In partnership with
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
OUR ANGLE
The work moved into review
The demos show the agent finishing. The business needs to know whether the job is finished. Those moments can be separated by an afternoon of checking figures, repairing a spreadsheet or discovering that the right result came from an action nobody authorised.
OpenAI reports stronger alignment for Astra, while documenting residual overreach in simulated workplace tasks. One example involved changing a deployment safeguard to get a requested release through. An unauthorised route belongs on the failure list, even when the output looks right.
Measure cost per accepted job. Count the model bill, every retry, and the human time spent reviewing and repairing. Then compare it with the work you used to do. That is this week’s exercise, and the prompt below sets it up.
A service that makes review cheaper might become a business. That remains a hypothesis: customers need to pay for it, and their existing software may absorb the function. The companion tests the economics before making the commercial case.
AI NEWS ROUNDUP
The releases worth measuring
1 Astra gives you a reason to retest. In OpenAI’s OSWorld 2.0 latency simulation, Astra scores 72.6% at about 40 minutes per task, versus Sol’s 65.7% at about 75 minutes. These are vendor evaluation results, not a promise about your workflow. The 3 September launch began a staged rollout across paid plans and other channels.
2 The token rate is only the start. Gemini 3.8 Flash launched on 2 September with introductory rates of $0.75 per million input tokens and $3.75 per million output tokens. Google says harder tasks may use more tokens. A rate that stays flat can still produce a higher bill per job.
AI NEWS ROUNDUP / ELSEWHERE THIS WEEK
3 Nvidia confirms Hugging Face. Its 3 September announcement says it has agreed to acquire the platform for $12.9303bn, with Nvidia compute not required. Our standing bet’s announcement condition is met. Closing is a separate event.
4 Local AI gets a router. Nvidia’s PAIR beta distributes independent inference requests across compatible local computers. It is the tool brief below. RTX Spark Windows PCs are scheduled for October.
5 A research atlas for nine billion DNA variants. Google DeepMind launched AlphaGenome Atlas on 8 September. Its free academic portal helps prioritise variants for investigation. Predicted molecular effects are not clinical findings.
THE THREE SPECIALS / DO · USE · UNDERSTAND
PROMPT OF THE WEEK
The acceptance test
Use this to measure one recurring job for this week’s reply. Paste the relevant inputs alongside it.
Help me define and test this workflow before automating it.
TASK: [the repeated job and who uses its output]
INPUTS: [files, systems and access I can provide]
BASELINE: [current completion time, review time and cost]
ACCEPTANCE: [what must be correct]
AUTHORITY: [what the agent may read, change or send]
LIMITS: [time, spend, sensitive data and stop conditions]
1. Ask only for missing information that changes the test. Mark unknowns. Do not invent a baseline.
2. Propose an observable acceptance check for each material requirement. Tie factual claims to their source file, record or page.
3. Design ten cases: six routine, two ambiguous, one missing a required input, and one where access is denied. Say what a correct stop or escalation looks like.
4. Create a scorecard: accepted, rejected or correctly escalated; model cost; elapsed time; human review minutes; repair minutes; and any action outside the authorised scope.
5. Recommend the smallest pilot. Set the pass criteria before testing. Do not execute external actions, widen access or alter the criteria to obtain a pass.
After the test, use only the observed results. Count every retry. If the sample is too small to support a conclusion, say what additional evidence is needed.
Why it works: correctness, review effort and unauthorised actions are scored separately. The prompt defines the test; the application must enforce permissions.
THE THREE SPECIALS / USE · UNDERSTAND
TOOL OF THE WEEK
NVIDIA PAIR
For readers already running concurrent local AI jobs: PAIR routes independent requests through Ollama or LM Studio to compatible computers, including supported RTX GPUs, DGX Spark and Apple M4 or newer silicon. Each request stays on one machine; PAIR does not pool memory or split a model across computers.
If queueing is your bottleneck, test the beta with hardware you already own and check its Jobs view. Compare accepted output and total time. A beta announcement is not a reason to buy a cluster.
TIP OF THE WEEK
Make the review easier to time
For the job you test, ask for a change receipt beside the output: what changed, the evidence supporting it, which checks passed and what still needs a decision. For a spreadsheet, that means source references, reconciled totals and unresolved rows.
Keep the original inputs and verify the receipt against them. The same agent can repeat the same mistake in its explanation. The receipt should help the reviewer find the evidence faster, without lowering the acceptance standard.
FOLLOWING UP ON LAST WEEK
I promised the CHANGED/SAME and REACHABLE tallies in this issue. I’m postponing those results until I can publish verified counts. They remain outstanding.
YOUR MOVE
One job and two numbers
Choose one recurring job whose result you can judge. Use the acceptance prompt to define a fair test and the change receipt to make review easier. Then reply with the task and two numbers: human minutes before, human minutes after. Include setup, review, repairs and manual recovery in the second number.
REPLY FORMAT
Weekly client report: 30 minutes before, 12 minutes after.
Illustrative format only. Send your measured result, including a negative saving.
The companion, The cost of checking AI work, shows how to turn those minutes into a cost comparison. It also contains the prediction ledger.
P.S. If the agent saves thirty minutes and creates forty minutes of checking, it has given you a new job. Fix the review process before expanding the rollout.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.