AI Assistant Basics

An AI assistant is useful when it reduces the effort of a bounded task without taking responsibility away from the person delivering the work. It can transform information, suggest a structure, compare drafts, identify missing fields, and help create a first attempt. It cannot independently know which facts are true, whether a customer authorized a decision, or whether an output is appropriate to publish.

This guide is for a small-service operator who wants practical speed without hiding quality problems. It covers safe use, task design, verification, cost, and customer-facing boundaries. It does not recommend a particular model or assume that a tool has any specific feature.

G01 helps select an opportunity, G02 keeps the test capital-light, G04 tests demand, and G12 will cover larger one-person workflows. This guide establishes the basic operating rules that every later AI-assisted workflow should follow.

Opportunity and Customer

The strongest first AI-assisted service opportunities are not "AI services." They are ordinary business results where draft production is a bottleneck but judgment remains valuable.

Useful categories include:

• Turning messy notes into a structured first draft.
• Classifying requests against fixed rules while flagging exceptions.
• Comparing a document with a checklist and identifying missing fields.
• Summarizing approved source material with links back to the source.
• Preparing several message variants for a human to choose and edit.
• Cleaning inconsistent labels in a data set while preserving an audit trail.
• Converting a completed process into a draft procedure for staff review.

These tasks share a useful property: the output can be checked against something. A customer can inspect the request queue, source document, approved policy, or data set and decide whether the work is correct.

Avoid beginning with autonomous decisions

Do not start with a service that:

• Sends external messages without review.
• Sets prices, approves refunds, or changes financial records.
• Gives legal, tax, accounting, medical, safety, or regulated advice.
• Decides who receives a job, benefit, loan, or other consequential opportunity.
• Publishes claims without checking sources.
• Handles credentials, payment data, health information, or privileged records in unapproved tools.
• Promises that generated work is error-free.

These uses may be technically possible. They are poor first services because errors are difficult to reverse, responsibility is unclear, and the customer may not understand the limits of the tool.

Find the human review point

A saleable workflow names the person who approves the result. For an internal report, the buyer may approve the final version. For a customer message, an authorized employee may approve every send. For a document checklist, a qualified professional may accept or reject flagged items.

The review point is part of the service. If the buyer wants the review removed without adding another control, narrow the work or decline that part of the engagement.

Treat the customer's outcome as the product

A buyer is not paying for a clever prompt, a model name, or an impressive demonstration. The buyer pays for a cleaner queue, a shorter turnaround, a more complete report, or fewer missed exceptions. Price and quality should be described in those terms.

Fit and Skills

AI assistance fits a person who is willing to read, compare, test, and take responsibility. It is a poor fit for someone looking for a machine to remove the need for judgment.

Core capabilities

Build these four capabilities:

• Specification: Define the task, inputs, rules, output format, and exclusions.
• Source control: Know which facts come from the customer, which come from a trusted reference, and which are suggestions.
• Verification: Check critical details against a reliable source.
• Escalation: Stop when the request is ambiguous, high consequence, outside the approved scope, or unsupported by the inputs.

Prompt writing is only one part of the skill. A clear process with a poor source will still produce poor work.

Understand three kinds of output

1. Transformation output: The model reorganizes text or data supplied in the task. Accuracy can often be checked against the input.
2. Suggestion output: The model proposes wording, categories, or next actions. A person must decide whether to accept them.
3. Research output: The model claims facts about the world. Every material claim needs an independent source because a fluent answer can be outdated, incomplete, or invented.

The third category carries the most risk. Never cite a source merely because the model supplied a citation. Open the source, confirm that it exists, read the relevant passage, and check that it supports the claim.

Assess whether the work can be verified

For each task, write:

• Source of truth.
• Required fields or facts.
• What would count as an error.
• Who can detect that error.
• What happens if the error reaches the customer.

If no one can define correctness, AI will accelerate disagreement rather than resolve it. First clarify the standard with the buyer.

Protect your own attention

AI can generate more output than one person can review. That creates a hidden cost. Estimate review time and cap the volume. A workflow that produces fifty pages in two minutes but requires three hours of checking may be slower and riskier than a smaller, better-controlled output.

Startup Budget

Use a modest tool budget and treat every recurring charge as an operating decision. Prices change, currencies differ, and plan limits vary. The following are illustrative ranges, not quotes.

Discovery stage

Illustrative range: 0 to 50 currency units.

• Notes and process maps.
• A free account only if its data terms are appropriate.
• Fictional test data.
• A basic calculator for review time and cost.

