Prime Ai Solutions
Read time: 4 minutes
Leader, Leader, Leader….Welcome back.
AI is saving your team time. Can you show where that time went?
Most companies cannot. This week: why that matters, what Anthropic's IPO filing says about your AI bill, and a prompt you can use today.
LATEST NEWS
Where does the time AI saves go?
The short version: Companies are saving time with AI but not turning it into results. The fix is to decide what "success" means before you start.
On Monday, Gartner published new research. Four numbers stand out.
3%. That's the share of companies getting more savings from AI productivity than they planned. Yet three in four board members expect AI to deliver them.
38%. That's the share of companies already over their AI budget. And 85% plan to spend more next year.
21%. That's how much of the time AI saves goes into more valuable work. Nobody can say where the rest goes.
78%. That's the share of HR directors who say their leaders don't agree on where AI creates value.
Apollo's chief economist, Torsten Slok, sees the same thing in company results. The 7 biggest tech firms are earning more per employee. The other 493 companies in the S&P 500 are not.
This is not new. Computers and electricity also took years to show up in productivity numbers. The gains came when companies changed how they worked, not when they bought the technology.
One company that got it right
Bayer's crop science team had a simple problem. Researchers waited hours for engineers to answer questions about their data.
So Bayer built a set of 22 AI agents that answer those questions in plain English. In the first phase, the agents answered about 80% of questions without an engineer. One diagnosis went from over an hour to under 10 minutes.
The important part: Bayer picked one problem and measured one number. It did not just count "hours saved".
Do this: Before your next AI project, write down three things. What must change? What number will prove it? Who checks the result?
Anthropic opened its books. Here's what it means for your AI bill..
The short version: AI is getting much cheaper to run, but companies are using much more of it. Your AI bill will grow unless someone owns it.
On Monday, Anthropic, the company behind Claude, filed to go public.
In 2023 it made $41m.
In the first half of 2026 it made $16.3bn!
Three lessons for your business:
Lesson 1: Don't trust the headline number.
The filing shows an $80.7bn loss for the half. But $72.4bn of that is a one-time accounting charge. No cash left the business.
In fact, Anthropic made its first operating profit in Q2: $974m.
Do this: Before you repeat any big number, ask two questions.
Is it cash? Will it happen again?
Lesson 2: AI is getting cheaper, fast.
In 2023, Anthropic spent $10.56 on computing power for every $1 it earned. By early 2026, it spent 55 cents. This week OpenAI also released a model at half the price of the one before.
Do this: Keep a list of AI ideas you dropped because they were too expensive or not good enough. Try them again every three months.
Lesson 3: Your AI bill is becoming a usage bill.
80% of Anthropic's revenue comes from customers who pay for what they use, not a fixed fee per person. Those customers now spend about 9 times more than a year ago.
Now remember Gartner: 38% of companies are over budget, and only 3% beat their savings targets. Spend is rising much faster than results.
Do this: Treat AI like your electricity bill. Give it an owner, a monthly limit, and track the cost of each task.
To be fair, the filing also shows big risks for Anthropic. It owes hundreds of billions of dollars in computing contracts, and five customers bring in about a quarter of its revenue.
STEAL THIS
The one-page AI scorecard
The short version: This prompt makes you answer the questions from stories 1 and 2 before you spend money. Copy it, fill in the brackets, and paste it into any AI tool.
You are a finance business partner.
You must defend this AI project to the board.
Be sceptical.
I want to use AI for: [the task, e.g. answering supplier queries]
Today the task takes: [hours per week, and who does it]
The problem this causes: [e.g. month-end close is 2 days late]
Our AI cost today: [monthly spend]
Do these five things:
1. Give me ONE number that shows this worked. Not "hours saved".
2. Tell me how to measure that number this week, before we start.
3. Ask me where the saved hours will go. If my answer is vague, ask again.
4. Estimate the monthly AI cost if usage grows 5 times. Suggest a limit and an owner.
5. List 3 ways this project could look successful but fail.
Example of a good number: "Share of research questions answered without an engineer." Bayer measured this.
Format: one page, plain English. Start with a table: the number, today's baseline, the target.What you get: a one-page plan you can take to your board. It has a clear target, a cost limit, and the risks already listed.
It also uses RACEF, the framework I teach: Role, Action, Context, Examples, Format.
QUICK HITS OF THE WEEK
GPT-6.1 Terra. OpenAI's new model can use a computer like a person: it reads the screen, clicks and types. OpenAI's example is entering invoices into accounting software.
It's on the $20 Plus plan.
One warning: OpenAI's own tests found it sometimes ignored its limits. Let it enter invoices. Don't let it approve payments.Deloitte, again. The Western Australian government wants money back for a Deloitte report. It had references that AI invented and a calculation error of about $260m. If your name is on the work, the AI's mistakes are your mistakes.
Claude for Legal: ready-made legal tools.
Yesterday, Anthropic showcased templates for contract review, citation checks and scans of upcoming regulation.
You start from an example, add your own documents, and share the result with your team.
The point: AI is moving from answering questions to doing the work.
Do this: Test it before you trust it.
Run the citation check on a document you have already checked by hand, and see if it finds the same problems.
Use a document with no client data, and check your firm's data rules before you upload client files!
Here's the one thing to take from this week. AI is not failing. Most companies never decided what success looks like. The 3% did.
If this was useful, forward it to one person who approves AI spend. The scorecard could save them from a bad budget decision.
-Umar, Prime AI | primeai.solutions
