EN
Idioma · Language · Lingua
Design the agent: prompt, context and rules.
For Abbie, for your platform or for your human team.
Abbi Studio, the main platform.
Abbi Studio is the main platform of Abbiamo.ai. Abbi Actions, Abbi Tasks and Abbi Flow live inside it; Abbie and Caddie, the AIE agents, are governed from it.
-
Inside Abbi Studio
Abbi Actions
Predefined actions with verifiable outcomes.
Explore the actions → -
Inside Abbi Studio
Abbi Tasks
Ordered combinations of Abbi Actions that deliver a complete outcome.
Meet Abbi Tasks → -
Inside Abbi Studio
Abbi Flow
Chains «Tasks» into a complete process, with conditions, detours and the point where a person steps in.
Meet Abbi Flow →
The file for the agent.
The agent's file: layered prompt, context package, tools, rules and limits, tests, goals and an audit sheet.
-
File
Layered prompt
-
File
Context package
-
File
Tools
-
File
Rules and limits
-
File
Tests
-
File
Goals
-
File
Audit sheet
The file for the role.
The role file: script, knowledge base, systems and permissions, rules, simulations, goals and role sheet. «Studio doesn't replace your training: it gives it content, structure and evidence.»
-
Role file
Script
-
Role file
Knowledge base
-
Role file
Systems and permissions
-
Role file
Rules
-
Role file
Simulations
-
Role file
Goals
-
Role file
Role sheet
Control by design or light guardrails.
-
Abbie and Caddie
Control by design
«Answers only with what it can back up».
-
Other platforms
Light guardrails
«We measure and show you when your agent answers without backing».
Three principles that don't depend on the model.
Seven consoles to define who the AIE agent is, what it knows, where it works and what it achieves. Abbie and Caddie, the Abbiamo.ai AIE agents, are not controlled from the inside: they are governed from the outside, with rules, sources, scripts, «Actions» and metrics managed by a business expert.
-
Principle
What must come out the same every time
Prices, rules, actions and limits live in the console, not in the generative part of the model.
-
Principle
Every decision is traced
Back to the rule or source that produced it. Every change keeps who, when and what.
-
Principle
Improvement without drift
The AIE agent does not learn on its own: it changes when a person approves a learning.
Governed through configuration, not code.
Two ways to govern a model. Only one can be audited and doesn't have to be redone with every new model.
-
From inside the model
Expensive, slow and not auditable
Retraining the model · adjusting weights · writing fragile prompts · redoing it with every new model.
-
From the outside, with Abbi Studio
Rules governed by a business expert
Rules that no level relaxes · approved sources with a validity period · approved scripts and answers · «Actions» with maximum limits and metrics for correction.
The model brings the language. The console brings the framework.
Seven consoles. Without a line of code.
One control question per console. Your team decides the answer.
In sequence
Apart
1Entity
Who is it, and whom does it represent?
Your brand, your tone. Always as an AI.
Define how Abbie introduces herself, how she speaks and what she will never do. Fixed rules are locked and cannot be turned off.
2Training
What does it know, and where did it learn it?
It answers with what you approve.
Each source has a validity period and is approved by your organization. To publish, a minimum accuracy level and validation by a second person are required.
- Payment agreement policyCurrent
- Frequently asked questions guideCurrent
- Product ratesApproved
3Deployment
Where does it act, and who authorized it?
Without the complete review, it does not launch.
Define the goal, audience, channels and schedule of each deployment. Who authorized it is recorded.
In Abbie, a deployment is a goal with its customer base, its script and its schedule.
4Evaluation
How do we know it does it well?
Nothing changes unless a person approves it.
A sample of conversations is scored by hand. If a rule is violated, the deployment can be paused.
Evaluation: a lite version of Abbi PitWall. It works on its own or connected to PitWall.
5«Actions»
What can it execute, and with what limit?
A limit that can only go down.
The catalog of operations Abbie can execute. Each one is defined only once, with its limit.
Abbi Actions are created and published only by Abbiamo.ai; the other consoles can restrict them, never expand them.
6«Tasks»
Which sequence does it complete, and what outcome does it deliver?
A sequence with a verifiable outcome.
Abbi Tasks combine published «Actions» into a sequence with a verifiable outcome; they are composed by Abbiamo.ai, partners and their clients.
- 1Receive a patient's test resultsAbbi Action
- 2Record them in their fileAbbi Action
- 3Notify the physicianAbbi Action
7«Flow»
Which process does it follow, with which detours, and where does a person step in?
The client defines the process; the agent follows it.
