The 4D Framework: How to Decide What AI Can (and Cannot) Do in Your Business

Updated October 18, 2026: We clarified what the law says about telling customers they are talking to AI, corrected the explanation of how AI tools behave and learn, and added sources, including a link to the original AI Fluency Framework.

First D: Delegation

Delegating to AI does not mean abandoning a task. It means assigning the right work to whoever handles it best — and sometimes that is an AI tool; other times it is you or someone on your team.

A practical way to sort your business tasks is into three groups:

Group 1 — AI can handle this on its own, once tested. Repetitive, standardized, well-documented tasks where a mistake is low-stakes and easily corrected. Example: answering the same ten questions your business gets every week — store hours, list prices, return policy, how your process works. AI only gets here after testing (Discernment) and a period in which you review everything it produces (Diligence). After that, the goal is that customers get an instant answer and you stop writing the same message twenty times a day.

Group 2 — AI drafts; a human decides. Tasks where AI accelerates the work but the final judgment stays with you. The AI drafts a service proposal; you review it before sending. The AI summarizes call notes; you confirm the summary is accurate before filing it.

Group 3 — Human only. Complaints and conflict resolution, emotionally sensitive conversations, decisions that are hard to reverse, anything where being wrong carries serious legal or financial consequences. AI has no role as the primary actor here — though it can help you prepare (gathering background, organizing context) before you step in.

The question that determines which group a task belongs to: what happens if AI gets it wrong? If the consequence is trivial and easily fixed, Group 1. If it needs review before reaching the customer, Group 2. If an error could damage an important relationship or have irreversible consequences, Group 3.

See also how to decompose a workflow for AI automation — the process of mapping your tasks makes Group 1 candidates much easier to spot.

Second D: Description

An AI tool works from what you tell it, and it fills any gaps with guesses. Give it vague instructions and you get vague results; give it precise instructions and results become much more consistent, though not guaranteed to be identical: models are non-deterministic by default, so outputs may differ from request to request (OpenAI API docs). You still review them.

The most practical way to think about this: imagine you are training a very smart new employee who just walked into your business and knows nothing about your context. What would they need to know to handle this task well, without making the mistakes that matter most to you?

A useful description has three parts, adapted from the framework's product, process and performance description:

The product: what exactly does the AI produce? A WhatsApp message no longer than three sentences? A formal email with greeting, body, and sign-off? A four-bullet summary? Be specific about format and length.

The process: how should the AI reason its way to that result? What questions should it ask itself, in what order? When should it tell the customer this case needs a human? Define the decision points explicitly.

The behavior rules: how it must behave and what it must never do. This is the part that is easiest to skip, and where problems tend to appear. Examples: "never confirm a delivery date without checking inventory," "never carry one client's information over into another client's conversation" (and don't put other clients' data in its context in the first place), "never invent a price — if it is not in your context, say you will check." Behavior rules protect your reputation.

Third D: Discernment

No description, however detailed, survives first contact with real examples unchanged. There are always cases you did not anticipate, language nuances the AI reads differently than you would, or information the tool simply does not have and decides to make up.

Discernment means evaluating AI outputs for accuracy and usefulness. In practice, that is not a one-time review — it is an iterative process using your own examples.

Before any AI tool touches a real customer, test it with real cases from your business history; as a rule of thumb, start with at least five. Actual conversations that already happened. Questions that already came in. Situations you already handled. Take out names, contact details and anything sensitive before you paste those cases into the tool. Review each output as if it were about to reach a customer.

Pay close attention to invented answers: does the AI answer confidently questions whose correct answer is not in its context? That is one of the most damaging patterns for customer trust. The best defense starts in the Description step: tell the AI explicitly what it must not invent and give it permission to say "I don't have that information" instead of guessing — Anthropic says this "can drastically reduce false information," but such techniques reduce invented answers without eliminating them (Anthropic, Reduce hallucinations), so keep reviewing. Catching the pattern requires testing in situations where the right answer is "I don't know."

Expect two or three rounds of adjustment before the behavior is consistent. That is normal. The AI does not learn your business from these tests; what improves is your description, which you refine between rounds. (Memory features, where available, save notes about you that you can view and edit, and the tool reads them in later chats; see, for example, Claude's memory help page.)

Fourth D: Diligence

Putting an AI tool to work is not a one-time event — it is an ongoing supervisory practice, at least until the tool has earned your trust through enough real-world history.

Three concrete practices that make the difference:

Review before it reaches the customer. At first, review every output the AI produces before it goes out. Over time, once history gives you confidence, you can move to sampling — reviewing a portion of outputs rather than all of them. But the starting point is full review.

Be transparent with your customers. State clearly when a response comes from an AI assistant, and always leave a clear path to a human. Depending on where you operate and what the AI does, the law may require it — see the FAQ below for examples as of September 2026 — but disclose anyway. Disclosure can have a cost, and being exposed by someone else can cost more trust: in a field experiment with more than 6,200 customers, disclosing a sales chatbot before the call cut purchase rates by more than 79.7% (Luo et al., Marketing Science, 2019), and across 13 experiments in work settings, people trusted those who disclosed AI use less, but trusted them even less when someone else exposed it (Schilke & Reimann, Organizational Behavior and Human Decision Processes, 2025). Say it up front, make the AI genuinely useful, and keep a person one step away.

Plan for failures before they happen. What do you do if AI sends something incorrect to a customer? Who catches it, how is it corrected, who follows up? What condition would cause you to shut the tool off immediately? Having those answers before you start — not after — separates a professional implementation from one that ends in a crisis.

For our own AI use we go further: a person approves anything that reaches a client or the public (Gugubrand AI policy). The human remains the center. AI amplifies; it does not replace.

What you can do today

1. 1Audit your repetitive tasks. Write down everything you do more than three times a week that involves text — answering questions, drafting messages, summarizing information. Sort each one into the three Delegation groups.
2. 2Pick one Group 1 task and write its three-part description. What it produces, how it reasons, how it must behave and what it must never do. No more than one page.
3. 3Test with at least five real examples before using it with customers. Use actual cases from your history. Review each output, note what breaks, adjust the description and run at least one more round — usually two or three in total — before the AI touches a real customer.

For a broader implementation roadmap, see how to implement AI in your business, and for the rules your own team should follow, how to write an AI use policy.

The 4D framework (Delegation, Description, Discernment and Diligence) is part of the AI Fluency Framework, developed by professors Rick Dakan (Ringling College of Art and Design) and Joseph Feller (University College Cork); the free course materials based on it were produced by Dakan, Feller and Anthropic under a CC BY-NC-SA license. The three task groups and the product/process/rules split in this article are Gugubrand's adaptation for small businesses, in our own words.