AI Studio

AI Studio is the authoring and testing surface for every AI agent in AppBO. Instead of configuring an agent's personality in one place, its knowledge in another and its booking behavior in a third, the Studio gathers the whole definition of an agent into one workspace: who it is, what it knows, what it is allowed to do, how it should sound, and when it should step back and hand over to a human.

The mental model is a rehearsal room with a one-way door. Work in the Studio happens against a draft of the agent. You can rewrite instructions, swap knowledge sources, change booking rules and run as many test conversations as you like without a single real contact seeing any of it. Only when you deliberately enable the agent on a channel does the configuration start answering live traffic. That separation is the most valuable thing the module gives you, because AI agents fail in ways that are obvious in a test and expensive in production.

Treat the Studio as where you define behavior, not where you define process. Process belongs in workflows. The agent's job is to converse, interpret and act inside boundaries you set; the workflow's job is to decide what happens to the record afterwards. Operators who push orchestration into the agent's instructions end up with an agent that is hard to debug and a build nobody else can maintain.

Agent identity and voice

• Agent name and the persona it presents as, including whether it discloses that it is an assistant
• Tone, reading level and response length guidance
• Language handling and whether the agent mirrors the contact's language
• Business context the agent always has: who the client is, what they sell, where they operate

Instructions and goals

• The core instruction set: the agent's objective, the order it should pursue things in, and what it must never do
• Qualifying questions the agent should work through before it offers to book
• Explicit refusals: topics to decline, claims it must not make, information it must not guess

Knowledge sources

• Documents, pasted text, FAQ pairs and crawled website pages the agent can draw on
• Which sources are active for which agent, so a sales agent and a support agent can read different material

Actions and tools

• Whether the agent may look up availability and book, reschedule or cancel appointments
• Whether it may update contact fields, apply tags or capture structured answers
• Which calendars, pipelines and fields it is permitted to touch

Guardrails and handover

• Handover triggers: explicit requests for a human, repeated failure to answer, sentiment or keyword conditions
• Escalation destination, whether that is an internal notification, a task or a pipeline move
• Operating hours for the agent and what it does outside them
• Message limits per conversation and a stop condition so the agent never loops

Testing

• A test conversation pane where you play the contact and watch the agent reason and act
• Review of past live conversations so you can find the exact turn where behavior went wrong

1. Decide the agent's single job before you open the Studio. One agent that qualifies inbound leads and books is a good scope. One agent that also handles billing disputes and support escalations is not.
2. Create the agent and give it a name that identifies it internally, not just to contacts, so a team managing several agents can tell them apart.
3. Write the business context block: who the client is, what they offer, service area, hours, and anything the agent would otherwise invent.
4. Write the core instructions as objective first, then sequence, then constraints. Put the refusals in their own clearly separated section so they survive later edits.
5. Attach knowledge sources. Start with the handful of documents that answer the questions you actually receive, not the entire website.
6. Set the tone and length guidance, then test immediately. Length is the setting operators most often get wrong, and it shows up on the first reply.
7. Enable only the actions the agent needs. If it does not need to write to the contact record, leave that off.
8. If the agent books, connect the specific calendar it should use and confirm the slot rules, buffers and timezone behave the way you expect before the agent is ever involved.
9. Define handover: what triggers it, who gets told, and what the agent says as it hands off. An agent that escalates silently is worse than one that never escalates.
10. Set operating hours and the out-of-hours behavior, plus a hard cap on messages per conversation.
11. Run at least a dozen test conversations: the easy path, the vague inquirer, the price shopper, the angry contact, the one who asks something the knowledge base does not cover, and the one who asks for a human on the first message.
12. Fix what breaks by editing instructions and knowledge, not by adding more rules on top of rules. Re-test after each change.
13. Enable the agent on one channel only, with a narrow audience if the platform allows it, and watch live conversations daily for the first week.
14. Once stable, document the agent's scope and handover rules for the client so their staff know what the agent will and will not do.

AI Studio sits upstream of almost everything conversational. The agents you author here are what the Conversation AI and Voice AI modules actually run; the knowledge you attach is what the AI Knowledge Base module manages; and the handover rules you set are what hands control to Workflows, Pipelines and Tasks.

Inbound lead qualification

A paid ad drives traffic to a funnel. The form creates the contact, a workflow tags it by source, and the chat widget hands new inquiries to the Studio-authored agent. The agent qualifies against the questions you configured, books into the sales calendar, and the workflow moves the opportunity to a booked stage and notifies the rep.

Speed-to-lead on missed calls

A missed inbound call fires a workflow that opens an SMS conversation. The agent takes over by text, finds out what the caller needed and books or escalates. The Studio is where you make sure that agent sounds like a person covering the phones rather than an autoresponder.

After-hours coverage

Outside business hours the agent answers, answers the common questions from knowledge, and books the first available slot the next working day. The operating-hours setting in the Studio and the calendar's availability rules have to agree, or the agent will cheerfully book a slot nobody is there for.

• Testing is not production. Test conversations are cooperative because you know the agent. Real contacts send one-word messages, photos, voice notes and nonsense. Always review real transcripts for the first week.
• Instruction bloat is the main failure mode. Every time an agent misbehaves there is a temptation to add another rule. Past a certain length the instructions contradict each other and behavior becomes unpredictable. Rewrite rather than append.
• The agent will fill gaps if you let it. Anything you do not state, and do not forbid it from guessing, is a candidate for invention. Pricing, availability, clinical or legal claims and policy details need either a knowledge source or an explicit refusal.
• Enabled actions are real actions. An agent allowed to book will book, including at times you did not intend if the calendar rules are loose. Get the calendar right first.
• Handover needs a human on the other end. Escalation that fires into an unwatched inbox or an unassigned task is the most common silent failure in AI builds.
• Draft and live can drift. After you enable an agent, remember that later Studio edits are changes to a live system. Re-test before you publish, not after.
• Knowledge changes do not rewrite instructions. If a policy changes, update both the source document and any instruction that restates the old policy.
• Scope creep across clients. Copying an agent between accounts carries the previous client's context. Check every context and knowledge reference after a copy.