Chapter 3. Choosing product forms and features

Learning objectives

• Explain chat, projects (knowledge spaces), custom assistants, connectors, agent features, deep research, scheduled tasks and APIs through analogies.
• Choose a suitable form based on duration, sharing scope, external material, actions and the need for integration into a product.
• Distinguish the scope of the current conversation, saved instructions, memory and project instructions.
• Explain inherited connector permissions, the difference between read and write access, and why agents need approval points.
• Judge forms by their properties even when product names change, and compare them using real work rather than feature charts.

Introduction

Nam, a manager at an accounting office, noticed something odd about how employees used AI. One pasted the same work-rules document into a new conversation every morning. Another downloaded client material each week and uploaded it again. A third manually visited ten websites to collect information, asking AI only to summarize the results. All three used a single chat interface for everything.

Their tasks had different structures. The first repeated work under the same rules; the second needed current material from outside systems; the third required actions across several steps. AI products offer different forms for these types of workflows. Choosing poorly can mean repeating the same setup every day, receiving answers based on old material or using unnecessarily heavy tools.

This chapter begins with selection criteria, then examines chat, memory, projects, custom assistants, connectors, agents, deep research, scheduled tasks and APIs. "Going deeper" explains how to distinguish fixed procedures from open-ended problems, and how to decide whether answering from company documents requires retraining a model. This chapter corresponds to Delegation in AI fluency.

Key concepts

Five questions for choosing a form

AI product names change often, but their forms fall into a few categories. Ask five questions when choosing.

1. How long will the work continue? Is it a one-off, or will the same material and rules be used repeatedly over several weeks?
2. Who will use it together? One person, a team or the whole organization?
3. Is outside material needed? Must it access current files in internal drives, email or messengers, or research the web broadly?
4. Will it take actions? Is an answer enough, or must it navigate websites, fill forms or send email? Must it run at set times without a person present?
5. Must it be built into our product or system? Is human use through an interface enough, or must software make high-volume automatic calls?

The table is only a starting point; forms are not mutually exclusive. Combine them when work spans several questions. For a proposal, for example, keep writing rules in a project, retrieve current client material through a connector and receive weekly progress summaries through a scheduled task. Put approval and review points wherever the combination takes action.

Chat: a single telephone consultation

Chat is the most basic question-and-answer form. Like a single telephone consultation, it suits work that begins and ends in the moment: brainstorming, short drafts, sentence editing and one-off questions. Most chat products allow tools such as file attachments, web search and image understanding within a conversation.

Chat's limitation is that context generally remains within that conversation. Memory may carry some information forward, but important rules and material are not guaranteed to transfer automatically. Using chat alone for recurring work means repeatedly pasting the same material; differences between people's versions lead to inconsistent results.

Several products offer temporary chat for conversations you do not want in history. OpenAI states that temporary chats do not appear in history, create or update memory, or contribute to model improvement, but may be retained for up to 30 days for safety purposes. Saving a temporary chat turns it into a regular conversation governed by account settings.

Current conversation, saved instructions, memory and project instructions

The expectation "I told it last time, so it should know" often fails. What the model remembers depends on what the product supplies as context. Distinguish four layers.

OpenAI distinguishes custom instructions, which users provide directly, from memory, which the product selects and accumulates across conversations and sources. Deleting a conversation may leave separately saved memory behind, so check both. Project instructions apply only within their project and take precedence over custom instructions.

For important work, explicitly state necessary conditions in instructions or project instructions instead of relying on memory.

Projects (knowledge spaces): a shared team drawer

A project is a workspace bringing related conversations, files and instructions together. Names differ by product; this book also calls it a knowledge space. OpenAI recommends reusing projects for recurring work such as weekly research and content drafting. Anthropic's Claude projects can likewise hold a knowledge base of uploaded documents and their own instructions.

Suitable work has three properties: it repeatedly uses the same material, follows the same rules and continues for several weeks or more.

Consider boundaries when sharing.

• Visibility: OpenAI states that all members of a shared project can view and download uploaded files. Material intended for only some members should never be placed in the shared space.
• Editing rights: Anthropic describes view and edit sharing permissions on Team and Enterprise plans. Decide who can modify instructions and material.
• Memory scope: OpenAI explains that shared projects automatically use project-only memory and do not access members' personal memories or context outside the project.

