AI ADVANTAGE WEEKLY · ISSUE 001 · MONDAY 19 OCTOBER 2026 · 9 MINUTE READ
Hello from the desk, and welcome to the first Sentient Sky. Every week we test one AI task in a real small firm and show you the receipts, including what went wrong. This week starts where every firm should start: the rules. We had AI draft ours, checked every line, and turned the result into a one-page policy you can copy.
INSIDE
The Build: we let AI draft our AI rules, then checked them
Intelligence: three things worth your time, two you can skip
The Number: how many of the smallest UK firms use AI
Your move: the 9pm Test
IN PARTNERSHIP WITH
Tired of news that tells you what to think? 1440 is a free 5-minute daily brief with just the facts. No spin, no clickbait. I read it, and it’s worth a look. - Niku
Every headline satisfies an opinion. Except ours.
Remember when the news was about what happened, not how to feel about it? 1440's Daily Digest is bringing that back. Every morning, they sift through 100+ sources to deliver a concise, unbiased briefing — no pundits, no paywalls, no politics. Just the facts, all in five minutes. For free.
THE BUILD
What it got right, what it missed, and the one-page policy you can adapt for your firm.
Checked by Niku Dutta against the sources in The Receipt
RECEIPT · THE BUILD · No. 001
TESTED 19 OCT 2026 · 1 FIRM, 5 PERSON · 1 AI TOOL
2 DRAFTS · 3 PROBLEMS FOUND · 7 SOURCES CHECKED
IN SHORT
We gave an AI assistant our firm's facts and one instruction, and asked for a one-page AI policy. A person checked every line before anyone signed it.
The first draft took 4 minutes. Checking it took 32. It left health data off the red list, and the check caught it. Two smaller slips are logged below.
WORTH IT IF
Anyone in your firm uses ChatGPT, Copilot, Gemini or Claude for work, and nothing is written down yet.
SKIP IT IF
You already have a written AI policy that staff signed in the last six months. Use the kit's check sheet to review it instead.
THE BUILD CARD
FIELD | ISSUE 001 |
|---|---|
Task | Write the firm's one-page AI policy (Task Map T-ALL-01) |
Firm | Sentient Sky (our own firm) |
Tool and plan | Claude, Team plan |
Data colour | Green: built from firm facts, no client data |
Risk tier | Low |
Set-up | 14 minutes in our test |
Run | 4 minutes to the first draft |
Monthly cost | Free for the template. Drafting works on a free AI plan; we used a paid one and recommend it for higher limits and stronger privacy and data terms. |
Drafts logged | 2 |
Human check | The owner reads every line against the kit's sources page, then signs |
Verdict | Use, with the check |
Test date | Monday 19 October 2026 |
Client data leaves the firm? No. You draft from your firm's facts, not client files. Who approves: the owner, who signs it.
§ What we tested
Most small firms using AI have no written rules for it. In a 2026 survey of North American accounting firms that use AI, 87% had no formal written AI policy.1 In a survey of US legal professionals, 57% of solo practitioners and 55% of small firms had no AI policy.2 Both are vendor surveys, so read them as a signal, not a census.
Our test firm this week is our own: Sentient Sky, a consultancy and the publisher of this newsletter. It is not an accounting or law practice, so we wrote the policy to work for those firms too, and the kit marks the parts that differ.
We listed every AI tool we use and what each one touches. We gave that list and our firm's facts to Claude (Team plan) with the instruction below. Then we checked the draft line by line against the published guidance listed in the kit.3 We logged every problem: what it said, what was wrong, and how we caught it.4
“Nothing went into the policy unread.”
§ How to set it up
STEP 1 OF 4
List every AI tool your firm touches
Ten minutes. Include the free ones people use on their phones. For each tool: who uses it, the plan (free or business), the account (work or personal), and what goes into it. The kit has the table.
STEP 2 OF 4
Paste in one instruction, word for word
Five minutes. Sign in with a work account, never a personal one. Paid is recommended, free works. Give it your firm’s facts and the tool list. No client files. Fill in only the parts in <angle brackets>. Leave the [ASK] and [CHECK WITH ADVISER] lines alone.
STEP 3 OF 4
Check the draft against the traffic light
Fifteen minutes. Ours took longer because we logged every problem. Read every line. Strike out any law, rule or regulator the draft names that you cannot find on the kit's sources page. Then check the client-data rules: red never goes in, amber only into approved business tools, green is fine.