Abbi Flow chains Tasks into a complete, end-to-end process, with conditions, detours and the point where a person steps in. It is defined and approved by the client or the partner in Studio; the agent follows it. The agent is not a flow, and it does not make one up. A Flow never expands what its Tasks and Actions allow, and it is not published without end-to-end testing.
This is how Abbi Studio is designed. Each capability is confirmed with your team in the demo.
Entity: who the AIE agent is.
The foundation inherited by every profile and deployment. Every change asks for confirmation, because it changes the AIE agent everywhere. «You decide who it is and what it will never do».
- 1.1 Identity
- Character, brand it represents, introduction, languages and contact channels.
- 1.2 Personality
- Warmth, formality and proactivity from 1 to 5; informal or formal address.
- 1.3 Voice
- Synthetic voice, speed, interruptions and forbidden words.
- 1.4 Creative freedom
- Strict, guided (default) or free within policies.
- 1.5 Global rules
- The lock: escalation, vulnerability, data it never asks for and the autonomy ceiling.
- 1.6 Regulatory framework
- Country, industry and preloaded regulations; conversation retention.
Training: what it can do.
A role is a template, not an «Action». This is where Abbie or Caddie are taught what they know, with sources the client approves. «It only knows what you teach it, from sources you approve».
- 2.1 Profile
- Role and experience level.
- 2.2 Knowledge
- Sources with a validity period (90 days by default) and an owner.
- 2.3 Intents
- At least 5 phrases per intent.
- 2.4 Expertise
- Procedures, decision trees and 10 or more example conversations.
- 2.5 Enabled «Actions»
- Only enables published «Actions», with equal or lower limits.
- 2.6 Validation
- 20 or more cases, with risk and questions without backing.
- 2.7 Answer hierarchy
- Source map: which source answers each data type.
Deployment: goal, audience and schedule.
In Abbie, a deployment is a goal with its customer base, its script and its schedule. In Caddie, the team, the process, the platform and the blocking points where he works. «It works toward a goal, on your schedule and within your regulations».
-
Simple mode · 6 fields
The template fills in the rest
Name · published profile · template · main goal · customer base · contact channel. The template fills in the rest with safe values.
-
Advanced mode · 12 blocks
Every part of the deployment
3.1 General · 3.2 Objective and KPIs · 3.3 Audience and data · 3.4 Script · 3.5 Possible outcomes · 3.6 Knowledge · 3.7 Channels and capacity · 3.8 Scheduling · 3.9 Compliance · 3.10 Deployment rules · 3.11 Escalation · 3.12 Pre-launch review.
-
Before launch
Mandatory review
It does not launch without complete compliance or with blocks in the review.
Evaluation: improvement without drift.
It measures what was achieved and what needs correcting, and returns learnings to the other consoles. In Abbie it measures results against the goal; in Caddie, the quality of his help and of the interaction. «You see what it achieved, and nothing changes unless a person approves it».
-
Human scoring
2% scored by hand
Of conversations; minimum 50 per week.
-
Rules
1 violation pauses the deployment
A single rule violation is enough to pause it.
-
Approval
0 changes without human approval
No learning is applied unless a person approves it.
-
Connected
Abbi PitWall
Evaluation: a lite version of Abbi PitWall. It works on its own or connected to PitWall.
Meet Abbi PitWall →
From a finding to an approved change.
-
Evidence
Finding with evidence
The Evaluation detects what needs correcting and backs it up.
-
Proposal
Proposed
The learning is proposed to the corresponding console.
-
Approval
Approved
Approved by whoever has permission.
-
Application
Applied in its console
The change is recorded in its console.
If it touches Training, the profile is revalidated before it is published.
«Actions»: each operation is defined once.
An «Action» is an operation named with a verb and an object («record promise to pay») that gives the same result given the same inputs and leaves a record. Only Abbiamo.ai publishes it. «Each action is defined once and no one can expand it».
-
Contract sheet
Everything declared before publishing
Name · inputs and outputs · resolution · encrypted connector · validations and error · maximum limits · maximum autonomy · version · tests · log · shelf.
-
Mode
Prepare
Assembles the operation without writing to the systems. It is Caddie's mode.
-
Mode
Execute
Writes and returns a verifiable output. Abbie uses it within her autonomy. If an «Action» can only execute, Caddie cannot use it.
Two modes, always.
Three rules apply on every screen.
-
Rule
Cascading inheritance
Each level inherits from the one above and can only restrict, never expand.
-
Rule
The strictest limit wins
When an «Action» is executed, the minimum among «Action», Entity, Training and Deployment applies.
-
Rule
Everything is recorded
Every change keeps who, when, at which level and what changed.
Maximum amount per promise to pay.
