Hi, it's Andreas here.
This week's edition is about AI ROI. The short version: there are three kinds, and saved time is not money yet.

This newsletter recently passed 40,000 subscribers. As a thank you, I arranged software deals for you: my software deals page has more than 600 offers. The free ones unlock with your email, the rest with the code HITL (you need to be subscribed here). Cohort members get additional deals, and they have already saved more than $100,000. I am planning to add more, so stay tuned.

Next, I’m evaluating a 12-week AI leadership program for the people who own AI in their company. A handful of readers asked for it, and so did people in the cohort.

It would be hands-on, because the gap I see most often is leaders who decide about AI without having built with it. The topics: change management, data analytics, models and agents, software development with agents, and enterprise deployment with risk and governance.

Would you join? One click below. If you vote yes, add your role and company in the comment box and you'll hear about it first.

In this issue:

Anthropic releases Claude Fable 5.2 and Mythos 5.2 at a lower price → Both run on the same model, up to 30% faster than 5.1, at $8 per million input and $40 per million output tokens, down from Fable 5.1's $10 and $50. Fable ships to all users with cyber and biology safeguards, Mythos only to verified organizations. For most coding and knowledge work, Anthropic recommends starting with Opus 5.5.

US agencies surveilled an executive close to Jensen Huang, the New York Times reports → They suspected the executive, linked to Nvidia customer Megaspeed, could be a Chinese asset and worried about Huang's access to President Trump. Nvidia disputes it.

Anthropic starts its IPO roadshow at $70 to $78 a share → The offering is 825 million shares, and Anthropic has applied to list on Nasdaq under the ticker ANT. At $78 it would be worth about $1.4tn fully diluted, CNBC calculates. The amended filing puts run-rate revenue at about $81B at the end of October, up $10B in a month.

AI stocks lose more than $2tn in a day after the US midterms → The Nasdaq 100 fell about 6% on Nov 5, its worst day since April 2025. Two days earlier, voters backed limits in 40 of 41 local data-center measures.

OpenAI ships GPT-6 Cyber and rates it Critical for cybersecurity → In a mid-October run, OpenAI says, its unreleased cyber model built 948 Linux root exploits for about $1.2M in API credits. Only approved customers get access.

Revolut says self-hosted open models cut its AI coding costs about 90% → GLM now serves roughly 2,000 engineers, on Revolut's own figures. Also new: Kimi K3.1 Pro open weights and Mistral Studio managed agents for on-prem.

Also this week:

“Where is the ROI?” is the question I hear most about AI, especially from CFOs, and especially now that budget planning is underway. I believe there is a plain reason for it. The cost of AI comes first and is easy to see: licenses, tokens, cloud bills, the people who build. The benefit comes later, and most of it is indirect. So the person who signs the budget sees a growing cost item, hears about saved time, and asks where the bottom-line impact is. That is fair.

But the question is always focused on one thing: getting the money back. What we spent, and what came back, as one number. My view is that there is no such number, because there are three different returns. Here is how I sort them:

Each return pays in a different kind of value: non-financial for the first, financial for the second, strategic for the third. I think about the money behind them in two colors. Blue money is value that does not reach the P&L by itself, such as time saved or a better customer score. Green money is revenue or cost reduction on a line you can point to. Cost avoidance stays blue unless it removes a line that is already in the budget, such as an approved hire.

Gartner research published in October found that 77% of board members expect AI to deliver productivity and cost savings, and 3% of organizations exceed their cost savings goals through productivity. As I read it, boards expect green money, and productivity on its own pays in blue.

I ran several workshops with companies in October. And when a reason for AI came up at all, it was productivity. But I never heard anything about how it was measured, or against what. A license for every employee buys you the first return. If the business case that got it approved promised savings, it promised the second.

Below: how blue money turns green, how I measure ROI on one workflow and what to use when it is not one workflow.

Take a team of ten where everyone saves five hours a week. That is fifty hours, more than one full-time job. This will not affect anything on the P&L.

It moves only when someone decides what the fifty hours are for. The team takes on more paid work with the same people, or work that went to a contractor comes back in-house. Both are decisions someone has to own.

If nobody decides, the team absorbs the time. And a few months later the saved hours are becoming a new normal.

More output has the same limit. If AI lets your engineers ship twice as much, customers still have to want it and pay for it.

Nevertheless, of the three returns, everybody asks of course only about the return on investment, and I get it. Companies have been burning real money on AI for three years. So how do you measure it?

There are many ways: revenue uplift, hours saved times an hourly rate, customer satisfaction, adoption rate, an AI literacy score. Across a whole company, most of them are too imprecise, because you cannot say what AI caused. Trust me, we spent hundreds of hours trying several approaches, and we always ended up measuring via one workflow at a time. It is well proven and, in my view, the cleanest way. A workflow has a start, an end, a volume and a cost, so you can measure it before and after. A license for every employee has only the cost.

