By Damian Mathews & The Last Mile Team
The other day Fish released a framework that reshaped my view on AI in the contact center. He argues AI since ChatGPT has moved through four distinct modes, each with a different orientation, a different compute profile, and a different ceiling on economic value.
Mode 1 is autoregressive. The model reaches backward into training data and predicts the next token. ChatGPT. Question and answer. Information retrieval through a better interface. It saves time. It does not, by itself, do new work.
Mode 2 is autoprogressive. The model reasons forward through a problem, building each step on the last. Reasoning models. Burns tokens by design. Solves problems retrieval cannot.
Mode 3 is autoexecutive. The model reaches outward into external systems. Agents in harnesses. Sends emails, writes files, processes refunds, books travel, coordinates other agents working in parallel. Performs consequential work in the world. The economic value here is categorically higher than anything before it.
Mode 4 is autotelic. The system pursues goals it sets for itself. Still mostly future tense.
The framework is useful because it makes a question precise. Which mode is your contact center actually deploying?
Most teams will say agentic. Most teams have not deployed Mode 3.
Walk into the average contact center AI deployment in 2026 and what you often find is Mode 1 in a new wrapper. The bot retrieves an answer from a knowledge base. It summarizes a ticket. It surfaces a relevant article for the agent. These are useful capabilities. They are also fundamentally the same capability the call center had ten years ago with a worse interface.
The value the industry keeps promising lives in Mode 3. An agent that opens a refund case, processes it, updates the CRM, notifies the customer, and closes the loop without a human writing a line of glue code. An agent that handles the long tail of contacts that used to require a human because the workflow was too unpredictable to automate. An agent that watches every conversation in real time and intervenes when something is going sideways.
This is real. It is also rare.
The gap between Mode 1 deployments and Mode 3 outcomes is the gap that explains why nearly half of enterprise leaders now describe AI adoption as a disappointment. They deployed search and expected transformation. They got search.
Closing that gap requires three things most vendors will not give you. A clear-eyed assessment of which mode your current deployments are actually in. A workflow redesign that puts humans where their judgment matters and removes them from places they are just data entry clerks. And an architecture that can move work between modes as the work itself changes shape.
That third one is hard. It requires someone who understands that the operating model has to evolve at the same speed as the technology. You need a partner that runs the build with you, inside your operation, with the experience already behind them.
We are about to launch something built for this exact gap (more on that extremely soon).
In the meantime, the diagnostic question is the useful one. Pull a list of every AI deployment in your operation right now. For each one, mark the mode, and be honest about it.
How many of them are Mode 1 wearing a Mode 3 sticker?
– Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
SoftBank is pouring up to €75 billion into French data centers, its largest AI infrastructure bet in Europe. The plan funds 5 gigawatts of capacity, starting with 3.1 GW across the Hauts-de-France region by 2031. At this scale, Europe’s compute map could shift toward France, though grid and permitting hurdles may slow the buildout.
Anthropic raised $65 billion and confidentially filed to go public, reaching a $965 billion valuation. That figure tops OpenAI for the first time, on a reported $47 billion revenue run rate. A near-trillion-dollar AI pure-play hitting public markets could reset how investors value the entire sector, assuming the listing lands.
Leading Edge
Proven moves you can copy today.
OpenAI’s frontier models and Codex are now generally available on AWS, landing inside Amazon Bedrock. GPT-5.5, GPT-5.4, and the Codex engineering agent now run in customers’ existing AWS environments, no migration required. For the many enterprises already standardized on AWS, this removes a real adoption barrier and is worth piloting.
Anthropic opened its powerful Claude Mythos security model to 150 more organizations across more than 15 countries. Early partners, including NATO, the EU’s ENISA, Okta, and Samsung, have already flagged over 10,000 serious flaws. Putting frontier models on defense could finally tilt the security math, though attackers get the same tools.
Off the Ledge
Hype and headaches we’re steering clear of.
Tech layoffs have already hit 142,000 this year, even as the companies cutting jobs stay profitable. Amazon, Microsoft, Alphabet, and Meta have committed roughly $700 billion in 2026 capex, much of it to compute. Calling it ‘funding AI’ does not change the math: firms are turning payroll into data centers.
GitHub Copilot moved its 4.7 million paying developers to token-based billing, and the backlash was immediate. Some heavy users report projected monthly costs leaping from about $50 into the thousands, while code completions stay free. After years of encouraging heavy use, the surprise bills are pushing some developers toward rivals like Cursor.