As material grows, products may retrieve only relevant passages instead of including all documents in context at once. Anthropic states that paid plans switch to a retrieval approach, RAG mode, as project knowledge approaches context limits, increasing capacity. RAG (retrieval-augmented generation) retrieves document fragments relevant to a question and supplies them to the model. Since it reads only retrieved portions, material it fails to retrieve may be absent from the answer. Project quality therefore depends even more on the uploaded material than on the model. Remove outdated and retired versions, assign update owners and put applicability and effective dates at the start of documents. For frequently changing material, connect the original through a connector rather than uploading copies.

Custom assistants: a dedicated employee who knows the manual

A custom assistant packages instructions, reference material and sometimes tools so several people can receive consistent help. Examples include employee-benefits guides and brand-style editing assistants.

OpenAI contrasts these with projects as follows. Custom assistants suit distributing stable, expert-curated knowledge across an organization; shared projects suit collaboration in which members build up material and conversations together. One distributes a finished manual; the other is a shared workshop.

Do not stop at a role description when creating an assistant. Design four parts separately: recurring instructions, reference material, available tools and work it will not do, such as referring case-specific legal judgments to the responsible officer. Uploading material does not ensure perfect understanding of every detail. Test representative questions before launch.

Connectors: entering only with your own access badge

A connector links AI to external services such as drives, email, messengers and calendars so it can retrieve information during conversation and sometimes take actions. Products may call them apps or connected apps. Many connectors use MCP (Model Context Protocol), a public standard. The official MCP documentation presents it as an open-source standard connecting AI applications to external systems, comparing it to USB-C, which connects devices in a standard way. Standardization allows one connection to be used across several AI products.

The most important property is permission inheritance. Anthropic explains that connectors inherit individual users' permissions: files or records inaccessible to a user in the source system remain inaccessible through the connector. OpenAI likewise states that connected apps in business workspaces respect an organization's existing permissions. Some apps require each user to connect an account; others use administrator-configured connections. Connectors do not create new permissions. They are like an employee badge that admits you only to authorized floors.

Two qualifications matter.

• Whose badge? Anthropic explains that a connector using shared credentials can access everything those credentials allow. A broadly privileged shared account undermines distinctions between individual users' access.
• Reading and writing are different permissions. Searching and reading documents poses different risks from sending email, changing schedules or deleting files. Anthropic states that Team and Enterprise administrators can set each connector tool to always allow, require approval or block. These restrictions do not grant broader permissions than the source system permits.

When current material is needed, connecting an authorized original is more accurate and safer than downloading and uploading weekly copies. As Anthropic advises, connect only trusted services needed for work.

Agent features: an assistant sent on an errand

Agent features let AI accept a goal, plan multiple steps independently, move between websites and apps, gather information and take actions such as filling forms or creating documents. OpenAI describes ChatGPT's Work, alongside Chat for quick conversations, as an agent for longer, multi-step work and finished deliverables. It explains that users can observe progress, redirect work and approve important actions.

Set three things when delegating to an agent, as when sending someone on an errand.

1. Goal and endpoint: What counts as finished—a comparison table, a draft, or stopping just before submitting a booking request?
2. Allowed scope: Which sites and apps may be used, and which actions are prohibited, such as purchases, registration or external sending?
3. Approval points: Where must it stop before hard-to-reverse actions and obtain human confirmation?

Deep research: a report comparing multiple sources

Deep research takes a complex question, develops a research plan, reads multiple sources and produces a report with citations. OpenAI explains that ChatGPT's deep research shows a plan users can revise, allows changes to direction and sources during execution, and returns structured reports with citations or source links. It recommends deep research for questions requiring synthesis across multiple sources, while ordinary chat may be faster for quick checks and short exchanges.

Deep research also proceeds independently through several steps, but it focuses on reading, unlike agents that make external changes. OpenAI states that deep research uses connected apps' read capabilities, not write capabilities, and app connections do not exceed original permissions. Many citations do not complete verification. Open original sources for a few key claims.

Scheduled tasks: recurring work with an alarm

Scheduled tasks repeat work at set times or under specified conditions. OpenAI states that ChatGPT can run one-off or recurring tasks, monitor changes and respond to supported app events such as new email or messenger messages. It also notes that tasks may pause for user review when reaching an action requiring approval.

