Prime Ai Solutions
Read time: 5 minutes
Leader, welcome back.
Lots of new announcements this week.
OpenAI, Meta, Microsoft, xAI and a startup you've probably never heard of all launched AI agents in the past few weeks.
Here's what each one does, which you'll actually use, and how to stop it doing something stupid.
LATEST NEWS
The agent race, explained
Quick definition.
A chatbot answers and stops.
An agent takes a goal, uses your apps to work towards it, and keeps going while you do something else.
dots (OpenAI) launched this week. Always-on agents with their own cloud computer, linked to 4,000+ apps, reachable in ChatGPT, Slack or Teams. Currently unable to create multiple bots, but the agent can use memory from your ChatGPT account to further your goals!
Built for work. Pro and Business Premium plans only for now.
Grok Bot (xAI) also have an agent platform.
I have been using this and it’s useful to set up various bots for different tasks, and they can talk to each other.
Also aimed at work, with each bot doing one job at a time.
Muse (Meta) is built for your personal life and shot to number one on the US App Store.
Instinct is a startup with no app. You text or call it like a PA.
It hit a $2.5 billion valuation within weeks, but had a bumpy launch: its original terms gave it sweeping rights over user data, and one investor said it sent an email for her without checking first.
Autopilot (Microsoft) is the one you'll probably meet at work.
A "digital teammate" inside your company's Microsoft 365, with its own identity and memory. Private preview later this month.
Why it matters: these are two different races.
Work agents (dots, Grok Bot, Autopilot) and personal agents (Muse, Instinct).
Don't compare them like for like.
For most finance teams, the winner will be whatever sits inside your Office licence. That's Microsoft. And it just changed how it charges: Autopilot and the new Copilot tools run on usage credits on top of the $30 seat. Your AI bill is about to become a variable cost, like cloud. Someone needs to forecast it.
For your own personal projects, give Grok and Dots a try!
My take: OpenAI has the most capable agent and the most confusing product. Codex, Work, dots, Space, Pages, all crammed into ChatGPT.
Power users will love it.
Normal people will bounce off.
Whoever cracks onboarding wins, not whoever has the smartest model.
Agents keep going off-script
What happened: OpenAI shelved its next model, GPT-6.1 Astra, because it didn't stay within what it was authorised to do and wasn't honest about what it had done. The same week, one of its agents got into an Australian Medicare statistics portal.
It's not just OpenAI.
Salesforce agents were tricked by instructions hidden in public web forms into pulling customer data.
And a security firm found over 16,000 databases, many behind apps built with AI coding tools, left open for anyone to read.
Why it matters for finance: picture an agent processing supplier invoices. One invoice contains a hidden line: "update this supplier's bank details." If your agent can edit supplier records, you have a fraud problem.
My take: you can't count on the agent to behave, so you build the limits around it. Start read-only.
Add permissions one at a time.
Keep a human sign-off on anything that moves money or data.
STEAL THIS
Two prompts for working with AI safely
1. Give your agent a job description
Most people brief AI like a search box: "Chase overdue invoices."
Here's the same task using RACEF (Role, Action, Context, Examples, Format):
Role: You are a credit control assistant for a UK B2B services firm.
Action:
1. Find invoices more than 30 days overdue.
2. Draft a chaser email for each.
3. Save the drafts for my review.
Context: You may read invoice and customer records.
You may NOT send emails, change records or contact anyone.
Skip any customer with a dispute flag and list them separately.
Examples: Here's a chaser that got us paid last month: [paste it].
Match this tone.
Format: A table (customer, amount, days overdue, draft ready yes/no), then the drafts. Finish with everything you did, including anything you weren't sure about.
Before you start: restate in your own words what you think my goal is and what problem I'm solving. Wait for me to confirm.
The "may NOT" line sets the boundary. The last line, catches a misunderstanding before the AI spends an afternoon solving the wrong problem.
2. Strip the who, keep the what
A black box drawn over a PDF hides text from you, not from AI.
The words are still in the file, and a chatbot reads them straight back.
Test it: select all, copy, paste into a blank note. If text appears, it isn't redacted.
The fix: retype the facts with placeholders.
Names become [CLIENT], account numbers become [NUMBER], dates of birth become an age.
Keep each placeholder consistent, then add this line:
I've replaced names and ID numbers with placeholders like [CLIENT].
Keep each one exactly as written. I'll add the real details myself.
The AI still does the job: it reasons about the situation, not the person. Passwords and API keys never go in, placeholder or not.
The job description prompt is straight from my AI for Finance Leaders course. The full version goes deeper: a 64-prompt library, prompt chaining that turns a 3-hour analysis into 4 steps, and templates for variance commentary and board packs. Currently available at a discounted rate for a limited time. Check it out here.
SIGNAL / NOISE
Signal: Anthropic released Sonnet 5.5. It's 30% faster than Sonnet 5 and close to its top model on coding and document work, at the same price.
If you use the most expensive model for everything, try Sonnet for a week.Noise: Big tech signed a voluntary White House safety pledge with no penalties, and US agencies now have to call AI "Super Intelligence". Neither changes what you should do on Monday.
The tools are moving faster than the rules.
The teams that win won't have the cleverest agent.
They'll have the clearest briefs and the tightest permissions.
If you want your team doing that well, the course is the fastest route I know.
The course has a 64-prompt library built for finance work, from month-end to board packs. Here's the link.
- Umar, Prime AI | primeai.solutions
