Reduce Owner Dependency: Use AI to Capture What Only You Know

Log every question staff bring you for two weeks, then capture the answers behind the most common ones: talk them through on voice notes or let an AI assistant interview you, have AI turn the transcripts into written rules and checklists, and put them in a shared assistant staff ask first.

Expect two to three hours a week of your time for six to eight weeks. AI does the writing up; you still have to do the explaining, and some things (relationships, the final say on big decisions) can't be captured, only handed over. Often, a good share of the questions reaching the owner aren't know-how at all but requests for permission, and a written spending limit fixes those overnight without any AI.

Follow me on Instagram@sagnikteaches

Find out what only you know: the two-week interruption log

Most owners underestimate how often they're the bottleneck, because each interruption is small. So measure it. For two normal weeks, jot down every question or request that reaches you from staff: a notes app, a sheet of paper in your pocket, or a quick voice note all work. Record five things:

Connect on LinkedInSagnik Bhattacharya

The type column is the important one, because each type is fixed differently:

Subscribe on YouTube@codingliquids

• Facts (prices, procedures, where things are) go into written knowledge. This is where AI helps most.
• Judgement calls (whether to take a job, how much to discount, when to escalate) become decision rules with thresholds. AI helps you get them out of your head and into words.
• Permissions (can I spend, order, agree) need delegated limits, not AI. A written spending limit for your senior engineer removes these overnight.
• Relationships (customers or suppliers who only deal with you) need introductions and time. No tool captures trust.

Permissions are often a bigger share than owners expect. In a five-chair hair salon, the owner's log might show that a third of the questions were variations on "can I?": can I give a regular a free fringe trim, can I let a client move her appointment inside 24 hours without the fee, can I order more of the toner that's running low. Three written limits clear nearly all of them: stylists may give away services worth up to $15, once per client a quarter; reception may waive the late-change fee once per client a year; the senior stylist may order stock up to $150 without asking. None of that needs an AI tool, and it's the fastest win in the whole exercise.

Pick the first topic to capture

Don't try to capture everything at once. Group the log's fact and judgement questions by topic (pricing, scheduling, suppliers, customer problems) and score each topic on two things: how many questions it produced, and how costly a wrong answer would be. Start with the topic that's high on both. For most service firms that's pricing, because it's asked about constantly and a bad quote costs real money. A topic that's frequent but harmless, like where the spare keys live, can be written up in ten minutes without any interview at all.

Scored for the plumbing firm in the worked example further down, the log's fact and judgement questions sort out like this:

Let AI interview you

Writing down what you know is slow and dull, which is why owners never do it. Being interviewed is faster, and an AI assistant makes a patient interviewer: it asks follow-up questions, pushes for examples and doesn't get bored. Pick one topic from your log, such as pricing standard jobs, and run a 20 to 30-minute session. Use voice mode in the ChatGPT, Claude or Gemini app if you think better out loud, or type.

You are interviewing me to capture how I make decisions in my business,
so my team can make the same decisions without asking me.

Topic: [e.g. how I price standard plumbing jobs]
Business: [one or two lines about what you do and who your customers are]

Rules for the interview:
- Ask ONE question at a time and wait for my answer.
- Start broad, then dig into specifics: numbers, thresholds, examples.
- When I give a general principle, ask for a real recent example.
- Ask what would make me decide the opposite way (the exceptions).
- Ask what a new member of staff would most likely get wrong.
- Every 8-10 questions, summarise what you've captured so far as
  draft rules and ask me to correct them.
- Don't suggest answers or best practices. Only capture mine.

When I say "finish", write up everything as:
1. Decision rules (When..., do..., unless...)
2. Numbers and thresholds
3. Worked examples I mentioned
4. Open questions I didn't fully answer

A few minutes into a session, the exchange tends to go like this. Take the owner of a small joinery workshop, being interviewed about when to accept rush orders (illustrative):

AI: You said you usually say no to rush jobs. Can you give me a recent one you said yes to?

Owner: Last month, a set of fitted shelves in two weeks instead of six. But that was a builder who sends us work every quarter.

AI: So repeat trade customers can get a rush slot. Is there a limit on size, or on how many at once?

Owner: Nothing over about three days' bench time, and never two rushes in the same fortnight, because it wrecks the schedule for everyone else.

AI: Would you charge more for it?

Owner: For private customers, yes, 20% on top. Not for the builder. He'd go elsewhere.

Four exchanges produced a rule the owner had applied for years and never said out loud: rush slots for repeat trade customers only, up to three days' bench time, one per fortnight, with a 20% premium for private customers who get an exception. The follow-up about a recent yes was what surfaced it. When you read the summary the AI produces every ten questions, watch for anything you didn't say. If a line such as "prioritise customer satisfaction on urgent requests" appears, it's the model's filler, not your rule, and should be struck out.

If you think best away from a desk, record voice notes when you're parked between jobs and transcribe them. Many phones can now transcribe recordings, and dedicated tools such as Otter.ai Pro cost $8.33 per user a month billed annually ($16.99 billed monthly). Paste the transcript into the assistant and ask it to draft rules in the format below, listing anything unclear as a question for you.