STEP 4 OF 4
Ten minutes. The owner signs. Everyone who uses AI at work reads it and signs the sheet. Put a review in the diary for six months from today, or sooner if a tool changes its terms.
THE INSTRUCTION, AS WE RAN IT
You are helping a small professional firm write its first AI policy. Use ONLY the facts I give you below. If something you need is missing, write [ASK: what you need] instead of guessing. Do not name any law, regulator, guidance or standard unless it appears in my facts. If you think one applies, write [CHECK WITH ADVISER: topic].
Write in plain English, under 450 words, for busy staff. Use exactly these headings: Who this covers; Tools we allow; Client data; The human check; Telling clients; Mistakes and incidents; Training and review. Under Client data, use three levels: red (never enter), amber (only in approved business tools, minimum needed), green (fine to use).
Then give me:
A. The policy.
B. An approved-tools table with these columns: Tool and plan; Signed in with; Allowed for; Never for; Trains on our data?; Checked on (date).
C. A list of every [ASK] and [CHECK WITH ADVISER] you wrote.
My firm's facts:
<Firm name> is a <type of firm> in <country> with <number> people. We use these AI tools: <tool, plan, who uses it, what for>. Our clients are <who they are>. We hold <types of client data>. The person who approves AI use is <name, role>.§ What went wrong
Every problem we found, typed, and how we found it. The counts come from our log: one firm, one afternoon.4
TYPE | WHAT WENT WRONG | TIMES | CAUGHT BY |
|---|---|---|---|
F06 Missing step or omission | Left health and other special category data off the red list, which the kit's traffic light treats as red | 1 | Line-by-line read |
F09 Length or format | Asked to list every [ASK] and [CHECK WITH ADVISER] note, it merged the two identical "date checked" cells into one line. | 1 | Line-by-line read |
F12 Other | Said free personal AI accounts are never used for firm work, then asked whether a free chatbot could stay in use for a firm purpose. | 1 | Line-by-line read |
Type codes: F01 Invented source · F02 Wrong figure · F03 Wrong name or client · F04 Wrong date or deadline · F05 Wrong jurisdiction · F06 Missing step or omission · F07 Overclaim · F08 Tone · F09 Length or format · F10 Data leak risk · F11 Refused or stalled · F12 Other.
Caught by: Line-by-line read · Source check · Ledger spot-check · Named reviewer · Second tool · Client · Not caught.
THE HUMAN CHECK
Nothing went into the policy unread. If you copy this, keep that rule for every AI output in your firm: a named person checks facts, figures, names and citations before anything reaches a client, a court, a regulator or a tax authority.
Verdict: Use, with the check. The draft took 4 minutes and invented nothing; the 32-minute check found the three things it got wrong.
Tell AI not to guess and it stops making things up, but it starts leaving things out, which is quieter and easier to miss.
KIT 001
THE KIT · FREE
The One-Page AI Policy kit
The policy template, an approved-tools list, the client-data traffic light, a human-check sheet and a staff sign-off sheet, with separate notes for UK and US firms.
Word, 13 pages. Free to adapt and share.
AD · OUR OWN PRODUCT
Go PRO. Stop reading about AI and start shipping it.
The free newsletter tells you what’s worth paying attention to. PRO hands you the tools to act on it:
Templates for client work and your own operations
Agent frameworks you can adapt instead of inventing from scratch
Agent builds, step by step, so you can see how they actually work
Research reports on what’s working and what isn’t
Annual members also get The AI Agent Sprint free: the 22-chapter playbook plus all its resources. You can use it to build an AI consulting practice, or to run your own business: a studio, a gym, a startup.
Monthly: $19 (book not included). Annual: $149 for founding members (then $199), and the book is included. Cancel anytime. A 14-day guarantee starts when you get access.
Why this is here: it's ours. Selling it pays for the testing, and it never changes what The Build says.
INTELLIGENCE
Three things worth your time
SRA · 8 OCT 2026 · LAW · CONSULTATION · Regulator
Solicitors may soon have to show they can check AI
The Solicitors Regulation Authority wants to add technology, including AI, to the Statement of Solicitor Competence. The draft covers checking AI output against primary sources and supervising the tools a team uses. The consultation closes on 3 December.5
Do this: if you supervise anyone who uses AI, read the supervision part, and answer it before 3 December if you have a view.
