Lesson objective
Explain how volume, handoffs and errors that nothing announces change the risk of an AI-assisted task once someone other than the owner runs it, and convert a measured correction rate into an expected monthly count.
Plain-language explanation
Why this matters now
Levels 1 and 2 taught one careful person, usually the owner, to run one AI-assisted task well: triage it, prompt it from a verified business file, check the draft against the source, keep the use record and route anything that belongs to a licensed professional to that professional. Level 3 starts at the moment that task leaves the owner's hands.
It happens in every growing business. A first employee starts answering enquiries. A virtual assistant takes over the quote replies. A seasonal helper sends booking confirmations through the summer. The tool can be exactly the same one the owner used. The risk is not the same, because the owner was carrying something the tool never had: the right answer, held in their head.
When an owner reads a drafted quote, they know the oven-cleaning price went up in March, that the Tuesday slot is always taken by the van's service, and that the customer on Elm Row was promised a discount last autumn. They catch the tool's mistakes against that knowledge without noticing they are doing it. A helper who has worked for the business for three weeks has none of it. They can only check a draft against what is written down, and in most small businesses very little is.
There is also no proven productivity gain waiting for the business that scales. Every figure this programme can cite is a laboratory task, a single large firm's rollout, a vendor-affiliated preprint or a self-report, and none describes a small business. The reason to run AI-assisted work as a disciplined system is the consequence of an unchecked output reaching a customer, multiplied by the number of times it runs.
Professional guidance (voluntary framework): NIST's AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, released 26 January 2023) describes itself as "voluntary, rights-preserving, non-sector-specific, and use-case agnostic." It binds no business. NIST's web page states that 1.0 is being revised as part of the White House AI Action Plan; no revised version had been published as of this lesson.
A question to sit with
For two years you have written every quote reply yourself and caught the tool's slips because you know your own prices. Next month a part-time assistant will send forty a week. What will she compare each draft against, and how would you ever find out if she compared it against nothing?
What goes wrong
Volume turns a rate into a count. A method that needs correcting now and then feels safe when it runs a few times a week. Run forty or sixty times a week, the same rate becomes a steady number of wrong quotes, wrong dates and wrong promises every month. Nothing about the method changed except how often it runs.
The helper never saw the right answer. This is the small-business version of the problem, and it is the one that hurts. An error is only visible to someone who knows what correct looks like. The owner knew. The helper was never told, so a draft carrying last year's price looks exactly as right to them as one carrying this year's.
Handoffs drop the exceptions. Every owner-run business runs on exceptions absorbed in someone's head: the customer who is always invoiced monthly, the street where the van cannot park before nine, the regular who gets a courtesy call before a price change. None of these is in the prompt. The owner applied them silently; the helper cannot.
Shared prompts make errors travel together. When the owner hands over "the prompt that works", every reply now comes from one source. A stale price typed into that prompt months ago reaches every customer who asks, and consistent output starts to look like quality when it is really one error repeated.
Nothing announces a confident error. Professional guidance (voluntary framework): NIST's Generative AI Profile (NIST AI 600-1, July 2024) defines confabulation as "the production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications')," and says confabulations "are a natural result of the way generative models are designed." A wrong price in fluent, friendly prose raises no flag. Only a comparison against the source finds it.
The first employee changes the rulebook, not just the workload. Some duties switch on at a stated headcount and some reach a business with one employee or none. Your Level 1 two-column duty list records which is which for your business. The week you hand work to a first employee is the week to reread it, rather than assume the list you wrote as a sole owner still holds.
The practical answer
Separate the risks that handover multiplies from the risks it creates. They need different controls.
| Risk | Present when the owner did it? | What handover does | Control that addresses it |
|---|---|---|---|
| An invented price, date or promise in a draft | Yes | Multiplies it | Checking each draft against a written source |
| Customer data pasted into an unapproved tool | Yes | Multiplies it | The data rules from Levels 1 and 2, written into the procedure |
| The checker has no answer key | No | Creates it | A dated reference sheet the helper checks against |
| Exceptions the owner applied from memory | No | Creates it | An exceptions list, kept with the procedure |
| Several private copies of "the prompt" | No | Creates it | One shared prompt with a version and one named keeper |
| Nobody owns the operation | No | Creates it | A named owner (Lesson 3 of this module) |
Four of the six rows are created by the handover itself. A better prompt fixes none of them.
Do the frequency arithmetic before handover. Take the correction rate you actually measured while doing the work yourself, multiply it by the volume the helper will run, and write the answer as a sentence.
If that sentence is unacceptable, the question is not whether the tool is good. It is which control changes the number: a reference sheet to check against, a narrower task, a heavier check on the replies that quote a price, or the owner keeping the task a little longer.
Write the answer key down. The single most useful thing an owner can do before handing over an AI-assisted task is to write down what the owner was checking against. For a quote reply that means the current price list with its effective date, the service area, the standard terms, the minimum charge and the exceptions list. Module 5 turns those documents into a knowledge base; for now, one dated page is enough.
Ask four questions of any task you plan to hand over. These are an AI Coalition Network recommendation, not a legal or professional requirement.
| Question | Why it matters | If the answer is worrying |
|---|---|---|
| How often will it run? | Frequency converts a rate into a count | State the expected monthly count of wrong outputs |
| Who will run it, and for which customers? | New people lack the owner's memory of prices and exceptions | Write the answer key before they start |
| What will the checker compare each output against? | A check with no source is a read-through | Name the document, with its date |
| What do all the runs share? | Shared inputs cause errors that arrive together | Put the shared prompt under a version and one keeper |
Hand over in stages. Let the helper draft while the owner still checks for a fortnight, then let the helper check while the owner samples. A defect found in week one costs a few corrected replies. The same defect found in month three costs an unknown number of quotes and a search through the sent folder.
