Quanta Bits - May 9, 2026

Hi everyone,

Spock is back.

Like a garden left untended, my OpenClaw setup had become a bit chaotic after a run of upgrades and versioning issues. The ironic answer, recommended by a few people in the AI builder world, was to use Codex to fix it. After a week of wrestling with it, Spock is healthy again and I am back to enjoying the clean garden.

That small experience connects to a bigger pattern. The more AI enters real work, the more obvious the maintenance work becomes. Someone has to keep the agents, tools, workflows, data, costs, and exceptions from turning into a messy garden.

That is where this week's essay goes: the new operating work AI creates, and the roles companies may need to make it useful.

AI Ops is becoming the missing layer. The repeated pattern this week was not "people need more AI tools." It was that someone has to operate the AI work: agent inventory, workflow ownership, context, data hygiene, evals, access, cost, exceptions, and adoption.

The AI layoff story is getting pushback. HBR, a16z, and Big Technology all pushed against the clean job-apocalypse story from different angles. Layoffs branded as AI efficiency can create mistrust, weak adoption, low-quality AI work, and hollowed-out talent pipelines.

Frontier labs are admitting deployment is the hard part. OpenAI and Anthropic moving into PE-backed enterprise services is a big tell. Model access is not enough when enterprise value gets stuck in workflow redesign, change management, and technical-plus-process work.

AI adoption is exposing old operations debt. In operating meetings this week, the same issues kept showing up: funnel visibility, lead fields, territory ownership, roadmap coordination, data quality, readiness checklists, and intake processes. Pipeline scoring on dirty CRM data. AI makes these more visible.

Agents are moving closer to customers and money. Voice agents, service agents, agent payments, agentic CRM examples, and agent-generated media all pointed in the same direction. Once agents touch customers, transactions, and brand experiences, ownership and review stop being optional.

The first cloud bills felt like progress. Teams could provision infrastructure quickly. Product groups could test ideas without a long hardware cycle. Then the bills kept growing.

I remember a client where cloud spend started doubling fast enough that the CFO had to pull the emergency brake. The questions were simple: what is this workload doing, can we run it more cheaply, and who owns the bill?

That is how CloudOps and FinOps became real disciplines. Cloud had become too useful to leave unmanaged.

I think AI is entering a similar phase. Costs matter, but ownership matters more: where is AI being used, what is it costing, and is the work actually getting better?

The AI bill is messier than the cloud bill. Cloud spend was attached to servers, storage, databases, environments, and projects. The AI bill includes tokens, software, employee time, review and rework, integrations, controls, bad records, broken handoffs, and workflows that never produce value.

Deloitte's Nicholas Merizzi and colleagues recently framed this well. AI costs shift with model choice, prompt length, workload design, infrastructure, networking, power, and cooling. The bill comes from many small design choices, not one procurement decision.

Cost matters, but it is not the whole story. Once AI touches real workflows, someone has to know which systems it touches, which data it trusts, who reviews the output, and who fixes the process when the answer is wrong.

Companies need budget alerts, usage dashboards, model catalogs, spend thresholds, and cost allocation. The FinOps Foundation now has a dedicated FinOps for AI category because AI spend is granular, unpredictable, and spread across SaaS products, model vendors, and hyperscalers.

Budget controls help, but the AI bill needs an operating owner. That should start as a shared motion led by the COO, CIO, or transformation leader, with Finance measuring spend, IT and Business Applications owning platforms and access, and business leaders owning outcomes.

That function has to ask questions between finance, technology, and the business:

Gartner made a related point this week: better ROI comes from the skills, roles, and operating models that let humans guide and scale autonomous systems.

That matches what I am seeing in practice. The companies getting value from AI are creating connective tissue around the tools: intake, approved use cases, access, workflow design, review paths, quality checks, measurement, and adoption. Cost is only the first visible symptom.

This is familiar if you have watched Sales Ops, Marketing Ops, RevOps, or Finance Ops mature. Sales Ops is not one job. One person may own territories, another forecasting, another lead routing, CRM hygiene, sales compensation, reporting, or process design. Together, they make the sales system work.

AI Ops may evolve the same way. I would start by naming the work:

These do not have to be five separate people. In a smaller company, one strong operator may cover several responsibilities for a while. The important part is protected capacity: time away from daily fires to define work that is still messy.

That is the positive version of this story. AI creates new work for people who already understand process, systems, exceptions, adoption, and measurement.

