollar bills). Solutions over-promise and under-deliver, companies sign up to the vision, and millions later, the project is written off or turns into something completely different that doesn't work at all.
Now for the part nobody plans for: what happens after the conflagration. If that's where you are, picture it. The flames in the dumpster have gone out, and you're looking at a Pollock-like mess of melted plastic and crushed C-level dreams.
Is there something salvageable in the wreckage? Perhaps. It depends on what you do next.
Many implementations fail because they try to do it all, and AI is not the end-all solution to everything that ails a business.
Instead of boiling the ocean, look at what your support staff solves over and over: callers checking when their appointment is, or asking for help creating an account. The smaller the use case, the better. Pick one where AI has a real shot at succeeding, decide up front what “working” looks like, and launch it. If it works, scale from there.
Many projects over-index on the model. “This model? Let's just deploy it, and we're good!” The real breaking points tend to show up where the AI meets the rest of the business. So take a minute to look at the end-to-end process:
Say you have a voice agent helping people book dentist appointments. You want it to do more than talk and drop things on a calendar. You need it to go into the systems of record and check or change things. When was this person's last cleaning? When should the next reminder go out? If all you're thinking about is the model layer, you're not adding a whole lot of value.
Voice AI is especially good at revealing whether you've done your due diligence on edge cases, and some hilariously public misfires show what happens when you haven't.
When Taco Bell rolled out voice ordering at its drive-throughs, people flooded social media with recordings of their encounters with the human-free “solution.” One man crashed the system by asking it for 18,000 cups of water. McDonald's ended its voice AI drive-through test in 2024 after the system misread orders, adding bacon to someone's ice cream and hundreds of dollars' worth of chicken nuggets to another person's order.
Hilarity aside, these rollouts created more work for staff, not less. They optimized for theoretical efficiency over reality, and they didn't account for the fact that people love a drive-through prank, whether a human or an AI is behind the speaker.
Taco Bell, at least, took a step back. The company has said it is coaching restaurant teams on when to rely on voice AI and when to monitor it and step in. They've joined a host of big names that are by no means scrapping AI, but are reevaluating a hot-out-of-the-gate approach.
Walmart's approach is instructive, too. Suresh Kumar, Walmart's CTO and Chief Development Officer, said the company is simplifying its agents so employees don't have to remember which one does what: “If I have an agent that helps you with your payroll and I have a different agent that helps you with identifying merchandising trends, you shouldn't have to remember that and switch between those two.”
An MIT report from last year found that AI bought from vendors specializing in a particular problem succeeded about 67% of the time. Internal builds succeeded about a third as often. Lead author Aditya Challapally noted that companies building their own tools were often hesitant to disclose their failure rates.
Custom builds carry their own risk. When a project tries to anticipate every possible problem and hand-build software for all of it, you tend to get a long project with a much higher bill.
The better bet is a solution that arrives with a clear playbook for your problem: finished agent configurations that can be deployed in days, company knowledge that connects without new infrastructure, conversational paths that come pre-configured. Then you're making small adjustments, not large-scale overhauls.
Throwing money at the problem, if that wasn't clear enough, is not the way forward. The organizations bouncing back have lowered their celestial ambitions and gotten specific about what they're building. Before you relaunch, ask:
The public record shows humans are still absolutely needed, so build them into the plan from the start.
Then be honest about whether the solution meets the needs of both sides. If the answer is no, hold off and work the problem until it holds up.
Co-Founder & CTO, 8x8
Bryan Martin is Co-Founder and Chief Technology Officer at 8x8, where he leads the company's technology vision across AI, cloud communications, and platform architecture. One of the original architects of 8x8's cloud infrastructure, Bryan has shaped the company's technical direction from its early days as a VoIP pioneer to its current position as a Communications Intelligence Platform. He holds 137 U.S. patents spanning semiconductors, video processing, computer architecture, and communications — and brings that same builder's instincts to how 8x8 approaches AI: specific use cases, clean handoffs, and outcomes that hold up under scrutiny. Bryan received Bachelor's and Master's degrees in Electrical Engineering from Stanford University.
Stay current on what matters in CX and IT. Subscribe to our LinkedIn newsletter for regular analysis on the decisions shaping customer experience, technology, and AI. Clarity you can act on.