A common misunderstanding is that running at the scheduled time and making correct judgments are different things. A summary may arrive on time and still need checking. Define execution time, the material's reference date, who checks abnormal results and when the task ends. Grant narrow permissions because it runs without a person present. OpenAI advises against sensitive personal information in task titles and instructions because anyone with a shared task link can read the full instructions.

API: components you assemble yourself

An API (Application Programming Interface) is a route through which software calls an AI model directly. Instead of a person using an interface, a company app or internal system sends requests and uses returned results. It suits embedding summaries in your own app, automatically classifying tens of thousands of inquiries daily or requiring fine-grained control.

Non-developer managers should understand four things in API adoption discussions.

1. Billing is separate. OpenAI states that API usage follows API pricing rather than the usage included in chat subscriptions. Buying a subscription does not make API use free (Chapter 5).
2. Non-urgent bulk work can cost less. OpenAI states that its Batch API costs 50% less than synchronous calls and completes within 24 hours for work not needing immediate responses. Anthropic also offers a 50% discount on batch requests.
3. Development and operations require work. Error handling, cost monitoring and rechecking after model changes are needed.
4. Accuracy and security do not arise automatically. Verification processes and data policies remain necessary.

Starting with an API for one or two people's recurring tasks is often unnecessarily heavy. If people can handle the volume through an interface, try a project or custom assistant first.

Going deeper

Workflows and agents: fixed paths and open paths

Anthropic distinguishes two AI system types. Workflows move models and tools along predefined paths; agents let models determine their own processes and tool use. A train follows tracks, while a taxi chooses a route after receiving only the destination.

Anthropic's advice is clear: find the simplest possible solution and increase complexity only when needed. Workflows suit well-defined work requiring predictability; agents suit open-ended problems where the number of steps is unknown. Agents spend time and money for potentially better results, and greater autonomy brings increased cost and risk of errors accumulating at every step. Anthropic therefore recommends sufficient testing in isolated environments, human checkpoints and stopping conditions such as iteration limits.

For work following the same sequence every month, fixed-step automation is cheaper and more reliable than adopting an agent.

Does answering from internal documents require retraining?

A common proposal is, "To make AI understand our company policies, we must retrain it on them." Training a model further on additional material is called fine-tuning. OpenAI describes two dimensions of accuracy improvement. If the problem is knowledge the model never learned, outdated knowledge or company-specific information, optimize context: retrieve necessary material through RAG (retrieval-augmented generation). If the problem is inconsistent behavior, such as unstable formats or unsuitable tone and style, consider fine-tuning. The starting point is always prompt improvement, with its evaluation set serving as the baseline.

Internal policy Q&A is usually the first case. Policies change and answers need supporting clauses. Retrained models need retraining after revisions and make it difficult to show which clause supports an answer. By contrast, supplying current policies to projects or custom assistants, or connecting originals through connectors, requires only document updates and allows citations. Before choosing an expensive approach, ask whether the problem is knowledge or behavior.

At work

Cozy Living's launch team: designing two months of work

Lee, a team leader at home-accessories company Cozy Living, is responsible for a new product launch. Three marketing and two product-planning staff will work together for two months. Lee organizes AI use through the five questions.

First, launch preparation lasts two months, and all five repeatedly use the same specifications and brand guide. Lee creates a shared project and adds core brand rules to project instructions. A suggestion to put the brand guide in everyone's custom instructions is rejected: those instructions affect other personal conversations, vary between members and are not used as members' personal custom instructions within the shared project. An unfinalized cost sheet is excluded because only the two product planners should see it. Every shared-project member can open and download uploaded files.

Second, inventory and sales figures change daily in an internal-drive spreadsheet. Instead of repeatedly downloading and uploading it, each person connects a drive connector through their own account to retrieve current files during conversations. Marketing lacks permission for the cost folder, so the connector cannot show it either. Lee asks the administrator to allow file search and reading but require approval for spreadsheet-editing tools, preventing AI from changing original figures while analyzing them.

Third, agent features compare competitors' new-product prices across shopping sites. Scope and endpoint are explicit: "Do not purchase or register. Stop after creating the comparison table."

Fourth, a scheduled task produces industry-news summaries every Monday morning. Lee checks the first few weeks personally before sharing results with the whole team.

