Your Company Needs A Harness, Not An Upgraded Chatbot

152% — the quarter-over-quarter growth of agent tokens through Salesforce's Agentforce in Q1, 28.6 trillion total, on $11.1 billion of revenue. That is what production AI looks like on an income statement when the buyer's harness is built. The same week, the Financial Times put hyperscaler AI ROI at minus nine percent for Microsoft, minus thirty-five at Oracle. Same models. Opposite outcomes. The variable was always going to be the harness, and the next chatbot upgrade does not change which side you must build it on. (Hint: it's yours.)

Tie every AI spend line to a KPI today. Cost per token is the invoice; revenue per head is the business. Dan Martell said it on X this week, and he is right. If your engineer spending two thousand dollars a month on Claude Code is shipping what three engineers shipped last year, you have the best deal in the history of corporate America — and you do not know it because the controller is reading the wrong column. Treat AI spend with the same rigor you treat payroll, this quarter. The companies cancelling their Claude subscriptions are responding to a measurement failure, not a cost problem.

Assign one named human to own AI implementation. Today. By end of day. Matt Stockton said it out loud this week: every team is busy with its actual job, nobody is tasked with how AI changes the shape of that job, and the capability overhang is years deep. What your stack can do today is far ahead of what your team is doing with it. The fix is not training. The fix is a named owner with the authority to redesign workflow and a budget to do it. Pick the person, write the memo, send it before close of business.

Take Opus 4.8's new Thinking effort selector seriously, because nobody else will. Anthropic shipped a five-step cost-performance dial (Low / Medium / High / Extra / Max) at the task level today. That is a new buyer-side control surface, and almost nobody is going to set it deliberately for the next ninety days. Run an audit of your five most expensive Claude workloads this week, set the dial to Medium on three of them, and watch the math. The point isn't the savings. The point is to learn the new control before your competitors do.

Keeping up with AI is hard. We know — we do it daily. Today's column is the audit one column at a time. The eight questions in the long-form are what we run at Outsider Labs. If you can answer all eight cleanly, you don't need us. If you can't, we need to talk.

Outsider Labs Advisory →

Friday, May 29

1. Salesforce just settled the bull-vs-bear AI debate by accident. (Feeds SEVEN.) $11.1B Q1, with Agentforce processing 28.6 trillion agent tokens (up 152% QOQ) and 3.8 billion agentic work units (up 111%). The same week the FT pegged hyperscaler AI ROI at minus 9 to minus 35 percent under best-case assumptions. Same models. Opposite outcomes. The variable is who built the harness around the chatbot. (Full analysis below.)

2. Opus 4.8 shipped its answer to the ROI question. The answer is still on Anthropic's side of the dashboard. Anthropic launched Opus 4.8 with sharper judgment, longer autonomous work, the same price, and a new Thinking effort selector that lets you tune cost-performance at the task level. Aligned News framed it as the lab's direct response to enterprise ROI complaints. They are not wrong about the features. They are wrong about whose side those features actually live on. (Hint: it isn't yours.) (Full analysis below.)

3. Anthropic took funding at a $965 billion valuation as Claude demand surges. The lab side of the harness just priced in at almost a trillion dollars on the strength of selling the half that pays them. The buyer side is still mostly free — and that is the single largest arbitrage on the board for anyone willing to do the work the labs are not selling.

4. CISA released guidance on the careful adoption of agentic AI services. Federal cybersecurity is now publishing the same harness-component checklist Tomasz Tunguz wrote out as a private essay last week — runtime, identity, observability, kill switch, the whole thing. The buyer-side governance question is officially public policy. If your board doesn't have answers, the regulators are about to.

5. AI-driven traffic to retail is up 393%, and most businesses have no idea what to do with it. Reported this afternoon by The AI Report. AI is now sending customers to your front door at four times the rate of last year, and the harness to convert them — pricing, fulfilment, recommendation, identity — was not built for visitors who arrived via a chatbot. The gap between traffic and conversion is the harness gap, rendered in dollars.

Same Week, Opposite Outcomes.

Two numbers landed in the same news cycle this week and read like they came from different planets.

Salesforce, Q1: $11.1 billion in revenue, with Agentforce processing 28.6 trillion agent tokens, up 152% quarter over quarter, and 3.8 billion agentic work units. That is production AI on an income statement, billed and collected at industrial scale.

The Financial Times, same week, picked up on X and run to a quarter of a million views by an account called Yoshik: AI ROI at the hyperscalers, under best-case assumptions, sits at minus 9% for Microsoft, minus 15% for Google, minus 28% for Meta, minus 35% for Oracle. Only Amazon barely positive. The cleanest bear-case data on the page this quarter.

Same technology. Same models. Same vendor catalog. Opposite outcomes. The variable is the harness, and what nobody is saying out loud is that there are two harnesses in every AI deployment. The one the lab built. And the one the customer didn't.

The labs (OpenAI, Anthropic, Google) shipped their side from day one. Identity, billing, retries, tool registries, sandboxing, observability of the run, MCP as connective tissue. So when an agent runs, it does not stop until the budget does, and the budget is yours. Every call has a meter on it that only one party can read in real time. The CIO staring at a six-figure Anthropic invoice is not a story about a model that broke. It is a meter only one side was set up to see.

Anthropic answered the ROI question today with Opus 4.8. Sharper judgment to cut confident mistakes. Longer independent work to cut the restart tax. Same price to dodge the revalidation tax. A new Thinking effort selector at the task level. Aligned News framed it as the lab's direct response to enterprise complaints. They are not wrong about the features. They are wrong about which side of the agentic transaction those features live on. The Thinking effort selector is a new knob on the lab's dashboard, and you still have to know which setting to pick, for which task, with which operator behind it.

Salesforce shows what the buyer side looks like when it works. Four years of building Agentforce + Data Cloud + Einstein Trust Layer in parallel — Tunguz's seven components, prebuilt and sold to the customer in a box. So a Salesforce customer turning Agentforce on inherits governance, identity, observability, cost controls. 28.6 trillion tokens move because the harness sits on both sides — theirs and yours.

But even Salesforce's harness is a starter kit. A pre-built harness is a chassis. The work that actually moves a P&L is the tuning — the data architecture mapped to your business, the agent ownership written against your org chart, the budget calibrated to your risk tolerance, the workflows shaped around the specific operators in your specific chairs. Capability is a commodity now, the same way an F1 chassis is. You can't buy any of that tuning, and you have to build it for you.

And the Verstappen story refuses to let you off the hook on one more thing. When the chassis itself changes underneath you, the team that gets the jump on the new car wins. New regs dropped this F1 season, Mercedes and Ferrari got there first, and the four-time world champion is fighting for crumbs because last year's tune doesn't fit the new car. Stand pat with last year's harness when the underlying capability shifts, and the same thing happens to you.

The engine has more horsepower now. The driver is the same person you were last week. The chassis is the same one everyone else got. Build the cockpit. Tune it to you.

The best driver/harness combination wins.

At COAI today: the full Signal/Noise — with the eight questions every leadership team should be able to answer — is live at getcoai.com.

— Harry and Anthony

Sources: