|
Hello π
If AI saves you an hour and a colleague spends that hour fixing what you sent, nobody saved anything. This week: who's doing the fixing, and how to cut it.
|
In today’s UPSKILL
| π§Ή |
The hidden cost of other people's AI work |
| π§ |
Tip of the week: send AI work back with what to check |
| π¬π§ |
How much UK financial firms let AI agents do |
| π οΈ |
Data checks in Excel, sources in Google Sheets |
| β‘ |
Prompts: agent limits, checking numbers, urgent requests |
| π‘ |
Quick hits: Deloitte's flawed AI-assisted report and more |
|
|
|
45% of AI users spend an hour or more a week fixing other people's AI work.
The short version: A survey asked 1,068 full-time US employees in computer-based jobs about their workload, hours and use of AI. AI is saving people time, but a good chunk of that time then goes on colleagues fixing the work.
|
45%
of AI users spend an hour or more a week fixing other people's AI-generated work
|
53%
of AI users get AI work from colleagues that needs significant correction at least once a month
|
|
What AI is doing to the working week
Survey of 1,068 full-time US employees
Use AI tools for work
All employees surveyed |
|
|
Get colleagues' AI work that needs significant correction, at least monthly
AI users |
|
|
Say the work expected of them has gone up since they started using AI
AI users |
|
|
Spend an hour or more a week fixing other people's AI work
AI users |
|
|
Source: Hubstaff, State of the Workweek 2026.
|
The details:
| β |
Of AI users who gave a figure, 76% save at least an hour in a typical week and 32% save four hours or more. |
| β |
Only 26% of AI users take on more work with the time they save. 39% finish earlier, and 15% say nothing really changes. |
Why it matters: It's easy to count only the time saved by the person using the tool. Agree three rules with your team this week: whoever sends AI work checks the facts and figures first, they say which parts AI drafted, and anyone who has to rework it sends it back with a note. The tip below does the note for you.
|
|
|
Presented by G3NR8
β‘ Free every week
Get UPSKILL in your inbox.
What's actually working with AI at work, plus prompts and skills you can use straight away. One email a week.
Sign up free → |
|
|
Send AI work back with a note on what to check next time.
The short version: When someone in your team sends you AI work that needs heavy editing, don't just fix it quietly. Put their draft next to your fixed version and get AI to pull out the two things they should check next time.
| Paste into ChatGPT, Copilot or Claude. Fill in the brackets. |
|
Someone in my team sent me a draft they wrote with AI. I had to rework it before I could use it.
Their draft: [paste, with names, customer details and confidential figures removed]
My fixed version: [paste, with the same things removed]
1. List what I changed, grouped by type: wrong facts or figures, missing context, tone, length, structure, anything made up.
2. Name the two types that took the most fixing.
3. Write a friendly note to them, under 80 words, with two things to check before they send AI work next time. No blame, no lecture.
4. For each of the two, suggest one line they could add to any prompt for this kind of work. |
How to use it: Run it the next time you rewrite something one of your team sent you. Use the note to start a two-minute chat, and share your fixed version as the example to copy.
|
|
|
π¬π§
|
UK picture · Survey |
80% of UK financial firms now use AI agents. Only 2% let them run with little human oversight.
The short version: A survey of AI in UK financial services found most firms now run AI agents, but almost all keep a person in charge. The usual set-up: the agent recommends, a person decides and acts.
|
86%
of UK financial firms that responded use AI, compared with 75% in 2024
|
2%
let AI agents act with broad autonomy, with little human oversight while they work
|
|
How much UK financial firms let their AI agents do
Share of UK financial firms. A firm can run agents at more than one level.
| Decision support: the agent recommends, a person decides and acts |
|
|
| Constrained execution: the agent acts within set limits and rules |
|
|
| Conditional autonomy: the agent acts alone but needs approval for some actions |
|
|
| Broad autonomy: the agent acts with little human oversight while it works |
|
|
Source: Bank of England and FCA, Artificial intelligence in UK financial services 2026.
|
The details:
| β |
The most common use is everyday productivity, such as coding assistants and summarising tools. |
| β |
The biggest barrier outside regulation is fitting AI into existing processes and decisions (23%), followed by legacy systems (22%). In 2024, a shortage of talent and skills was in the top three. This year it isn't. |
| β |
40% name information security or operational resilience as a major constraint on using AI from outside suppliers. |
| β |
Only 39% monitor the impact of their AI on customers. |
Why it matters: Firms with some of the strictest regulators in the UK are using agents widely, and they still keep a person in charge of most decisions. Agents are arriving in everyday office tools too. When someone proposes one where you work, ask which of these four levels it will run at, who approves what it does, and who answers for its mistakes. Prompt 1 helps.
|
|
|
π οΈ
|
Trending AI tools |
Two spreadsheet features for checking data before you use it.
 |
#1
π A data check in Excel
|
Microsoft's latest Copilot update adds a clean data skill in Excel. Copilot scans a workbook for problems that are easy to miss, such as duplicates, inconsistent formatting and numbers stored as the wrong type, then suggests fixes in the file. You review and approve every change. It's part of Copilot in Excel.
| Try it Before you build a report from a spreadsheet someone sent you, ask Copilot in Excel to check it for data problems, then approve the fixes one at a time. |
|
 |
#2
π Sources for every cell in Google Sheets
|
Google says Fill with Gemini in Google Sheets now shows the web source behind each cell it fills in. It can also handle up to 3,000 cells in one request. It is rolling out on the web now and needs a Business Standard or Plus, Enterprise Standard or Plus, or Google AI Pro or Ultra plan.
| Try it Use it to fill a column of public facts you'd normally look up by hand, such as each competitor's headcount. Open five of the sources. If any is wrong, check them all. |
|
|
|
|
β‘
|
Practical AI · Prompts |
Three prompts for when AI is doing more of the work.
| 1 |
Decide how much an AI agent is allowed to do |
| Paste into ChatGPT, Copilot or Claude. Fill in the brackets. |
|
We want an AI agent to do this task: [describe it, no customer or personal details].
What it would touch: [systems, documents, or people it would contact].
What a mistake would cost: [money, customers, compliance, reputation].
Checks we already have in place: [approvals, policies, sign-off limits].
Use these four levels: 1) it recommends and a person decides and acts; 2) it acts within set limits and rules; 3) it acts alone but needs approval for some actions; 4) it acts with little real-time human involvement.
Recommend the lowest level that still saves real time, and say why.
List the actions that need a person's approval, what should be logged, who answers for its mistakes, and what results would move it up or down a level.
Finish with three questions to ask whoever is building or selling it. |
Why it works: you agree the limits and who owns the mistakes before anyone builds it.
| 2 |
Check the numbers in a report before you send it |
| Paste in the tables, not the whole report. Remove anything not yet public. |
|
These are the tables from a report I'm about to send: [paste the figures with their column letters and row numbers, or say each table starts at cell A1. No personal or customer data].
Don't check the sums yourself. Spot what needs checking and let the spreadsheet do the maths.
1. Find every total, subtotal, percentage and change figure.
2. For each one, write the Excel or Google Sheets formula I can use to check it against the figures it's built from.
3. Flag any number that appears twice with different values, any mix of units or time periods, and any group of percentages that should add up to 100, with a formula to check it.
4. Give me a checklist in the order I should work through it. |
Why it works: the spreadsheet does the sums, so you're not trusting a chatbot's maths.
| 3 |
Agree a check for urgent requests from senior people |
| Paste in, fill in the brackets, share the result with your team. |
|
Write a short rule for my team of [number] in [function] for urgent requests that seem to come from a senior leader, by email, chat, voice note or call.
Requests it covers: [e.g. payments, changes to bank details, sending files, password resets].
Include:
1. Warning signs to look for
2. How to check: call back on a number from our own directory, never one given in the message
3. A line someone can say to a senior leader to pause the request without it feeling awkward
4. Who to tell if it looks fake: [our IT security contact or channel]
Our existing policy, if we have one: [paste it, or say none]. Don't contradict it, and flag anything I should check with IT security.
Plain English, short enough to pin in the team chat. Make it clear nobody gets in trouble for checking. |
Why it works: people skip checks when the boss is in a hurry. A written rule gives them permission.
|
|
News, Insights & Trends
Four more stories worth knowing.
| π€¦ |
Bad AI: Western Australia's government is seeking its money back from Deloitte for a A$166,000 report on merging three universities. The report had AI-hallucinated references, and in one place two costs adding up to A$659m were totalled as A$399m, a A$260m gap (Guardian Australia) |
| πΌ |
Microsoft has opened voluntary exit packages to about 60,000 eligible staff, around 27% of its workforce, in the US, Canada, India and the UK, CNBC reports. CNBC says Microsoft is trying to balance a costly AI infrastructure build with capital efficiency. (CNBC) |
| π |
Senior managers are the most likely to act on a cloned CEO voice before checking it: 35% of them, against 16% of clerical staff, among office workers in Singapore. Prompt 3 above helps (SecurityBrief Asia) |
| π§° |
24% of US employees have built an app with AI that they use for work, and 53% of business leaders say their organisation has found at least one app that IT and leadership never knew about (Quickbase) |
|
|
|
G3NR8 is an impact consultancy for the exponential age. We help organisations implement AI for maximum returns on investment.
| β |
Train your teams: we've trained 150+ large organisations to use AI on their real work, from hands-on workshops to full AI activation programmes. |
| β |
Get your leadership team aligned, with a shared plan for AI and clear decisions on what happens next. |
| β |
Improve operational efficiency with OEX, our Operational Excellence platform. |
|
|
|
