By Damian Mathews & The Last Mile Team
Almost every CX leader I talk to has one. An AI pilot that worked. The demo landed, leadership was impressed, everyone agreed it should go live.
And then it just sat there. Three months later it’s still sitting there, and nobody can quite explain why. The usual explanations are that the model isn’t good enough yet, or the budget dried up, or the team got pulled onto something else. Occasionally those are true.
Far more often, the pilot stalled the moment someone said the most reasonable-sounding sentence in the room.
“Great. Let’s clean it up and ship it.”
That sentence is where most CX AI projects quietly go to die. It’s worth understanding why, because the reason isn’t obvious.
Fish wrote a piece recently called Stop Trying to Clean Up Your AI Prototypes that names the problem better than I’ve seen anywhere. His core point is that a prototype and a production system look like the same thing at two stages of completion. They’re actually different disciplines built for different purposes.
A prototype is built to learn. It’s optimized for speed and discovery, and it earns that speed by deferring everything production can’t. Edge cases. Error handling. Security. Integrations. What happens when the customer says something nobody planned for.
A production system is built for consequence. It has to hold up when a real customer is on the line, when the input is messy, when the thing breaks at 2am. AI makes both builds faster. They’re still different builds. So “clean it up and ship it” assumes the two are the same activity at different points on a progress bar.
They aren’t.
Trying to harden a prototype into production is like turning the architect’s scale model into the actual building. The model did its job. It got everyone to agree on what to build. You don’t then pour concrete into it.
Fish’s sharpest point is about direction. Struggling teams take the prototype and expand outward from it, bolting on security, wiring in integrations, patching edge cases one at a time. Every fix is constrained by choices the prototype made for entirely different reasons.
The better motion is to build inward. Start from the production environment, the one with your real systems, your compliance rules, your escalation paths. Then rebuild the validated idea inside it, using the prototype as a reference for what it should do rather than the foundation it’s built on.
Here’s the catch, and it’s the real reason so many pilots freeze. Building inward requires a production environment to build into.
If you don’t already have the integration patterns, the testing and simulation setup, the review and quality gates, and the governance in place, then the only thing you can build from is the prototype itself. Expanding outward isn’t a mindset failure at that point. It’s the only option available.
This is exactly what Kerry means by “Build the Last Mile First,” the fifth of the five plays in A1B: Customer Zero to AI-First. The other four plays get you to a validated prototype quickly. Show Don’t Tell, the Creativity of the Crowd, Possible-Practical-Profitable, and Deliver on the Vibe are all about discovering the right thing to build, fast.
The fifth play is the one everybody skips, and the one that decides whether anything ever ships. Prepare the production environment first, so that when the prototype proves the idea, you have somewhere real to build it.
For a CX leader, the practical version is a set of questions to ask before the next pilot, not after it. Where would this actually run in production? What does it need to connect to? Who reviews its output, and how does a bad answer get caught? When it hits something it can’t handle, where does it hand off?
If those answers don’t exist yet, the pilot will impress everyone and then stall, no matter how good it is.
The frozen pilot on your roadmap probably isn’t stuck because the AI fell short. It’s stuck because there was never a production environment prepared to receive it. So “clean it up and ship it” was the only path forward, and that path doesn’t lead anywhere.
What would it take for your best pilot to actually go live?
— Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
OpenAI taped out its first custom chip, a Broadcom-built inference processor it calls Jalapeño. The two firms took it from design to tape-out in nine months and aim to deploy by the end of 2026. Owning the silicon could ease OpenAI’s compute crunch, though a first-generation accelerator has to prove itself in production before anyone declares independence from Nvidia.
Google started rolling out Gemini 2.5 Deep Think, a parallel-reasoning mode for its Pro model. It pairs a two-million-token context window with multi-path reasoning that Google says leads current science and math benchmarks. Leaderboard wins rarely survive real workloads, but a model that can hold an entire codebase in context looks like genuine pressure on rivals.
Leading Edge
Proven moves you can copy today.
xAI’s Grok 4.3 arrived on Amazon Bedrock, making xAI the third major lab on the platform. It offers a one-million-token context window and adjustable reasoning, aimed at support, coding, and financial document tasks. Teams already on Bedrock get another frontier option with no new contract, though they will still want to test Grok’s enterprise guardrails first.
Assort Health raised $120 million to scale AI that answers patients’ phone calls. The Series C values it near $1.2 billion as voice agents now handle much of the inbound scheduling at some US hospitals. For any contact center the lesson is plain… where calls are repetitive and staffing is thin, voice automation finally appears ready to earn its keep.
Off the Ledge
Hype and headaches we’re steering clear of.
Tech layoffs are running near 1,115 jobs a day in 2026, and many employers are pointing at AI. Analysts flag a wave of ‘AI washing,’ where cuts driven by overhiring or soft revenue get repackaged as automation gains. Pinning it on the robots writes a clean press release, but boards may eventually ask where all that promised productivity actually went.
A US judge scrapped a trial after lawyers on both sides cited cases AI had invented. It caps a 2026 surge of sanctions, with courts fining attorneys thousands for fake but confidently cited precedents. Most of these filings failed at the easiest step, basic verification, and ‘the AI wrote it’ is shaping up to be a career-limiting excuse.
PS: Fable 5 is back!!!!
