Prime Ai Solutions
Read time: 4 minutes
Some people give AI a task at night. In the morning, the work is done.
Most people go back and forth with a chatbot all day.
The model is not the difference. The difference is one question: how does the AI know the work is right?
Below: what the best AI builders do, and a prompt you can use today.
LATEST NEWS
What the people with results do differently
1. They give AI a way to check its work
The problem: AI says the commentary is done.
At 11pm, you find the wrong number.
AI stops when the work looks finished. If nothing checks it, you are the check.
What the best builders say: Give AI a way to check its own work.
Andrej Karpathy says AI is strongest where work is verifiable.
Boris Cherny, who created Claude Code, calls it his most important tip.
He estimates it makes results 2 to 3 times better.
Dharmesh Shah, co-founder of HubSpot, makes it simple.
A grade is the AI's opinion: "Is this clear? 9 out of 10."
A check is pass or fail: "Is it under 150 words?"
The AI cannot argue its way past a check.
My take: finance teams have an advantage.
Most of our work is already pass or fail. The balance sheet balances, or it does not. We already have a word for this. Reconciliation.
But most teams ask AI for a grade: "Check this looks right."
Ask for checks instead. They work for sales proposals and board papers too.
I'm not saying this from the sidelines.
In the last month I've built three things.
A database that matches recruiters with job seekers.
An accounting agent, which works as your book keeper + on demand, now in private beta
And, for fun, a tool that renders and furnishes properties.
But a check has a limit. It can only check against what is written down.
2. They give AI the real rules
The problem: your policy says a manager approves credit notes.
But everyone knows anything over £10k goes to the FD.
Everyone knows one big customer always pays at 60 days, and that is fine.
AI does not know this.
It flags the wrong customer in your aged debt review.
Then it checks its work against the policy, and the check passes.
The answer is still wrong.
What the best builders say: We call this tribal knowledge. It is the know-how a company runs on that nobody wrote down. AI can only use what it can find.
My take: skip the big documentation project.
Start with month-end, before Christmas.
Before your most experienced person goes on leave, give AI 20 minutes with them:
You are interviewing our [financial controller] about how month-end really works.
Ask one question at a time.
Focus on exceptions, workarounds and anything that is not in the close checklist.
After 10 questions, write a list of the rules you learned.
Next to each rule, write who told you.Then paste that list into a Claude or ChatGPT Project. Every prompt you run in that Project can now use the real rules.
Now put the two ideas together. That is this week's prompt.
STEAL THIS
The month-end commentary that checks itself
Most people write this:
"Write the October variance commentary. Make sure it is accurate."
"Make sure it is accurate" asks for a grade. The AI will say yes.
This version from my course uses both ideas.
Run it in the Project with your rules.
The AI must pass the checks to finish.
If an answer is not written down, it must ask.
Write the October variance commentary from the attached management accounts and budget.
Before you say it is done, do these checks. Show me a pass/fail table.
TIE IT OUT
- Every number in the commentary is in the attached file. List any that are not.
- Every variance over £[10k] or [5]% has an explanation. List any you missed.
- For every line, actual minus budget equals the variance you quote.
- The commentary is under [200] words.
SOURCE IT
- After each explanation, give the tab and row it came from.
- If the reason for a variance is not in the file or my notes, write REASON NEEDED. Do not guess.
READ IT LIKE THE BOARD
- Read it as a board member who is not in finance. List each word they would not understand. Then fix it.
Fix all failures. Do the checks again. Then give me the final version.What comes back (illustrative):
Numbers match source file ........ PASS
All variances over £10k explained . FAIL: Travel +£14k, REASON NEEDED
Variance maths ................... PASS
Under 200 words .................. PASS (176)That FAIL is the valuable part.
The AI did not invent a reason. It asked.
Add the answer to your Project, and next month it knows.
Checks only find what your rules cover. You still sign it off, so agree board totals in Excel yourself.
The full version goes deeper: the RACEF framework, a 64-prompt library, variance commentary templates, board pack narratives and forecasting playbooks. Currently available at a discounted rate for a limited time. Check it out here.
SIGNAL / NOISE
Signal: only 3% of the FTSE 100 show what AI earned them. Flexera reviewed their 2025 annual reports this week (via Accountancy Age). Only 2% disclose what they spend on AI.
None show it as its own cost line.
Start now: track your total AI cost and one result.Noise: "Everyone already uses AI agents."
A new Bank of England and FCA survey says 80% of UK financial firms already deploy them. Put that next to the 3% above.
Using AI is not the same as getting value from it.
That's it for this week.
One question connects all of it: how will the AI know it is right?
If this prompt saves you time at month-end, the course has 64 more prompts for finance teams. Take a look here.
-Umar Prime AI | primeai.solutions
