What is a Digital Mind? A plain-English guide for expert businesses

A Digital Mind is an AI representative built from the knowledge, instructions and boundaries an expert or business approves. It answers visitors on the owner's behalf, follows a written Playbook, and can request human help, with AI replies continuing until a teammate takes over. It is not the human expert and shouldn't be presented as one.

What does a Digital Mind do?

It does six things, and each one maps to a part of the product that the owner sets up.

Knowledge is what the Mind knows; the Playbook is how it behaves. Plandora's own guidance compares the pair to a new team member: knowledge is what they know, and the Playbook is how you've told them to work. Search Mind knowledge stays on in the Playbook. Search web is a separate switch, off by default, that an owner can turn on when current public information helps, and extra usage charges may apply. With it on, a Mind can draw on the open web as well as approved knowledge, so check which setting a given Mind uses. Audience features and some channels depend on your plan or rollout, so confirm what your account includes before you plan around them.

How does it answer from your knowledge?

Systems that answer from a body of documents often use a technique called retrieval-augmented generation, or RAG. The paper that named it, by Patrick Lewis and colleagues, was accepted at NeurIPS 2020 (opens in a new tab)^[1]. It paired a pre-trained language model with a searchable index of Wikipedia passages, so the model reads retrieved text before it writes. The authors started from a known gap: large models store facts in their parameters, and "providing provenance for their decisions and updating their world knowledge remain open research problems." On the language-generation tasks they tested, the authors reported that retrieval-augmented models produced text that was more specific, diverse and factual than a parametric-only baseline.

That is the general technique, not a description of Plandora's internals. For an owner, the practical start for fixing a wrong business fact is to correct the source, then retest. Say the Mind quotes an outdated cancellation period: edit the Drive document or FAQ answer that holds it, or for an uploaded file, upload the corrected version and move the old one to Trash. Then look for another answer that still says the old thing, and ask the question in several phrasings in the chat preview. A document you write saves as you type and is learned right away, and anything you move to Trash is ignored. Correcting the source does not guarantee the next reply is right, because a language model still writes it. For what each source type accepts and how syncing works, see how a Digital Mind learns your expertise.

What are Teach, Share, Respond and Improve?

Plandora's documentation organises the ongoing work into four verbs. They repeat as a loop rather than running once.

Playbook rules have to be short: each section allows 4,000 characters and the whole Playbook 12,000. Writing a Playbook your Digital Mind can follow goes through each section.

Before sharing, the owner can use the chat preview in the editor to try a typical question, one the Mind should decline and one that needs a person. Finishing setup publishes the Mind at its public URL, and later valid Playbook edits save and publish automatically, so check the Live status after each change. To keep working privately, unpublish.

Share is where visibility comes in. A link in a bio is one route to your Mind; being found and quoted accurately by search engines and AI assistants is another, and generative engine optimisation for expert businesses covers it.

Improve is less abstract than it sounds. In Insights, a question marked No answer or Partial answer offers an Add an answer action that opens the FAQ in Drive with that question already filled in, so the fix is often one new answer. Answer analysis is plan-dependent.

Here is an illustrative example, not a real conversation. A prospect asks, "Is implementation support included?" The FAQ never covers it, so the reply is thin and Insights marks the question as a Partial answer. The owner selects Add an answer, writes the scope they approve, and asks the same question again in the chat preview, phrased two or three ways. For how this loop connects with intent and follow-up, see the expertise operating system.

What happens in one conversation, from question to handoff?

1. A visitor arrives. They can land on a public profile with the owner's name, bio and links, plus Starters, which are prompts they can tap to begin chatting. They can also arrive straight in a conversation through a Quick Chat link or QR code, or meet the Mind on the owner's website. Visitors can chat without signing in up to a limit, then must sign in to continue.
2. The visitor asks a question. The Mind answers from the knowledge it has learned, follows the Playbook and replies, consulting a Specialist if the Playbook calls for one. It can reply when the question arrives rather than when the owner is next free, as long as the plan's AI credits last (see the limits below). How much that matters is the subject of speed to lead with a Digital Mind.
3. The conversation builds context. If the Playbook asks for details, such as a preferred date, the Mind saves them to the contact through a Data Collection. Where Stages are activated, it can move the contact into an Automatic Stage when a reply gives enough evidence. The won Stage is always manual, so a person confirms the outcome.
4. A person is requested. The Playbook can name who handles which kind of question, so a Boundaries line might say that refund requests go to a named person. A handoff request notifies the team and flags the conversation, and the Mind keeps replying until someone takes over.
5. A person takes over. Taking over, or simply sending a reply, pauses the Mind on that conversation, and Release to AI hands it back. Designing that moment well is the subject of human handoff for AI conversations.

What does the owner see?