Each level can only lower the limit. What Abbie can commit to is the strictest of them all.
- «Action»
- 50,000
- Profile
- 20,000
- Deployment
- 10,000
- Abbie can commit
- 10,000: the strictest limit.
Illustrative figures.
Answers only with what it can back up. If it has no backing, it says so.
A language model doesn't consult the truth: it predicts the most likely text. If it doesn't know, it fills the gap with something plausible. In a bank or a hospital, a plausible but false balance or deadline is a compliance failure. That is why we don't ask the model to be careful: we take the decision away from it.
-
Rule
Whoever writes doesn't decide
The model provides the language. Which source to use, in what order and when to stay silent are decided by the rules in Abbi Studio.
-
Rule
Not all data is worth the same
Critical data (price, balance, rate, deadline, policy) comes only from sources that back it up.
-
Rule
Either there is backing, or there is no answer
There is no middle ground in which the AIE agent makes assumptions.
Seven levels, in order.
It first classifies the question; then it checks in this order and stops at the first source that has the data.
- 1 · Databases, live
- Client data through read-only «Actions»; includes deterministic calculations. Critical data: yes. Default: if there is a connector.
- 2 · Deployment data
- The deployment's base and offers; if it contradicts level 1, level 1 wins. Critical data: yes. Default: on.
- 3 · Knowledge base
- Approved manuals, policies and rates, with a validity period. Critical data: yes. Default: on.
- 4 · Connected applications
- Apps, microapps and management software the company operates with: calendars, case numbers, requests in progress. Critical data: yes. Default: if there is a connector.
- 5 · Authorized official websites
- Only by exception and when at least three agree; official domains and sources the entity authorizes. Critical data: never a price, a rate or a balance. Default: off.
- 6 · Internal consultation with a human agent
- A supervisor confirms without transferring the conversation. Critical data: yes. Default: optional.
- 7 · The configured route
- Without backing, it says so and follows the configured route: a human agent or a rerouting process. It never answers from the open internet or from the model's memory.
Each level can only restrict the previous one.
A deployment can turn the web off, never turn it on if the entity doesn't allow it.
- 1 · Entity
- The ceiling: which levels the AIE agent can use, which data are critical, which external sources and which web domains are authorized. Who: Abbiamo.ai or the distributor.
- 2 · Training
- The source map: which source answers each data type and at which level. Who: the partner, with the client's approval.
- 3 · Deployment
- Deployment data (level 2), per-deployment restriction, internal consultation with a human agent (level 6) and the «without backing» message. Who: the partner or the end client.
- 4 · Evaluation
- The evidence: answers by level, partial answers, answers without backing and critical data blocked. Who: quality.
- «Actions»
- Level 1 live: queries and calculations as read-only «Actions»; the catalog of external sources. Who: Abbiamo.ai.
Deterministic in substance, natural in form.
-
Step 0
Classify the question
Before searching, it identifies the intent, the data type and whether it is critical. That decides which levels it can use. If the question is ambiguous, it asks for clarification once.
-
Partial answer
Only what has backing
If the question has several parts, it answers the ones that have backing and says clearly which ones don't. The part without backing is never filled in by assumption. Everything is recorded: classification, level and source of each answer.
-
Deterministic in substance
Always the same, given the same question and the same sources
How the question is classified and which sources it checks, in order · which data it can give and from which level it comes · when it stays silent and escalates · what gets recorded.
-
Natural in form
Can vary
The wording and the tone, within what console 1 sets. Same data, different words.
AI brings the raw material. Abbi Studio bounds what it brings and designs what it lacks.
-
Cognitive · High
Memory, attention and analysis
Conversation and history; intent, keywords and tone; classifies and diagnoses the case. On its own, AI can state things without backing. Console 2: sources, intents and source map.
-
Metacognitive · Partial
Self-regulation and calibration
Within rules and script; knows when it doesn't know and says so; improves only if a person approves. On its own, AI doesn't learn in real time. Consoles 1, 2 and 4: threshold and answer hierarchy.
-
Linguistic · High in text, partial in voice
Communication and argumentation
Clear, in Spanish and English; objections handled with approved answers; steers toward the objective. On its own, AI improvises if it has no script. Consoles 1 and 3: voice, languages and script.
-
Creative · Partial, deliberately bounded
Form, options and substance
Varies the wording to sound natural; combines options from the approved catalog; never creates offers. On its own, AI doesn't innovate beyond its training. Console 1: level of creative freedom.
-
Emotional · Partial
Functional empathy and self-control
Acknowledges before proposing; stays calm in the face of anger; detects vulnerability and escalates. On its own, AI doesn't feel; it never fakes it. Console 1: warmth and protocol for abuse.
