AI clone vs chatbot vs Digital Mind: what's the difference?

The three labels overlap and products under each differ, so a name says little about what a product can do. A chatbot may be a script or a generative system, an AI clone is built to represent one named person, and a Digital Mind is Plandora's AI representative. Compare them by documented capability.

For a buyer, the useful comparison is accountability: what the system may say, what it must refuse, and what happens while a person is needed.

What do the five labels mean?

Product marketing blurs these labels, so these are working definitions, not a market taxonomy. Any product can combine them.

• Rule-based chatbot. Software that follows predefined rules, decision trees or scripted responses. IBM (opens in a new tab)^[1] defines a chatbot as "a software application that communicates with people through text or voice" and calls the rule-based kind the traditional one.
• General-purpose AI assistant or custom GPT. A large language model you chat with for drafting, research and analysis, sometimes configured with your own instructions and uploaded files.
• AI clone (also described as a digital twin or digital double). An AI built to represent one named person in text, voice or video. Axios (opens in a new tab)^[7] reports that executives feed in their writings, keynote speeches and interviews, and sometimes meetings, to build text and voice chatbots and video avatars from that material. This is the market term, and the rest of this article uses it only for this category.
• Customer-service AI agent platform. Software aimed at support teams that answers and resolves tickets using help-centre content, order data and a help desk.
• Digital Mind. Plandora's product: an AI representative built from the knowledge, instructions and boundaries an expert or business approves. It answers visitors, collects permitted details and can request a person.

The labels overlap in practice. IBM says many AI chatbots connect with business applications, databases and knowledge bases, and can transfer more complex conversations to a human. Delphi, which Axios calls an AI cloning platform, says in its Hosted Actions help page (opens in a new tab)^[2] (updated 18 September 2026) that an action can "tag and update contacts", "notify you when a human should step in" and connect to a CRM. The same page uses the term "Digital Mind" for Delphi's own product. A category label therefore cannot tell you whether a product has knowledge controls, contact records or a handoff. The product has to.

How do you compare them?

A table by category would need every cell to hold for every product, and none can. This one sets out what Plandora documents for a Digital Mind and points each row to the scorecard question to put to any vendor, whatever the label. Where the right-hand column says "varies", no specific documented capability was checked at that vendor's own pages, apart from the examples above.

The table does not rank products. In the Plandora column, Audience, Automations and some channels can depend on plan or rollout.

Where do these tools go wrong?

Every product in this space can fail in specific ways. The two cases below are illustrations, not reported incidents.

A scripted flow loops. A visitor asks something the script never anticipated. The flow asks for an account number the visitor does not have, then asks again. The US Consumer Financial Protection Bureau's June 2023 report on chatbots in consumer finance (opens in a new tab)^[3] says chatbots "may be useful for resolving basic inquiries, but their effectiveness wanes as problems become more complex." It also describes customers stuck in loops "without an offramp to a human customer service representative." The report covers chatbots in general, so the loop is not only a scripted-tool problem. Recovery: give every step an exit to a person, and read the conversations where visitors reach the fallback or repeat themselves.

A generative reply invents an entitlement. A visitor asks whether a refund is still possible after the cancellation window. The reply says yes, although the policy says no. Recovery: correct or remove the outdated or conflicting source, route refund decisions to a named person in the Boundaries, retest the question in several phrasings, and have a person contact the visitor. In Plandora, the owner can find that conversation in the Inbox and take over.

Three pieces of evidence show how real the second case is.

• A September 2025 paper by Kalai and colleagues (opens in a new tab)^[4] argues that language models hallucinate because training and evaluation reward guessing over admitting uncertainty. That suggests a model left to itself leans towards answering, so refusing or handing over has to be designed in.
• A preregistered study of commercial AI legal research tools, published in the Journal of Empirical Legal Studies in 2025 (opens in a new tab)^[5], found that tools from two major vendors each hallucinated more than 17% of the time. Stanford's May 2024 summary of the study (opens in a new tab)^[6] says it used over 200 open-ended legal queries and counted incorrect answers and answers citing sources that do not support them. The tools hallucinated less than GPT-4, but the authors called the vendors' claims overstated. The rate will not transfer to your use, and the study compared whole products, so it does not isolate why they did better than GPT-4.
• Axios reported in June 2025 that executives were building AI clones from their writing, speeches and interviews. The reporter noted that training on one person's material "could mean fewer inaccuracies, but it doesn't mean none." The piece closed on the risk that AI clones could trigger "a backlash over authenticity."