The line "Don't suggest answers" matters. Left alone, AI assistants tend to fill gaps with generic advice, which then looks like your rule when it isn't. Three short sessions on one topic usually capture far more than one long one, because you remember things between sessions.

For work you do on a screen, such as how you set up a job in your software or check a supplier statement, record your screen while talking it through, then have AI turn it into steps. Creating SOPs from screen recordings with AI covers the method.

Turn the transcripts into decision rules staff can apply

A rule is only useful if a member of staff can apply it on a Tuesday afternoon without calling you. Use one format for every rule:

RULE [number]: [short name]
When:     [the situation]
Do:       [the action, with numbers]
Unless:   [the exceptions]
If unsure: [who to ask, or what to do by default]
Example:  [a real case, with details changed]
Owner's reason: [one line on why, so people can handle new cases]

Paste the interview write-up into the assistant and ask it to convert everything into this format, one rule per decision, flagging any rule with no number in it. Then edit them yourself.

The raw material is usually messier than people expect. This is the kind of voice note the plumbing firm's owner recorded from the van, word for word from the transcript:

"Yeah, so the cash thing. If they want money off for paying on the day, fine, five per cent, but not on the little jobs, under four hundred it's not worth it. And never the landlords, they're on account anyway, they get invoiced. Oh, and if we've messed up, like the job ran over because we brought the wrong part, then don't agree anything, ring the office first."

Nothing in that is wrong, but a new engineer couldn't find the rule in it quickly. Converted into the fixed format, and after one correction from the owner (the AI's first draft made the discount 5% on every job, having missed "under four hundred"), it becomes:

RULE 7: Cash-on-the-day discounts
When:     A customer asks for a discount for paying on completion.
Do:       Offer up to 5% on jobs over $400. No discount under $400.
Unless:   It's a landlord or letting agent account (they're invoiced,
          never discounted), or the job overran because of our error
          (then call the office before agreeing anything).
If unsure: Say you'll confirm by text within the hour; call the office.
Example:  Boiler service plus valve swap, $460. Customer offered to pay
          by card on the day; engineer offered 5%, $23 off.
Owner's reason: Getting paid the same day saves chasing time. Under $400
          the chasing saving is smaller than the discount.

"Owner's reason" is the line that turns a rule into judgement. Staff who know why can handle the case the rule didn't anticipate. Aim for 15 to 30 rules per topic; if you're past 50, split the topic. If you'd rather start from procedures than decisions, writing SOPs with AI from rough notes is the companion method.

Put it where staff will ask it first

Rules in a folder get read once. Rules inside an assistant get used daily, because asking a question is easier than searching a document. The simplest set-ups for a small team:

• A project in ChatGPT or Claude with your rule documents uploaded as files and a short set of instructions. Projects are available on every ChatGPT plan, and Claude's free plan allows up to five. For a team, use a business plan, which lets you share the project and excludes your content from model training by default.
• A Gemini Gem (becoming a skill from November 2026), which lets you add instructions and knowledge files, and is available on every Gemini plan including the free one; it also appears in the Gemini side panel in Google Workspace.

One caution: OpenAI is retiring custom GPTs on every ChatGPT plan. New ones can no longer be created, existing ones stop running on 11 December 2026, and OpenAI's migration turns them into plugins. If an older guide tells you to build a GPT for this, use a project instead. Give the assistant instructions along these lines:

You answer questions from staff at [business] using ONLY the attached
rule and knowledge files. For every answer:
- Quote the rule number you're relying on.
- If the files don't cover the question, say "Not covered: ask [name]"
  and don't guess.
- If two rules conflict, say so and give both.
- Never make up prices, times or policies.

For how to structure the underlying documents so answers stay accurate as the collection grows, see how to build a knowledge base your AI assistant can rely on.

"Ask the assistant first", and the gap log

The habit that makes this work is simple: staff ask the assistant before they ask you. When the answer is missing or wrong, they add a line to a gap log (a shared document is fine): the question, what the assistant said, and what you told them. Once a week, spend 20 minutes turning gap-log entries into new or corrected rules and uploading the updated files.

For the first month, expect the gap log to be long. That's the system working: every entry is a question that will never reach you again.

Worked example: a plumbing firm owner who priced every job

An illustration. Say the owner of an eight-person plumbing and heating firm is asked about 47 questions a week. The two-week log shows 94 in total: 41 facts, 28 pricing judgements, 15 permissions and 10 relationship matters. The owner hasn't taken a full week off in three years.

• Weeks 1 and 2: the interruption log.
• Week 3: permissions fixed without AI: a $500 spending limit for the senior engineer, and the office manager can approve any refund under $150.
• Weeks 3 to 5: three AI interview sessions on pricing, recorded with voice mode, produce 22 pricing rules and a price band for each standard job. The office manager starts quoting standard jobs within those rules.
• Weeks 4 and 5: the 41 fact questions become four knowledge files: price list, supplier accounts, van stock, common customer questions.
• Week 6: a shared Claude Team project goes live with the rules and files. Staff ask it first; the gap log collects 19 entries in the first fortnight.
• Weeks 6 to 8: the owner introduces the office manager to the five commercial clients who only ever called the owner, and copies them in on routine emails.