-
Moral and ethical · Requires explicit design
Lock, transparency and regulations
Rules that no level relaxes; always says it is AI; hours, lists and regulation. On its own, AI has no ethical judgment of its own. Consoles 1 and 3: lock and regulatory framework.
-
Autonomous · Requires explicit design
Decides, acts and graduates
The next step, within the mandate; Abbie executes, Caddie prepares; suggests, confirms or acts on its own. On its own, AI doesn't know its limit without a mandate. «Actions» and consoles 2 and 3: degree of autonomy.
-
Abbi Studio
Bounds and designs
It bounds what AI brings. It designs what AI lacks.
Abbie and Caddie use the same screens.
Abbie talks with the customer: she executes within her autonomy and works toward a deployment's goal. Caddie works alongside the human agent: he prepares, and the human agent authorizes each «Action».
- Entity
- Abbie: autonomy ceiling according to mandate. Caddie: ceiling fixed at «prepares».
- Training
- Abbie: sources, intents and procedures. Caddie: the same, plus control rules and a quality rubric.
- «Actions»
- Abbie: executes within her autonomy. Caddie: prepares; the human agent authorizes each one.
- Deployment
- Abbie: a deployment with a goal and an audience. Caddie: team, process, platform and blocking points.
- Evaluation
- Abbie: results against the goal. Caddie: quality of his help and of the interaction.
Each actor operates its own part.
Abbiamo.ai, distributors, partners and end clients, each with their part of the seven consoles.
- 1 · Entity
- Abbiamo.ai: base, personality and global rules. Distributor: country, language and regulations. Partner: introduction, voice and channels. End client: form of address and forbidden words.
- 2 · Training
- Abbiamo.ai: role templates. Distributor: roles by industry or country. Partner: trains, validates and enables «Actions». End client: loads sources and approves.
- 3 · Deployment
- Abbiamo.ai: templates. Distributor: local templates. Partner: designs and launches. End client: operates in simple mode.
- 4 · Evaluation
- Abbiamo.ai: global quality and engine. Distributor: results across its network. Partner: optimizes and applies learnings. End client: sees results and scores.
- «Actions»
- Abbiamo.ai: creates, tests, versions and publishes. Distributor and partner: consult and request. End client: consults.
- «Tasks»
- They are composed by Abbiamo.ai, partners and their clients, with published «Actions».
- «Flow»
- It is defined and approved by the client or the partner in Studio; console 3 chooses the «Flow» for each deployment.
Assisted collections, console by console.
-
Entity
Abbie for bank X
Spanish, voice and WhatsApp, collections in Mexico, ceiling «executes with confirmation».
-
«Actions»
Record promise to pay
Abbiamo.ai publishes «record promise to pay», up to 50,000.
-
Training
Collections agent
«Collections agent» profile; the «Action» at «suggests», cap 20,000; validation approved.
-
Deployment
Portfolio at 30 to 60 days
Goal: 30% promises at closing; cap of 10,000 per promise.
-
Evaluation
After two weeks
What was said matches what was recorded; moving up to «executes with confirmation» is approved.
-
Outcome
10,000 per promise
The minimum among the «Action», the profile and the deployment. Every promise is recorded in the bank's system.
Illustrative figures.
One sentence per screen.
-
Entity
You decide who it is and what it will never do.
-
Training
It only knows what you teach it, from sources you approve.
-
Deployment
It works toward a goal, on your schedule and within your regulations.
-
Evaluation
You see what it achieved, and nothing changes unless a person approves it.
-
«Actions»
Each action is defined once and no one can expand it.
Products and services, one shared approach.
-
AIE agent
Abbie
Talks with your customers. Acts in your systems.
Meet Abbie → -
AIE agent
Caddie
Caddie guides you. You decide, within the rules.
Meet Caddie → -
Inside Abbi Studio
Abbi Actions
Predefined actions with verifiable outcomes.
Explore the actions → -
Inside Abbi Studio
Abbi Tasks
Ordered combinations of Abbi Actions that deliver a complete outcome.
Meet Abbi Tasks → -
Inside Abbi Studio
Abbi Flow
Chains «Tasks» into a complete process, with conditions, detours and the point where a person steps in.
Meet Abbi Flow → -
Verifier
Abbi PitWall
Verifies, diagnoses and improves your agents in live operation.
Meet Abbi PitWall → -
Services
Microapps, consulting and implementation
Design and rollout based on a diagnosis.
See services →
Enter Abbi Studio.
Answers only with what it can back up. Each action is defined once and no one can expand it.
Sign in →or write to us at hola@abbiamo.ai