Do not transfer customer data into a free consumer product during discovery.

First paid workflow

Illustrative range: 20 to 200 currency units per month.

• One approved AI account.
• Secure storage for approved inputs.
• Password manager and multi-factor authentication.
• A way to keep source references and review notes.
• A reserve for usage limits, corrections, and subscription overlap.

Choose the plan based on controls and the expected output volume, not only the headline price. Check whether an advertised consumer plan excludes the confidentiality or administrative controls the customer requires.

Evaluation of a paid upgrade

Before upgrading, measure:

• Minutes saved per deliverable.
• Errors caught or missed.
• Cost per completed deliverable.
• Maximum output and review volume.
• Whether the higher tier provides needed access, privacy, or administrative controls.
• How the customer will be affected if the subscription expires.

A subscription that saves one hour on work priced above the monthly cost may be justified. A subscription bought before the first customer is a speculative expense.

Include model-independent costs

Budget for the work around the model:

• Writing samples and instructions.
• Source documents and verification.
• Human review.
• Data cleanup.
• Correction and escalation.
• Procedure maintenance.
• Secure account administration.

These are often the largest costs. Automation that ignores them creates an inaccurate margin estimate.

Skill Acquisition

Learn to design a task in a form another person could review. This is more reliable than memorizing prompts.

Use a seven-part task brief

1. Role and objective: What result is needed and for whom?
2. Inputs: Which approved documents or fields may be used?
3. Task: What transformation, comparison, or draft is required?
4. Rules: What must always or never happen?
5. Examples: What does one good and one bad output look like?
6. Output format: What structure, length, and fields are required?
7. Uncertainty: What should the assistant do when information is missing?

An illustrative task brief:

"You are assisting with an internal exception list. Use only the attached request log. Produce a list with request ID, customer, requested date, missing field, and urgency. Use the supplied urgency rules. If the request conflicts with the rules, mark it 'review required' and state the conflict. Do not infer missing dates or customer names. Return only the requested fields, ordered by urgency and then request ID."

This task does not ask the model to contact a customer or decide whether to accept a job. It transforms and flags.

Build a small evaluation set

Create five to ten fictional or sanitized test cases, including:

• A normal case.
• A missing value.
• A contradictory instruction.
• A duplicate record.
• An unusual but valid case.
• A request outside the rules.

Run the task on the set after every material prompt or model change. Count omissions, invented facts, format errors, and missed escalations. A demonstration that works once is not a quality system.

Learn to cite and verify

For research tasks, require a source line for every material claim. The review record should show:

• Claim.
• Source location.
• Person who checked it.
• Date checked.
• Confidence or unresolved question.

Reject unsupported statements from a customer-facing draft. If a source is necessary but unavailable, leave a visible placeholder rather than filling the gap with plausible text.

Practice refusal and escalation

Write short escalation messages:

"The supplied documents contain two different completion dates. I have paused the report and listed both versions with their sources. Please confirm the authoritative date before I continue."

"The request asks for a recommendation that falls outside the approved scope. I can summarize the relevant facts and decision criteria, but a qualified person must make the recommendation."

Practice these phrases before a real exception occurs.

Setup

Set up a small operating environment with clear access, records, and review.

Approve data categories before tools

List the customer data the workflow will touch:

• Public business information.
• Internal process notes.
• Customer contact information.
• Transaction data.
• Employment records.
• Health or safety information.
• Legal or privileged material.

For each category, decide whether the tool, agreement, and customer authorization permit processing. If the answer is not clear, exclude the category until qualified advice is obtained.

Create a secure account baseline

• Use a dedicated business identity where appropriate.
• Enable multi-factor authentication.
• Use unique passwords in a password manager.
• Restrict administrative access.
• Review connected applications and data-sharing settings.
• Record the service provider, plan, owner, renewal date, and purpose.
• Export important records in a usable format.
• Define what happens when the account closes.

Keep a work log

The log should include:

• Customer and engagement.
• Approved inputs.
• Tool and model used, if this matters to the customer.
• Task brief version.
• Output status.
• Verification completed.
• Human approver.
• Exceptions and corrections.
• Delivery and deletion dates.

Do not place unnecessary sensitive content in the log. Use IDs and source paths where possible.

Define publication and communication rules

Write:

• Who may send AI-assisted work externally.
• Which messages require approval.
• Which claims need a source.
• How AI assistance is disclosed when relevant.
• What happens when a model gives a conflicting answer.
• How customer questions about AI use are answered.