Start from the number you want to move: more capacity, a cost line that goes away, a faster cycle or more revenue. "Where can we use AI?" is here the wrong first question. Then pick a workflow where you can prove what happened. Its numbers already exist, few other things change at the same time, and it sits close to revenue or cost. That is why many companies start in sales, customer service or finance.

Very important: before you start, name one owner in the business who runs the workflow and answers for its numbers. A central AI team can help with budget, tools and people who build, and IT has to open the systems and the data. But the head of the function should still own the workflow and report the result. Then use the metrics the function is already managed on. You do not need new ones.

Then you need to follow six steps: choose the workflow and its owner, measure it before, redesign it with AI, let it run for 30 to 60 days and measure it after, count what the AI costs to run, and put the result in money. The before numbers are usually a cost, a time or a quality measure. The hours people spend checking the output count as well. In the example below they sit inside the team hours.

The same six steps for supplier invoices, with example numbers: 4,000 invoices a month, 12 minutes each before the redesign and 4 minutes after.

Work on the invoices drops from 800 hours a month to 267. Of the hours that go away, 320 were temp hours, so the temp contract ends and $8,000 a month leaves the P&L. After $1,500 for the AI, $6,500 a month is green money. The 213 hours the team gets back stay blue until someone decides what they are for, and so do the two days of cycle time. Add the cost to build: at $45,000, it pays back in about seven months.

Most AI spend is not one workflow. For a whole function, watch the same two to four metrics and compare them two ways: before and after AI came in, and heavy AI users against light ones. Other things change at the same time, and heavy users are often your best people already, so the gap overstates what AI did.

For licenses across the company, turn the question around: what has to be true for this to pay? Say AI costs $200 a month per person and a person costs $120,000 a year. That is 2% of payroll, so a team of 50 has to get done what 51 did before. You cannot trace every saved minute, but you can say whether that looks true. It stays blue money, and now you know how much of it you need.

Across the whole company, the proof should also get harder with time. In the first six months you mostly have what managers observe. From six to twelve months, single workflows should have changed: a vendor is gone, or a team takes on more scope. From one to two years, it should show in team budgets and hiring. After that, in new revenue or the company's main numbers.

Before an AI initiative receives a budget, you should clarify three points related to it:

Then judge each initiative by its own measure. An employee return gets a fixed budget, like laptops and training. Judge it on whether people use the tools and, more than that, on what they now do with the time. Stop asking it for a P&L case. An investment return needs the baseline and a P&L line with a date on it. A future return gets a small budget and a kill date.

The same three lines work for one use case on one team. Before you roll out an assistant, write down what the saved time is for, in a sentence a CFO can use: "We close the month two days earlier and no longer pay for temp help."

One camp funds nothing until it proves a saving. The other measures nothing because the return looks obvious. Both treat three returns as one. In the grey zone between them, each return gets its own test.

Tools for everyone are worth the money. People build the skill there, and without that skill nobody in the company can do the redesign. But they are a cost of doing business and should be budgeted like one. I run AI workshops, and a workshop buys the same first return. It pays when the people in the room go back and change how a process runs.

The money is in the second return, and that is redesign work. Someone has to decide which steps go and what the people do instead. That is the deployment work I wrote about last week. It is why the return shows up late, and only where someone owned the change.

Anthropic explains how agent memory works → A wiki built over 128 "dreaming" sessions scored 48% alone, below plain search at 53%. Search and wiki together reached 61%.

DeepLearning.AI and Arize's course on evaluating AI agents → Laurie Voss covers tracing with open-source Phoenix, turning failures into test sets, and measuring whether an LLM judge is reliable. About 15 hours.

Amazon's memo on what it now expects from engineers → It calls token usage and AI tool adoption "deeply flawed" as metrics, while noting they still correlate with the outputs that matter.

Alexey Grigorev's Coding Agent Setup Kit → A free 14-page template with an AGENTS.md, skills, a session prompt and a project process. Email required.

SPONSORED BY ALTERYX
Alteryx's free Nov 20 webinar on checking AI answers inside the chat
→ A live look at its November release: data lineage and trust signals in the AI chat. Its VURA framework asks if outputs are visible, understandable, repeatable and auditable.

I take on a small number of advisory engagements each quarter for AI programs that are stuck, or for a second read of an AI proposal before you sign. Reply to this email with the one initiative you would want a second pair of eyes on. And if you want to build agents rather than read about them, the waitlist for the fourth Agentic AI Cohort is open, and you can start the Build Sprint today.

Which return did your last AI business case promise: employee, investment or future? Reply with one word. I read every reply.

— Andreas