COURTS AND TRIBUNALS JUDICIARY · 9 OCT 2026 · LAW · Official
The judges' AI rules read like a small-firm policy
The judiciary's new position statement says AI supports judges and does not replace them: "Human oversight is essential." Judges may use only approved tools with UK data residency, enterprise data protection and web access switched off, and AI must not decide the outcome.6
Do this: the same four ideas are the spine of this week's kit. Approved tools only, data kept somewhere you know, a person checks, a person decides.
HMRC · 27 MAY 2026 · TAX · STILL TRUE, EASY TO MISS · Official
No AI agents in Government Gateway
HMRC's rules for its online services say automation tools "must not be used to enter data into and navigate Government Gateway". Trade press reports this covers browser agents and screen scraping, and that HMRC may block access.7
Do this: if anyone is trying a browser agent on client tax work, keep it away from HMRC's online services. Submissions go through MTD-recognised software.
SAFE TO SKIP
Another "AI employees" funding round. Ema raised $77M on 23 September for agents it sells to large enterprises.8 Built for firms with an IT department.
EU AI Act labelling headlines. The 2 December 2026 date for marking AI-made content mostly binds the companies that build AI systems, not firms that use them.9 A headline about a rule that mostly doesn’t bind you; if you publish AI content for EU clients, ask your adviser.
NEW BOOK ⚡ PRE-SALE NOW OPEN
The AI Agent Sprint + The AI Sprint Codex
The professional system for building, delivering, & scaling AI consulting expertise
AI isn’t hard. Translating it into something that works on Monday is the skill, and this book teaches it.
The AI Agent Sprint is a 22-chapter playbook by Niku Dutta, who won his first AI client after 72 hours of panic-Googling. Build an AI consulting practice with it: discovery calls, value-based pricing, agent builds, finding clients, scaling and exit. Or use it to run your own business. Founders, CXOs, small business owners, yoga studio and gym owners can all apply the frameworks and 600+ templates to their own operations.
The AI Sprint Codex lives online. It updates quarterly. Every platform recommendation, rate benchmarks, and regulatory note gets refreshed so the intelligence around the frameworks never go stale. The AI Agent Sprint gives you a system that doesn’t expire. The AI Sprint Codex gives you the system and tools that never falls behind.
Digital editions are $67 at pre-sale launch price. Read Chapter 1 free. A 14-day guarantee starts when you get access.
THE NUMBER
29%
of UK businesses with fewer than 10 staff used AI in late September 2026. For firms with 10 or more staff it was 38%, and for all businesses 30%, up from 9% when the ONS first asked in September 2023.10
For a small firm: seven in ten of the smallest firms report no AI use. Start with the rules, then one task you can check.
Why do some surveys say nearly half? They ask different firms different questions. An insurer's survey of 1,842 UK owners between 30 July and 7 August put it at 47% (vendor data).11 We use the official figure and label the rest.
ADD YOUR FIRM TO THE NUMBER
Does your firm have a written AI policy today?
One click. We'll publish the totals in a future issue, with no names.
YOUR MOVE
The 9pm Test
What task keeps you at your desk at 9pm? Hit reply and tell me in one line. Each month we'll pick one, test it in a real firm and show you the result. Your reply stays between us: we never publish a reader's words, name or firm without a written yes.
RATE THIS ISSUE
How useful was this issue?
One click records your answer. You can add a comment on the next page.
FROM THE DESK
I started this because most small firms using AI have no written rules for it, and I wanted to show the working on writing them.
This newsletter won’t tell you AI is magic, and it won’t give you a tool tour. Each week we test one task in a real small firm and show you where it went wrong (that’s the useful bit).
Ads are labelled, your replies stay yours, and a person reads every line before it goes out. I’m that person.
Thank you for being here at issue one.
Niku Dutta, Founder, Sentient Sky
MSc, MBA
ON THE DESK NEXT
The Build: answering client emails with AI, and the review checklist we used
From your replies: the first 9pm Test, chosen by you
THE RECEIPT
Sources, in footnote order. Vendor data is labelled.
Clio: 2026 Solo and Small Legal Trends Report (vendor data)
The kit's UK and US sources page, checked 19 October 2026
SentientSky test log, 19 October 2026 (summary on request)
Simply Business: 2026 SME Insights (vendor data)