Disclosure should match the actual role. A buyer may need to know that records are processed by an external service, especially when the agreement or applicable rules require notice. Do not claim that work is "done by AI" if a person performs the consequential judgment, and do not hide material automated processing behind a person's signature.

Test failure and recovery

Before the first customer:

• Lose access to the tool in a test.
• Export the current output and notes.
• Recreate the result with a manual fallback.
• Confirm the backup location.
• Check that a model outage does not make delivery impossible.

The customer buys the result, not access to one provider.

Concrete Offer

Sell a result with a review checkpoint. Make AI assistance an internal method unless the customer specifically values it.

Offer A: The verified first draft

"I will turn your approved source material into a structured first draft for internal review. The output will include source references, a list of missing information, and a separate exceptions section. A designated person on your team will approve the final version before external use."

This works for reports, procedures, briefs, and routine communications.

Offer B: The exception finder

"I will compare up to [number] records with your fixed checklist and return a list of missing fields, conflicts, and items requiring human judgment. I will not approve the records or make the final decision."

This works where consistency matters and the customer already owns an authoritative system.

Offer C: The assisted delivery block

"Up to [number] hours each month to prepare [specific output] from your approved inputs, including one source-verification pass and a weekly exception log. Work outside the listed fields or volume is quoted separately."

The price should be tied to the completed output and service capacity, not to the number of messages sent to the model.

Outreach

AI is a crowded topic. Outreach should focus on the customer's process and risk rather than the novelty of the tool.

Diagnose the bottleneck

Ask:

• Which repeated task consumes time without requiring a new decision?
• Where does a person currently read, classify, copy, compare, or draft?
• Which errors become expensive?
• What must a person approve?
• Which information cannot leave the approved environment?
• What does a correct output look like?

If the buyer cannot answer these questions, start with a paid workflow diagnostic rather than a demo.

Show a safe sample

Create a sample with fictional data that resembles the customer's document structure. Show:

• Input conditions.
• Output.
• Source references.
• Exceptions.
• Review notes.
• A deliberate example of information that was not invented.

Do not use the buyer's documents in a public model or an unapproved account to create the sample.

Avoid exaggerated claims

Do not promise:

• Errors will reach zero.
• Jobs can be replaced without measurement.
• A model understands the business.
• A generated citation is evidence.
• The customer will keep all sensitive data private without reviewing the provider's terms.

Credible outreach names the review process. Buyers who have handled operational risk usually understand why.

Use a paid diagnostic

A short diagnostic can:

• Map one workflow.
• Identify approved data categories.
• Create an evaluation set.
• Measure manual and assisted delivery time.
• Define the human approval point.
• Recommend whether AI is appropriate at all.

A recommendation not to automate is a valid result. It protects trust and may lead to a different manual or process-improvement engagement.

Pricing and Unit Economics

Price the delivered result, not the length or number of prompts.

Calculate the real assisted cost

Use this sequence for one deliverable:

1. Input preparation and cleanup minutes.
2. Prompt or workflow setup minutes allocated to this output.
3. Generation or processing time.
4. Human verification minutes.
5. Correction and exception minutes.
6. Customer communication and handoff minutes.
7. Tool cost for the activity.
8. Payment, storage, and administration cost.
9. Risk and quality reserve.

Add these to the required contribution. The model call itself may be cheap while review makes the workflow uneconomic at the proposed price.

Illustrative arithmetic

Suppose a report draft has 10 currency units of direct tool and storage cost. Input cleanup takes 25 minutes, generation takes 5 minutes, verification takes 55 minutes, and handoff takes 15 minutes. The total working time is 100 minutes, or about 1.67 hours.

At a fee of 140, cash contribution before tax and general overhead is 130, or 78 per working hour. But if the buyer requires 20 reports each month and the seller can support only 12 with the existing review standard, volume is the binding constraint. Raising volume without increasing review capacity would reduce quality.

This is illustrative and does not imply a market rate.

Price review capacity

Review is the core service, not a hidden extra. A customer who wants a lower price may accept a smaller source set, a shorter output, or fewer exception categories. A customer who wants unlimited source material and no review limit is asking for a different risk profile.

Separate transformation from consequential advice

A task that reorganizes approved facts can be priced as preparation. A task that recommends a legal, tax, financial, medical, employment, or safety decision is a specialized service with separate qualification and responsibility. Do not let a low-cost drafting fee disguise regulated advice.

Track quality and speed together

Review:

• Minutes per output.
• Error rate found in checks.
• Escalation rate.
• Rework after customer review.
• Subscription and usage cost.
• Duplicate or abandoned tools.
• Customer outcome, such as fewer missing fields.