The Inbox gathers conversations from the connected chat and messaging channels in one place (email is outbound only, so a customer's email reply does not appear there), with quick filters for Unreplied, Mine and AI handling. A contact's details show their Stage, tags, properties, the answers collected through Data Collections and private notes that the contact never sees. Reading these is how the Improve step starts, and Insights adds counts on top: visitors, top questions, answer coverage and which Quick Chat links lead to conversations.

Those counts need careful reading. Quick Chat counts entries, not unique people, and country data is approximate because it comes from IP addresses. A count points you to a question; the conversation tells you what happened.

What is a Digital Mind not?

It is not the human expert. Plandora's FAQ is direct about this: a Digital Mind should not be presented as the expert, and a visitor should not treat the conversation as private correspondence with that person.

It is not a general-purpose AI assistant. A general-purpose assistant is built to help with a wide range of topics. A Digital Mind is Plandora's configured AI representative, built around a specific owner, purpose, approved knowledge, response instructions and business workflow. It is designed to stay grounded in that context, to recognise when it should not answer and to hand off when a person or another process is needed. It is not a general-purpose source of truth. The word chatbot is looser: IBM (opens in a new tab)^[6] notes that chatbots range from scripted rule-followers to AI systems connected to knowledge bases, so the labels overlap. The longer comparison is in how a Digital Mind compares with chatbots and other AI representatives.

It is not a replacement for judgement in consequential decisions. AI responses can be incomplete or inaccurate. Financial, legal, medical and employment decisions still need qualified human review, whatever the Mind says.

Who is a Digital Mind for?

It is built for expert businesses: consultancies, agencies, coaches and educators, and professional-services firms. The shared pattern is that the value sits in a few people's judgement, and the same early questions keep arriving before anyone is ready to book. Plandora's page for experts describes that pattern in more detail, and sector pages cover areas such as skincare, supplements, telehealth, financial advisory and pet care. On the regulated ones the pages keep the Mind to approved information: the telehealth page says it never diagnoses or triages symptoms, and the financial advisory page says it never gives personal recommendations. This article is not medical, financial or legal advice.

A Digital Mind fits when you can write down what you are willing to say, when a person can take over, and when someone will read the conversations and fix the gaps. It fits poorly where most conversations turn on consequential decisions, or where nobody can own the review. Each personal Plandora account has one Digital Mind. Under an Enterprise contract, each person who joins keeps their own Mind, the organisation can have a Mind of its own, and an owner can create extra managed Minds if the contract includes them.

What are the limits?

AI replies can be wrong, and a Digital Mind is no exception. The US National Institute of Standards and Technology (NIST) uses the term confabulation, defined in its Generative AI Profile of July 2024 (opens in a new tab)^[2] as the production of "confidently stated but erroneous or false content." NIST adds that the risk arises when people believe the false content, often because of how confident it sounds, and that it may be especially important to monitor where generative AI feeds consequential decision-making.

Retrieval narrows the problem without removing it. In a 2024 preprint (opens in a new tab)^[4], researchers at Stanford and Yale ran 202 legal research queries through three commercial tools that retrieve from legal databases. Across the three tools, hallucinated responses ranged from 17% to 33% of all responses, as the authors coded them, which was lower than for the general-purpose GPT-4 model they compared against. These are different products in a harder domain, and the figures come from the May 2024 preprint. The peer-reviewed version, in the Journal of Empirical Legal Studies in 2025 (opens in a new tab)^[5], says the vendors' tools each hallucinate "more than 17% of the time". The study shows that grounding is not a guarantee. It does not tell you how often any particular Mind errs.

Telling people they are talking to an AI matters, but it is not a fix for accuracy. NIST's AI Risk Management Framework of January 2023 (opens in a new tab)^[3] treats transparency as part of trustworthy AI and then cautions: "A transparent system is not necessarily an accurate, privacy-enhanced, secure, or fair system."

Availability has a limit too. AI replies, training and Automation runs pause when the plan's monthly AI credits run out, and skipped replies are not replayed when the pool refills. Conversations the Mind could not answer show in the Inbox as "Your Mind couldn’t reply", and the team has to answer them by hand.

The owner has levers for all of this. Set Boundaries, test before launch, read conversations afterwards and keep a person on call. If an answer is missing, add it in Knowledge; if the Mind behaves oddly, clarify the Playbook. In Insights, Answered is a system assessment, not proof that the visitor was satisfied or the answer was correct, so read the conversation before you judge quality. Only add material you have the right to use, and keep confidential, regulated or sensitive personal information out unless your agreement and workspace configuration expressly support it. Why AI replies go wrong covers the causes and the controls in more depth.

How do you start?

Give the Mind one clear first job, such as explaining your services, answering product questions or helping someone book. Write the five questions you are asked most into the FAQ in Drive and check the Goal, Boundaries and Flow in the Playbook. After launch, read the first conversations in the Inbox and, where Insights answer analysis is available, the questions marked No answer or Partial answer, then fix the gaps in Knowledge or the Playbook. You can add more jobs later, once the first one holds up in real conversations.