Keep the referral lines where they were. Handover does not move them. A question about what a contract means, whether a dismissal is lawful or how an expense is treated for tax still goes to an attorney or the signing tax return preparer, and anything inside a licensed trade still goes to the licence holder. A helper, like the tool, never answers those; the procedure should say so in its first lines.
Remember this
- The owner checked drafts against knowledge in their head; a helper can only check against what is written down.
- Convert your measured correction rate into an expected monthly count and write it down before handover.
- Handover creates risks the owner never had: no answer key, lost exceptions, private copies of the prompt and no named owner.
- A confident wrong answer raises no flag; only comparison against a dated source finds it.
- No figure shows a proven AI gain for a small business; the case for discipline is the consequence of an unchecked output.
Small business example
Copperfield Home Cleaning is a fictional four-person cleaning company. For two years the owner, Dana Whitlock, answered every web enquiry herself: she pasted the customer's request into an AI assistant with a saved prompt, got a draft quote reply and fixed it before sending. Over six weeks she kept a tally and corrected 6 of 150 drafts, a rate of 4%. In May she hires a part-time assistant, Marisol Ferreira, to take the replies over, about 45 a week. Dana emails her the saved prompt and says the tool is usually right. Three weeks later a customer rings to book an oven clean at the price in her quote, $85.00. The price has been $95.00 since March. Dana opens the saved prompt and finds the old price list typed into it the previous year. She had been overriding it from memory on every reply without noticing. Marisol never knew there was anything to override. Dana checks the sent folder: of the 135 replies Marisol sent in three weeks (45 × 3), 22 quoted an oven clean at $85.00, an under-quote of $10.00 each, or $220.00 across the 22 if every customer books (22 × $10.00). Dana decides to honour the quotes already sent. She then makes four changes. She removes all prices from the prompt and writes a one-page reference sheet headed with its effective date. She adds an exceptions list: monthly-invoiced customers, the two streets with morning parking limits, and the regulars who get a call before a price change. She gives the prompt a version number and makes herself its only keeper. And for the next fortnight she reads every reply before it goes, then moves to a daily sample of five. Her handover note now reads: "At 4% and 180 replies a month, expect about 7 wrong quotes a month until the reference sheet proves otherwise."
Responsible-use warning
Do not treat the fact that a method worked for you as proof that it will work for a helper. The owner's own use tests the method under the one condition that disappears at handover: a checker who already knows every price, promise and exception. A shared prompt carrying an old figure, an exceptions list that exists only in your head and a helper told the tool is "usually right" are all invisible while you run the task yourself. Write the answer key before you hand the task over, not after a customer finds the gap.
Human-review requirement
The owner remains accountable for every reply the business sends, whoever drafted it. At handover the owner decides what the helper checks against, writes or approves the reference sheet and the exceptions list, and keeps the shared prompt under one named keeper. The helper owns each check they perform and records what they compared the draft against. Any question about a contract's meaning, an employment decision or a tax treatment leaves the workflow for an attorney or the signing tax return preparer, and anything inside a licensed trade goes to the licence holder, never to the helper or the tool.
Practical exercise
Use a fictional business or anonymised material with no real customer data.
- Choose one AI-assisted task you do yourself today and plan to hand to someone else, such as quote replies, booking confirmations or invoice reminders.
- Record how many times it runs in a normal week now, and how many it will run once the helper has it.
- From your last 30 drafts, count how many you corrected before sending. Calculate the rate, multiply it by the planned monthly volume, and write the result as one sentence.
- List everything you check each draft against. Mark which items exist in writing, with a date, and which live only in your head.
- Write the exceptions list: every customer, place or situation you handle differently from the standard reply.
- List every shared input the helper will use: prompts, templates, saved replies and settings. Beside each, write who will keep it and how a change will be recorded.
- Using the risk table in this lesson, mark each risk in your task as multiplied or created by handover.
- Mark any point where a helper could answer a question that belongs to an attorney, a tax preparer or a licensed trade, and write the referral line that will stop it.
- Write the staged handover plan: how long you will check every draft the helper produces, and when you will move to a daily sample.
Knowledge check
-
Why can the owner catch errors that a new helper will miss, using the same tool and the same prompt?
Answer guidance: The owner checks each draft against knowledge held in their head: current prices, promises made and exceptions. A new helper was never given that answer key, so they can only check against what is written down, and a wrong draft looks as right to them as a correct one.
-
An owner corrected 6 of 150 drafts and the helper will handle 180 a month. What should the handover note say?
Answer guidance: The rate is 6 ÷ 150 = 4%, and 180 × 0.04 = 7.2, so the note should say the business expects about 7 wrong drafts a month and that the check must find them before a customer does.
-
Name three risks that handing a task to a helper creates rather than merely multiplies.
Answer guidance: Any three of: the checker has no written answer key, exceptions the owner applied from memory are lost, several private copies of the prompt drift apart, and nobody is named as owner of the operation.
-
How does NIST AI 600-1 describe confabulation, and why does that matter for a helper checking drafts?
Answer guidance: It defines confabulation as confidently stated but erroneous or false content and says it is a natural result of how generative models are designed. A confident wrong price raises no flag, so only a comparison against a dated source finds it. The profile is voluntary.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.