If speed improves while rework and risk increase, the workflow is not ready to scale.

Delivery Example (Illustrative)

The following is an illustrative composite, not a real case study.

A regional training provider receives requests from employers for customized course outlines. The sales coordinator must combine a request form, approved course objectives, public policy information, and internal pricing rules. The current process takes several hours per outline and produces occasional mismatches.

The seller proposes a four-week paid preparation service for the internal outlines. It covers one request form and approved catalog extract at a time. The AI assistant classifies the request, identifies missing information, and drafts sections that copy approved objectives. It does not set prices, promise availability, or write policy claims from general knowledge.

The coordinator reviews every outline against a source checklist. Prices are inserted from the approved price sheet by a person. Any claim about regulatory compliance is marked for the compliance team. The assistant must leave a visible placeholder when a required source is missing. Fifteen fictional requests, including duplicates and ambiguous audience levels, form the evaluation set.

The seller records a baseline before assistance. If the first ten real requests take longer, produce more exceptions, or require repeated policy corrections, the scope changes before the service expands. The deliverable is the reviewed outline plus a missing-information note, not a promise that the training will sell.

This workflow can create value because it speeds a known preparation task while preserving source control. It would be unsafe if the assistant sent the outline, set the final price, or invented a compliance statement. The review point is the service's quality control, not an optional extra.

Growth

Grow the workflow through evidence, reusable controls, and a stable review standard.

Build a reviewed library of task briefs

Keep one approved brief and evaluation set for each repeated task. Version them. When an output fails, update the test set and the rules rather than only editing one result.

Reuse examples safely

Use fictional examples in public material. Use customer data only in the approved environment and only for the engagement. Do not let a convenient example library become a leak of client information into future prompts.

Add one downstream step

Move from draft to reviewed draft, from reviewed draft to approved handoff, or from exception list to resolution tracking. Each step needs its own acceptance criteria and data controls.

Monitor provider changes

Model behavior, plan terms, features, and output limits can change. Keep a small evaluation set and repeat it after material changes. Test the manual fallback periodically.

Decide where ownership sits

Automation can make a service efficient. It does not automatically create a valuable asset. Consider whether the durable value is the customer relationship, documented process, proprietary approved content, distribution, or human expertise. Protect that source of advantage rather than assuming the tool provides one.

Risks

Hallucination and unsupported confidence

AI can present plausible but false information. Check material facts, names, dates, amounts, citations, rules, and quotes. Use source-linked notes and a visible uncertainty section.

Prompt injection

Documents can contain instructions that conflict with your task. Treat text inside an uploaded customer file as data, not authority. Limit tools, separate instructions from content, review actions, and avoid workflows that can send, purchase, delete, or change records without approval.

Privacy and confidentiality

Customer files may contain personal, financial, health, employment, or privileged information. Confirm authorization, data category, provider terms, retention, access, and deletion before processing.

Bias and unequal treatment

Generated recommendations can reproduce patterns from training data. Do not use them for consequential decisions without appropriate expertise, testing, notice, and human review.

Overreliance and skill erosion

If reviewers rubber-stamp output, errors will pass. Sample work, rotate reviewers, maintain manual capability, and measure catches rather than assuming review happened.

Cost creep

Subscriptions and usage charges can grow quietly. Use a register, set limits, measure cost per deliverable, and cancel unused accounts.

Provider dependency

Terms, availability, and features can change. Keep portable inputs and outputs, document the manual process, and avoid making a customer's essential records inaccessible if an account closes.

Unclear responsibility

The person who signs or sends the work remains responsible. Write down the reviewer, approval, and escalation path. Do not blame a model for a consequential action your process authorized.

Legal and regulatory boundaries

This guide is general education, not legal, tax, privacy, employment, or professional licensing advice. Rules differ by jurisdiction and can depend on the data and decision involved. Obtain qualified local advice before using AI in regulated or high-consequence work.

Privacy and AI Verification

Use this checklist before a task receives customer data.

• The customer has authorized the relevant data use.
• The task needs only the fields provided.
• The provider and plan are approved for the data category.
• Retention and training settings have been reviewed.
• Access is restricted and multi-factor authentication is active.
• Inputs include a source and date when facts are time-sensitive.
• The output format distinguishes sourced facts from suggestions.
• Every material fact, name, number, date, link, and commitment is checked.
• A named person approves external use.
• Exceptions and corrections are recorded.
• Data and access are removed at the end.