The market signals point the same way. Microsoft's Agent 365 launch post talks about inventory, ownership, least privilege, compliance, lifecycle management, monitoring, and operations for agents. Cisco's March research note says 85% of surveyed customers are experimenting with agentic AI, but only 5% have broad production deployment.

BCG's work on scaling AI agents makes the same point: companies need to redesign workflows, decide where agents fit, and build cross-functional teams to sustain them. Aaron Levie, CEO of Box, put it more bluntly: implementing agents in enterprises will create more work than people imagine.

Most companies do not need an AI Ops function tomorrow. But if you can afford it, create a small team or protect one person's time to map what is already happening.

That is how many companies started with FinOps and CloudOps: someone inventoried workloads, mapped owners, studied dependencies and usage patterns, and explained why spend moved.

AI needs the same first motion. Inventory the agents, copilots, workflows, and AI features. Map the systems they touch, data they read or write, cost, usage, dependencies, and output review.

Then assign the first owners: agent inventory, cost and usage patterns, workflow quality and exceptions, and access, logs, and system dependencies.

The first team does not have to solve everything. It has to create the operating map.

The first phase of AI adoption was experimentation. The second is operations. The opportunity for operators is to become the people who make AI useful, measurable, and worth keeping.

IBM puts numbers on the C-suite rewire. IBM's 2026 CEO Study says top-quartile AI-maturity CEOs report 17% higher revenue growth, and companies that redesigned technology, finance, HR, operations, and cross-functional collaboration are 4x more likely to deliver on business objectives. The operating point: CEOs say 86% of employees have AI skills, but only 25% use AI regularly at work.

AWS gives agents a way to spend money. Amazon Bedrock AgentCore payments lets agents pay for services using Coinbase and Stripe rails, with user authorization and per-session spending limits. That turns cost control into an operating question: who decides what an agent can buy, spend, and get reviewed?

OpenAI and Anthropic move into deployment services. TechCrunch reported that both labs announced PE-backed AI ventures built around forward-deployed engineering capacity. The tell: the hard part is fitting AI into real workflows.

CrowdStrike names the patch sound barrier. In a May 8 DealBook interview, CrowdStrike CEO George Kurtz warned that frontier models could surface "hundreds of thousands, if not millions, of patches." The issue is operating throughput: banks, telecoms, and critical infrastructure cannot patch everything at once without breaking availability promises.

Anthropic packages finance agents as a vertical stack. Anthropic launched 10 financial-services agent templates, plus Office add-ins and data connectors for pitch building, Know Your Customer screening, reconciliation, and close. This is AI Ops by industry: templates, connectors, benchmarks, and clearer ownership inside finance operations.

NetSuite makes ERP more agent-ready. Oracle NetSuite launched SuiteCloud Agent Skills, six knowledge packages for AI coding tools building NetSuite customizations. The signal for CIOs: enterprise systems are starting to ship agent-readable infrastructure, not just AI features.

Floating data centers, literally. Panthalassa raised $140 million to build wave-powered floating data centers for AI workloads. It may stay niche, but it is a wonderful little reminder that when compute hits power and cooling constraints, the market starts trying stranger and maybe cleaner answers.

Dir. Nia DaCosta | 109 min | ★★★★☆

I mentioned back in February that the promos for this fourth film in the 28 Days Later zombie series had pulled me into catching up on the whole run. That investment finally paid off: this is the best one so far. The first half took a while to click and felt like a familiar post-apocalyptic B-movie, but once Ralph Fiennes' Dr. Kelson becomes the emotional center, the film gets stranger and more humane. The best idea is Samson, an infected alpha treated with care, sedation, and patience. Suddenly the rage virus becomes less of a zombie premise and more a story about psychosis, loss of self, and maybe recovery. Still too graphic for my taste, but the second half made the series feel like it has something real to say about what kind of people survive.

Agent maintenance is now an operational function, even for one-person shops. Codex fixed my broken OpenClaw setup after I had spent a week trying to get OpenClaw to fix itself. Same week, it submitted an invoice through a third-party app, filling forms and uploading the file. This newsletter is being assembled with Codex/GPT-5.5 instead of Opus 4.7 because the workflow is more predictable: less shaping, fewer detours, more "yes, that is what I meant."

The point is not which model wins this month. It is that even at a one-person scale, I now spend real time owning agent setup, agent breakage, agent cost, and agent switching. Multiply that by an enterprise and you get the essay above.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