Why AI Gives Different Answers to Different People

Two people can ask Insulin the same question and get different answers because its personal surfaces answer with the asker’s own connected accounts, knowledge bases, model providers, memory and conversation. An organization agent takes the per-person accounts, knowledge bases and model providers out, and Planner, in an organization channel, is documented to answer the same whoever sends the message.

---

“It told me something different.” Anyone rolling out an AI assistant to a team hears it sooner or later. Two people ask the same question — what the standard discount is, who owns an account, what a customer agreed to on last week’s call — and get two answers. Now someone has to decide which one is right, and whether the assistant can be trusted at all.

In Insulin the usual explanation is not an unreliable model: the two people were not asking with the same inputs. Here is what changes with who asks, how to find the input behind a difference, and where to ask when the team needs one answer.

Why does AI give different answers to different people?

Because two people asking the same question are rarely asking with the same inputs. An answer depends on what the assistant can reach, which documents it searches, which model writes it, what it remembers and what was said before the question — and on Insulin’s personal surfaces, each of those comes from the person asking.

A per-person input is anything an assistant answers with that belongs to the person asking rather than to the organization. The personal surfaces use them on purpose. Explaining why the built-in assistant is not offered in organization channels, the documentation puts it plainly: “Insulin is your personal assistant and runs with your connected accounts, which should not be used to answer in a room your teammates read.”

So a different answer usually means different material, not a worse assistant. Even identical inputs can produce differently worded answers, so the table below covers the causes a team can control: the inputs.

Same question, different answer: what changes with who asks

Five inputs change with who asks, and each behaves differently on the four surfaces where people put questions to Insulin.

Input	The built-in Insulin assistant	Your personal agent	An organization agent	Planner, in an organization channel
Integrations	Your own user-scope connections	Your user-level connections selected for it	Organization-level connections only	None, permanently
Knowledge bases	Automatic: every knowledge base you own — or your own ticked selection	Your knowledge bases attached to it	Organization knowledge bases attached to it	The channel’s knowledge bases
Model	Your connected providers first, then Fours-hosted	Its Default model, from your providers plus Fours-hosted	Its Default model, from the organization’s providers plus Fours-hosted	The organization’s model configuration
Memory	What it remembers about you, recalled automatically	The docs don’t say	The docs don’t say	Never your memory
Conversation	Your own thread	Your own thread	Your own thread	One shared channel thread

Read down the first two columns and every input belongs to the person asking; read down the last and none does. An organization agent sits in between: its connections, knowledge bases and model providers are the organization’s, but each person still talks to it in their own thread. One decision applies to everyone: whether Fours-hosted models are offered at all is set for the whole organization in its AI model policy.

How each input changes the answer

Each one changes what the assistant has to work with before it writes a word.

Integrations: whose accounts it can reach

The built-in assistant uses the allowed user-scope integrations of whoever is asking. Personal connections such as Gmail, Google Calendar, Google Drive and Gong are per user, so if you have connected Google Drive and your colleague has not, your assistant can reach your Drive and theirs cannot. A personal agent uses the user-level connections selected for it; an organization agent can use only organization-level integrations, never a user-level one; and Planner has none at all — permanently, by design.

Knowledge bases: whose documents it searches

By default the built-in assistant runs on Automatic, searching every knowledge base you own, read-only — so two colleagues on Automatic are each searching their own, rarely the same set. Ticking any knowledge base in its settings switches to an explicit selection that takes over completely and is personal to you.

The gap people miss is that Automatic covers the knowledge bases you own. Organization knowledge bases shared with you appear in the list with a Shared badge but are searched only once ticked, and because ticking switches you to an explicit selection, tick your own as well if you still want them searched.

Custom agents work the other way round, searching only the knowledge bases attached to them: your own on a personal agent, organization ones on an organization agent. Controlling which agents see which documents covers how attachment works.

Model: whose providers answer

There is no per-conversation model picker: Insulin selects the model, and an agent’s ownership decides whose providers it can draw on. The built-in assistant and personal agents add the providers you have connected and never reach the organization’s keys; organization agents add the organization’s. Either way, connected providers are preferred and the Fours-hosted models are the fallback, so if you have connected a provider and your colleague has not, your turns can run on your provider’s models while theirs run on the hosted pool. Who picks the model, and what happens when it fails covers selection and failover in full.

Memory: what it remembers about the person asking

Long-term memory is persistent knowledge an agent carries across conversations — your preferences, decisions and recurring instructions — and recall is automatic. Memory is self-owned: it holds what agents remember about you, so the same question can bring back different memories for each of you. The documentation describes memory for agents in general, not agent type by agent type, so the table leaves two cells as the docs leave them; on Planner it is explicit: Planner never reads your memory. If a memory shaped a wrong answer, edit what your AI assistant remembers.

Conversation: what was said before the question

Short-term memory is the current conversation — its earlier messages and tool results — and it is context for every answer that follows. A question asked twenty messages into a thread about one region is answered in that context; the same question in a fresh conversation is not. Each conversation is a thread between one person and an agent, an organization agent included. Only a channel shares one, because everyone in the channel sees the same messages.

Inbox is per person in the same way: it drafts replies in your writing style, learned from your own sent mail, using a knowledge-base selection that belongs to your account.