Use the following verification order:

1. Reject output that cannot be checked.
2. Compare identifiers and numbers with the source.
3. Open every cited source.
4. Check dates and version numbers.
5. Identify claims that require professional judgment.
6. Mark uncertainty and missing evidence.
7. Have the authorized person approve the final result.

30-Day Action Plan

Days 1 to 4: Choose one bounded task

• Select one repeated preparation, classification, comparison, or drafting task.
• Define the customer, result, human approver, and excluded decisions.
• Name the source of truth for each required fact.
• Write what a correct output contains.

Days 5 to 8: Set the data rules

• List the data categories involved.
• Confirm which tool and plan may process them.
• Review access, retention, training, and deletion terms.
• Exclude content you cannot safely process.

Days 9 to 13: Build the task brief

• Write the seven-part brief.
• Create five fictional test cases, including one conflict and one missing value.
• Define the required output structure.
• Practice asking the assistant to flag uncertainty rather than fill gaps.

Days 14 to 17: Test and measure

• Run the evaluation set.
• Record omissions, invented facts, format errors, and missed escalations.
• Measure manual review time.
• Revise the brief and rerun the tests.

Days 18 to 21: Package a small offer

• Write a one-page scope with volume, inputs, outputs, review, and exclusions.
• Calculate delivery time and cost.
• Prepare one fictional sample.
• Offer a paid diagnostic or pilot to five to ten relevant buyers.

Days 22 to 27: Deliver a controlled pilot

• Use only approved inputs.
• Keep a source and review log.
• Stop when the volume or data category exceeds scope.
• Compare assisted delivery with the baseline.

Days 28 to 30: Decide and document

• Review quality, time, cost, and customer acceptance.
• Keep, revise, or stop the workflow based on evidence.
• Document the manual fallback.
• Record one next improvement that addresses the largest measured problem.

Templates

Seven-part task brief

• Objective and audience: [text]
• Approved inputs: [list]
• Required transformation: [text]
• Always rules: [list]
• Never rules: [list]
• Output format: [fields and order]
• Missing or conflicting information: [behavior]
• Human approver: [role]

Evaluation log

• Test case: [name]
• Input condition: [normal, missing, duplicate, conflict, out of scope]
• Expected result: [text]
• Actual result: [text]
• Error category: [omission, invention, format, escalation, other]
• Severity: [low, medium, high]
• Brief change required: [text]
• Retest result: [text]

Source verification record

• Deliverable: [name]
• Claim or field: [text]
• Source: [file, page, record, or approved reference]
• Checked by: [person]
• Date checked: [date]
• Status: [supported, corrected, unresolved]
• Reviewer approval: [name and date]

Customer AI-use question

"Please tell me how this information is handled: which data enters an external service, how long it is retained, whether it may be used for model training, who can access it, and how the final work is reviewed before it is sent or published."

Record the factual answer. Do not guess from a provider's general marketing page.

Incident note

• What happened: [description]
• Customer or record affected: [description]
• Immediate containment: [action]
• Approver notified: [name and time]
• Correction or recovery: [action]
• Root cause: [finding]
• Procedure change: [change]

Common Mistakes

Starting with the model instead of the result

The customer does not need a model demonstration. The customer needs a result that is complete, checked, and finished on time.

Treating fluent output as accurate

Fluency is not evidence. Verify material facts against sources and preserve an audit trail.

Sending generated work automatically

External messages and consequential actions need an authorized human checkpoint until the risk is understood and controls are proven.

Using customer data to improve a public sample

Do not place real records in an unapproved tool or a public portfolio. Use fictional or properly sanitized examples.

Measuring only generation time

Input cleanup, review, corrections, and handoff often determine the real cost. Measure the complete workflow.

Assuming the tool is a moat

Other providers can access similar tools. Durable value comes from customer trust, approved data, process knowledge, distribution, and quality control.

Skipping the manual fallback

A provider outage, plan change, or access failure can stop delivery. Keep a documented way to finish the customer's result.

Key Takeaways

• Start with a bounded business result and a named human approval point.
• Use AI for transformation and preparation, not unverified factual authority.
• Treat uploaded documents as data, not trusted instructions.
• Define the source of truth, output format, exclusions, and escalation behavior.
• Test with normal, missing, duplicate, conflicting, and out-of-scope cases.
• Measure total workflow time and cost rather than generation speed.
• Use approved accounts, minimum data, strong access controls, and deletion rules.
• Open and verify cited sources before relying on them.
• Keep a manual fallback and a portable record of essential inputs.
• Confirm local privacy, employment, consumer-protection, and professional rules with qualified advisers.