For a product built around resemblance to a person, authenticity can be both the appeal and the risk. If a visitor believes they are talking to the person, a wrong answer lands on the person. Plandora's public FAQ takes a different stance: a Digital Mind "is not the human expert and should not be presented as one", and consequential financial, legal, medical and employment decisions still need qualified human review. For why replies fail and what to change, see why AI replies go wrong.

How does a Digital Mind differ from an AI clone?

Not by a capability only Plandora has. The Delphi example shows that an AI clone vendor can document business actions too, so a feature list cannot settle the question alone. What Plandora states is the design intent. Its FAQ defines a Digital Mind as an AI representative that can answer questions, guide a visitor, collect permitted details and help move a conversation to the next step, and says it is not the human expert. What is a Digital Mind? covers the model in full.

In the product, that job is built from named parts. Knowledge comes from Drive and Social media, the Playbook holds Goal, Role & style, Boundaries and Flow, and Contacts, Data Collections and Stages keep audience context. How a Digital Mind learns your expertise explains the knowledge side.

Voice is where the two overlap most, and it sits in tension with the authenticity risk above. Voice is optional in Plandora, and without it a Mind answers in text. With a trained voice, visitors can play a reply spoken aloud, and the public FAQ says a voice model based on the owner's likeness is plan-dependent, with provider requirements and consent to confirm. A trained voice sounds like the person, so disclosure matters more. In the current public chat view, the header shows a small AI tag beside the Mind's name; check how each channel you use discloses it, and say plainly in your profile and Playbook that visitors are talking to an AI representative.

A Digital Mind still generates text with a language model, so the error evidence above applies to it too. Its replies draw on three layers: the model's own knowledge, the sources the owner selects and, if switched on, web search. Choosing sources does not mean every generated statement was approved.

Ten questions to ask before you buy

1. Which knowledge layers does it use, and who approves each? Ask whether it relies on model knowledge, on sources you select, on web search, or on a mix, and for the list of sources and how each one gets in.
2. Can you see, update and remove what it knows? An outdated policy should be removable without a rebuild.
3. What does it do when it does not know? Ask questions outside the material and note whether it declines, hands over or guesses.
4. Which limits can you set, and how do you test them? Name the topics and decisions reserved for a person, then test them with paraphrases, follow-up pressure (such as claiming a manager has already approved the request) and out-of-scope requests, and record every failure. A pass in a preview shows observed behaviour, not a guarantee: instructions written to a model are not independently enforced restrictions unless the vendor documents a separate mechanism.
5. How does a person take over, and what happens while you wait? Ask whether the product pauses, queues the request or keeps replying until someone takes over. In Plandora the request notifies the team and the Mind keeps replying, so the Boundaries should say what it may tell the visitor meanwhile. Look for a shared inbox with history and a way to hand back to the AI. See human handoff for AI conversations.
6. Does it disclose that it is AI? Ask how visitors are told and whether it ever presents itself as the person. Disclosure rules vary by jurisdiction, and AI disclosure and the EU AI Act covers one regime.
7. Where does it run, and do conversations meet in one place? List the channels your audience uses, check each is covered and ask which of them reach one shared inbox.
8. Does it keep context and intent? Check whether it can save details you ask for, link chats to a contact, mark progress and sync to your CRM, or starts every chat blank.
9. Where does the data go? Ask the same five things of every vendor: whether content trains models, how long records are kept, how deletion works, who can read conversations, and what is committed in the contract. Plandora's FAQ advises confirming model-training and retention terms with sales, and keeping confidential or regulated personal data out unless your agreement supports it.
10. What does it take to set up and run, and what does it cost? List the components: platform fee, usage, channel provider fees, voice or video generation, and your time reviewing conversations and updating knowledge. Ask what happens at the limit. In Plandora, AI replies, training and Automation runs pause when credits run out, and skipped replies are not replayed.

When is a simpler tool enough?

A rule-based flow is enough when the questions are few, stable and exact: opening hours, order status, booking steps. IBM says traditional chatbots work well for predictable requests but struggle with questions outside their programmed responses. A general-purpose assistant is enough when you are the only reader and can judge the output.

Skip a Digital Mind if every answer needs lawyer-approved wording, if your enquiries are mostly order lookups in a ticketing system, or if nobody will read conversations. Any AI that talks to the public needs a person who reviews what it said. In Plandora, Insights can flag questions as No answer or Partial answer and offers an Add an answer action that opens the FAQ with the question filled in. That answer analysis needs a paid plan, and Answered is a system assessment, not proof the visitor was satisfied.