What else can change the model between two answers?

Two things, and neither needs anyone to change a setting: failover and the Fours-hosted tier step-down.

Failover happens when a model call fails: the turn retries on the next provider, and if it then succeeds, the reply ends with a short italic note naming both models.

The tier step-down is the one nothing on screen announces. The Fours-hosted pool has a Tier 1 quality model and lighter Tier 2 models, and once your organization’s spend on Fours’ own key passes a point Fours sets, Tier 1 stops being offered for the rest of the billing period. No banner appears and no request fails, but answers may read as less capable. A colleague answered by their own provider is not affected, because a connected provider’s models are always preferred over the hosted pool. The same question asked before and after the step-down can be answered by different models.

A model being retired is a slower change of the same kind; what happens to your agents when an AI model is retired covers it.

How do you find the input behind a difference?

Compare the two answers in this order — the checks on the reply itself first, then the settings behind it.

1. Make sure you asked the same agent. A personal agent is private to its owner, so a colleague’s agent with the same name is a different agent. In Chat’s search, every agent result is tagged org, user or system.
2. Read the end of each reply. An italic note there means the turn failed over to another model.
3. Look for a memory recall indicator. It shows how many memories were retrieved and what they say.
4. Compare the citations. Answers drawn from a knowledge base cite the source document; if the citations differ, compare which knowledge bases each assistant searches.
5. Ask each assistant what it can reach. Every agent can list its integrations, say whether each is user- or organization-level, and check whether each connection works.
6. Compare the defaults. Check the row marked Default in the built-in assistant’s settings, or a custom agent’s Default model.
7. Ask again in a new conversation. That takes out whatever the earlier messages in the thread contributed.

If every input matches and only the wording differs, that is ordinary variation between two answers, not a difference in what the assistant could see.

Where should a team ask when it needs one answer?

On a surface that does not take its inputs from the person asking. That removes the reasons two answers should differ; it does not promise word-for-word identical replies. Two surfaces qualify, and they remove different amounts.

An organization agent with attached knowledge bases

An organization agent uses only organization-level integrations, searches the organization knowledge bases attached to it, and draws on the organization’s connected providers plus the Fours-hosted pool. Put the material the team must agree on — pricing rules, policy, the definitions people argue about — in those attached knowledge bases rather than in each person’s own. Only org admins can create an organization agent, and one can be shared with the whole organization at once.

Two things stay per person: each colleague still talks to it in their own thread, and the documentation does not say how an organization agent’s memory works, so do not count on it either way.

Planner, in an organization channel

Planner is the one surface the documentation says answers the same whoever asks. In an organization channel, a message that @mentions nobody goes to Planner, the coordinator that belongs to the channel rather than to any member. The channels documentation states the rule directly: “Planner never reads your personal integrations, knowledge bases, or memory — its answers do not change depending on which member sent the message.” It draws on the organization’s model configuration rather than any individual’s, it reads the channel’s knowledge bases, and everyone in the channel sees the same messages.

Three things to know before you route a question there:

• Planner has no integrations. When a request needs one, it declines and names the member agent that can do it.
• The guarantee is Planner’s. @mention a specialist and that organization agent answers directly, with an organization agent’s inputs.
• Channel turns do not pause for approval. The tool-approval card does not appear in a channel.

Creating an organization channel takes organization administrator access, and only organization agents can be members; how channels work covers members, roles and Planner’s reach.

Frequently asked questions

Which answer is right when two colleagues get different ones?

Neither by default. Compare what each assistant drew on — the documents it cited, any memory it recalled, its model, its connections and the conversation before the question — and keep the answer grounded in the source your team treats as authoritative.

Does an organization agent give everyone the same answer?

Not guaranteed. It removes per-person connections, knowledge bases and model providers: it uses only organization-level integrations, the knowledge bases attached to it and the organization’s providers. Each person still has their own thread with it, and the docs don’t say how its memory works.

Does Planner answer the same no matter who asks?

That is what the documentation says. Planner never reads your personal integrations, knowledge bases or memory, and “its answers do not change depending on which member sent the message.” It runs on the organization’s model configuration, in organization channels only.

Why doesn’t the assistant search an organization knowledge base shared with me?

Because Automatic, its default, searches only the knowledge bases you own. Shared organization knowledge bases appear in its list with a Shared badge; tick them to include them. Ticking switches to an explicit selection, so tick your own as well if you still want them searched.

Can the same person get a different answer to the same question later?

Yes. A failover, named in an italic note at the end of the reply, or the Fours-hosted tier step-down can change the model; new memories and documents change what it draws on; and identical inputs can still be worded differently.

Takeaways

• Different answers usually mean different inputs: on the personal surfaces, the asker’s own connections, knowledge bases, model providers, memory and thread.
• Automatic searches only what you own. Shared organization knowledge bases are searched once ticked.
• Check the reply before the settings: the failover note, the memory recall indicator and the citations.
• An organization agent removes per-person connections, knowledge bases and model providers, but each person keeps their own thread.
• Planner is the one surface documented to answer the same whoever asks. Aim for consistent inputs, not identical wording.

When a question matters to the whole team, move it off the personal surfaces. Insulin agents for business workflows shows how an agent is scoped, personal or organization-wide, and shared with everyone who asks the same questions.