No API is used. This is not integration into a customer app, and five people can handle the volume through the interface. The criterion is the feature matching the work's shape, rather than the most powerful feature.

Starbridge Events: supplier research

Cho at event agency Starbridge Events must find an audio-equipment rental supplier for next month's corporate workshop. Three conditions apply: it must be in the Seoul metropolitan area, have at least 20 wireless microphones, and send a technician on the event day. The team leader wants a comparison of five candidates and draft quotation-request emails to three by Friday.

Cho divides the work into three parts.

1. Finding and comparing candidates: Multiple supplier websites and reviews need comparison, with traceable sources for the team leader. Ordinary chat search seems insufficient because it provides only a few sources, so Cho chooses deep research. The proposed plan omits the metropolitan-area and technician requirements; Cho corrects it before starting.
2. Checking equipment: Each website locates its equipment list differently, requiring navigation. Cho delegates this to agent features, prohibiting inquiry-form submission and registration.
3. Quotation-request emails: Letting the agent send them would be quick, but external email represents commitments in the company's name. With the mail connector's send tool requiring approval, Cho requests drafts only, then reads, edits and sends them personally.

One supplier is reported as offering technicians, but its citation is an event review from three years earlier. Cho marks it "Needs checking" and adds the question to the quotation request. Delegating everything could extend dangerously into external sending; doing everything manually would miss Friday's deadline. The crucial boundary is delegating reading and comparison while requiring human approval for external actions.

Common mistakes

1. "More powerful features are better." An agent for a one-off question or an API for two people's recurring work can simply add cost and risk. Choose the lightest form matching the task.
2. "Enabling a connector exposes all company material." Connectors follow the connected user's permissions. Conversely, sharing one broadly privileged account opens everything it can access. Avoiding connectors as inherently dangerous can create the greater problem of circulating outdated copies.
3. "Uploading to a project trains the model." Uploads supply reference context for that project. Uploading and training are different; plans and settings determine training use (Chapter 4).
4. "Answers are accurate in spaces containing documents." Models can still err or mix in general knowledge, and retrieval can miss portions. Verification remains necessary.
5. "Memory means I need not repeat conditions." Memory is only selected information, and its contents may be unclear. State important conditions in instructions or project instructions.
6. "Scheduling ensures correct handling." Scheduling controls execution time, not judgment accuracy. Assign result reviewers and an end date.
7. "Delegating everything to an agent is fastest." Hard-to-reverse payments or sending without approval increase risk rather than simply speed.
8. "Company-document answers require retraining." For a knowledge problem, first retrieve and supply material. Consider fine-tuning for inconsistent behavior.

Summary at a glance

Check the official docs

• OpenAI — Projects in ChatGPT: Project-instruction scope, sharing and file visibility, projects versus GPTs.
• OpenAI — Custom GPT retirement and migration FAQ: Custom GPT retirement schedule and migration to Plugins.
• Google — About the transition from Gems to skills: Gems-to-skills transition schedule.
• Anthropic — What are projects?: Knowledge bases, instructions, sharing permissions and retrieval mode.
• Anthropic — Use connectors to extend Claude's capabilities: Inherited permissions, shared-credential risks, per-tool allow/approve/block controls.
• OpenAI — Deep research in ChatGPT / Scheduled tasks in ChatGPT: Deep research, scheduled tasks and waiting for approval.
• Anthropic — Building effective agents: Workflows and agents, starting with the simplest solution.

Self-check

3-1. List the five questions used to choose an AI product form.

3-2. Distinguish a shared project and a custom assistant in one line each, using whether knowledge is stable or built up together.

3-3. Classify each proposed agent action as "May be delegated," "Check before execution" or "Human approval required," and explain why: reading an internal wiki, making a provisional meeting-room reservation, sending discount coupons to 200 customers, and confirming a schedule shared with a partner.

3-4. A mail connector's tools can be set to always allow, require approval or block. Where would you place mail search, drafting replies, sending mail and permanently deleting mail? Explain why.

3-5. Make a table distinguishing custom instructions, memory and project instructions using two columns: who sets them and where they apply.

3-6. Give one situation where deep research is better than ordinary chat and one where it is less suitable, explaining why.

3-7. Choose one recurring team task and answer the five questions. Based on those answers, design the form or combination of forms and approval points.

Hands-on tasks