By week 8, questions reaching the owner are down to about 18 a week, mostly genuinely new situations and the remaining relationships. Owner time spent: roughly 20 hours over eight weeks. The following month the owner takes a full week off, and the gap log records four questions, none urgent.

What AI can't capture for you

• Trust. A client who has dealt with you for ten years won't transfer to a chatbot. Transfer relationships person to person, gradually. AI can help with the first step, though: ask it to draft a short introduction email for you to edit and send, along the lines of "From next month [office manager] will be your first contact for bookings and quotes. She's been with us six years and knows your properties. I'm still here for anything contractual, and I'll join your next review meeting with her."
• Accountability. Anything that legally needs a qualified person's sign-off stays with a qualified person. The rules can describe the process; they can't sign the certificate.
• Genuinely new situations. Rules cover the repeat cases. The unusual ones still need someone with judgement, which is why "Owner's reason" matters: it's how you grow that judgement in others.
• Your authority. Staff need to know they're allowed to decide. Say it out loud, and don't overturn a reasonable decision made within the rules, even if you'd have done it differently.

How to tell it's working

1. Interruptions per week, re-counted for one week each month. The plumbing example went from about 47 to 18.
2. Assistant accuracy: once a week, pick ten recent questions and check the answers. Aim for nine correct, with the tenth correctly saying "not covered". In the plumbing firm's third week, the check found eight right and two wrong, and the two failures were typical. One quoted last year's call-out charge, because the old price list had been uploaded alongside the new one; deleting it and putting the date in each file name fixed that. The other answered "how big a deposit should we take on a bathroom refit?" with a confident "typically 25 to 50%", although no rule covers deposits. General-sounding questions tempt the assistant to answer from its own knowledge, so the owner wrote a deposit rule and added "questions about money not covered by a rule are always Not covered" to the instructions.
3. Decisions made without you, such as quotes sent within the rules without your review.
4. The week-away test. The real proof is a week off with your phone switched off, and the business still running.

This work also changes what the business is worth to anyone else. A firm where the know-how lives in documented rules is easier to value, hand over or sell than one that lives in the owner's head; getting your records ready to sell your business covers that side. And the rules make every other AI tool better, because giving AI your business context gets much easier once that context is written down.

Questions owners ask about capturing their know-how

Will staff actually ask an AI assistant instead of me?

Only if it answers well and you back the habit. Load it with rules that cover the common questions first, then when someone asks you something it covers, send them to it and fix it if the answer was wrong. Within a few weeks the easy questions stop reaching you. If the assistant keeps giving poor answers, the rules behind it are thin, not the staff.

Is it safe to put my pricing rules into ChatGPT or Claude?

Use a business plan that doesn't train on your content by default, such as ChatGPT Business or Claude Team, and keep customer personal details out of the rule files; rules about how you price don't need names. On a personal plan, switch off the model-training setting in privacy settings first. Treat your pricing logic as confidential and limit who can open the project.

What if my know-how is mostly instinct?

Instinct is usually a set of rules you've never said out loud. Talking through real recent examples brings them out: why you took this job and turned down that one, why this quote was higher. The AI interview prompt is built for that, asking for cases and exceptions rather than general principles. Expect the first session to feel awkward and the third to be productive.

Further reads

• How to Document Your Processes Before Adding AI — The wider process documentation this builds on.
• How to Redesign Job Roles Once AI Handles Routine Tasks — Reshape roles once staff take on your decisions.
• How to Train ChatGPT on Your Business Information Safely — Loading business information into ChatGPT safely.
• AI Knowledge Base Options for Small Businesses Compared — Options if the shared assistant outgrows a project.
• How to Choose and Support an AI Champion in a Small Team — Someone other than you to keep the rules current.
• How to Onboard New Hires Onto Your AI Tools and Rules — Use the captured rules to train new starters faster.
• Who Should Own AI in a Small Business? Roles and Responsibilities — The four roles AI needs in any small firm, how they're split in a barber shop, an optician and a garden centre, and a responsibilities chart to copy.
• What Business Data Should You Start Collecting Now for AI? — Seven datasets worth capturing from today (enquiries, quotes, job actuals, questions, complaints, prices, feedback), with the fields that make them usable.
• How to Use AI on Your Own Admin First, Then Roll It Out — A six-week owner-first plan: pick three of your own admin jobs, log what AI saves and gets wrong, write a playbook, then hand one task to one person.
• Automating a Broken Process: Why It Backfires and What to Fix — Why automating a broken process backfires, the five-pass fix to do first, and how to split the cleaned-up process between rules, AI and people.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center (Projects in ChatGPT; custom GPT retirement FAQ); Claude Help Center (projects); Gemini Apps Help (Gems); Otter.ai pricing page. Checked September 2026. The worked example is illustrative.