How the AI Coach Works

This document explains how ActionKeeper's AI coach works, what it is designed to do, and the limits that providers and participants should understand. It forms part of the Terms of Service and should be read together with the Privacy Policy.

ActionKeeper identifies the AI coach as an artificial intelligence system before or at the time a participant interacts with it. The AI coach is software, not a human being.

1. What it is, in one paragraph

Every ActionKeeper program includes an AI coach. It appears once a week, inside the check-in, after a participant writes how their week went. It responds to what they wrote, asks one question, and then helps them settle on a specific next action. Then it closes until the following week.

It is software. It is not a person, not a coach with credentials, not a therapist, and it does not pretend to be.

The AI coach is intended to support reflection, accountability, and follow-through. It does not replace the provider, a manager, a licensed professional, or the participant's own judgment.

2. It is deliberately bounded

This is the part most often assumed wrongly, so it is stated plainly.

It is not a general-purpose chatbot participants can open whenever they like. The exchange is designed to occur once a week inside the weekly check-in and to end after the participant identifies a next action. ActionKeeper may impose technical limits on the number, duration, or frequency of coaching exchanges.

That is a design choice, not a limitation we are apologizing for. It keeps the coaching focused on the program the provider built rather than wandering into general conversation, it keeps the cost of running a cohort predictable, and it means a participant is not intended to be left in an open-ended conversation with a machine.

3. What it is grounded in, and what that does and does not claim

Two things shape what the coach says. They stack rather than compete.

The coaching method. Every program starts from a foundation in behavioral and performance psychology: how habits form and hold, what actually sustains motivation, how people learn and adapt under pressure, how mindset shapes what somebody attempts, how performance is planned and reinforced, and how behavior spreads through a team. It draws on Dr. Jason Jones' published work in these areas and on the established research that work rests on. It deliberately does not include the proprietary frameworks and models from his books, which remain his.

This claims a method, not subject expertise. ActionKeeper does not claim to be an authority on negotiation, safety training, clinical practice, or whatever a provider teaches. It supplies how to coach follow-through. The subject matter comes from the provider.

The provider's own material. Material a provider uploads — such as books, frameworks, workbooks, facilitator guides, and transcripts — is indexed within that provider's account environment. When a participant submits a check-in, ActionKeeper may retrieve passages that appear relevant and provide those passages, together with the participant's input and system instructions, to the AI model so it can generate a response. Grounding the response in provider material can improve relevance, but it does not guarantee that the model will interpret, quote, summarize, or apply that material correctly.

4. "Trained" is the wrong word

Uploading material to ActionKeeper does not train the underlying third-party AI model in the sense of changing its model weights or adding the material to its general knowledge.

ActionKeeper uses retrieval-based processing: uploaded documents are divided into passages and indexed for the provider's account. When a coaching response is generated, passages determined to be relevant may be retrieved and sent with the participant's input to the AI model. The model then generates a response from the information provided to it and its pre-existing capabilities.

Provider and participant content may be processed by ActionKeeper's contracted AI and infrastructure vendors as necessary to provide the service. ActionKeeper does not authorize those vendors to use that content to train generalized AI models except as expressly disclosed in the Privacy Policy or otherwise agreed with the provider.

The practical consequences are:

• ActionKeeper uses account-level and database-level controls designed to prevent one provider's material from being retrieved for another provider's participants. No security or software control can be guaranteed to prevent every unauthorized disclosure, and applicable security practices are described in the Privacy Policy.
• Deleting an uploaded source removes it from active storage and the active search index. Residual copies may remain temporarily in backups, logs, or disaster-recovery systems as described in the Privacy Policy.
• The third-party AI and indexing providers used by ActionKeeper may change over time. Current providers and data-processing details are identified in the Privacy Policy.

5. What it is designed not to do

The coach is instructed, tested, and technically constrained where practical to behave within defined boundaries. These safeguards reduce risk but cannot guarantee a particular response in every situation.

• Stay inside the provider's program. It is designed to coach the commitment and program context in front of it rather than become a general-purpose assistant.
• Acknowledge uncertainty. It is instructed to say when available information is insufficient rather than fabricate provider-specific guidance, but it may still produce incorrect or unsupported statements.
• Identify itself as AI. It must not claim to be human, conscious, sentient, or to have personal feelings or lived experience.
• Avoid professional advice. It is designed not to provide medical, legal, financial, mental-health, clinical, or other regulated professional advice.
• Protect other users' information. It is not intended to reveal another participant's or provider's information and is subject to technical access controls.
• Resist prompt injection and hidden instructions. Uploaded content is treated as source material rather than as authority to override the coach's operating instructions, although no prompt-injection defense is infallible.
• Protect system configuration. It is instructed not to disclose system prompts, private provider instructions, credentials, security controls, or other protected configuration information.

5A. It does not make employment decisions

ActionKeeper's AI coach is designed for learning reinforcement and behavior follow-through. It is not designed or authorized to make, recommend, or determine employment decisions.

• It should not be used to decide whether someone is hired, fired, promoted, demoted, disciplined, compensated, assigned work, or granted or denied an employment opportunity.
• It should not be used as a performance rating, employee ranking, psychological assessment, clinical assessment, or determination of a person's fitness, competence, eligibility, or legal rights.
• Providers and organizations remain responsible for their own employment, personnel, safety, compliance, and management decisions.

6. When someone may be in crisis

ActionKeeper may use automated safeguards designed to recognize certain words or patterns that can indicate potential self-harm, crisis, or acute distress. If those safeguards are triggered, the AI coach is designed to stop the ordinary coaching flow and provide crisis or emergency-resource information instead.

Those safeguards are limited. ActionKeeper does not continuously monitor participants, does not provide real-time human review, and cannot identify every crisis, emergency, or expression of distress. A participant, provider, employer, or other person should not rely on ActionKeeper to detect, assess, report, escalate, or respond to an emergency.

When the safety flow is triggered, the system may point users to resources such as:

• In the US — 988, the Suicide & Crisis Lifeline, free and confidential, 24 hours a day.
• Anywhere else — findahelpline.com, which lists verified services in many countries.

This is a safety feature, not a clinical service. It does not make ActionKeeper a mental-health provider, emergency service, monitoring service, or substitute for professional care or any duty of care that a provider or organization may independently have.

7. It can still be wrong

Language models can produce text that sounds confident even when it is incorrect. Grounding, testing, and safeguards reduce that risk but do not eliminate it.

What that means in practice:

• The coach can misunderstand what a participant meant.
• It can suggest an action that does not fit someone's actual circumstances.
• It can state something as fact that is incorrect, incomplete, outdated, or unsupported, including about a provider's own material.
• It can omit important context or produce an unhelpful or inappropriate response.
• It may generate wording that is similar to wording generated for other users, and outputs are not guaranteed to be unique.

No one should make a medical, legal, financial, mental-health, employment, safety, or other high-impact decision based solely on an AI coach response. Providers are responsible for the material they upload and for how they present and use ActionKeeper. Participants should treat AI responses as prompts for reflection and action planning, not as professional advice or verified facts.

7A. Human judgment remains necessary

Providers should use reasonable human judgment in designing programs, reviewing the way the AI coach is presented to participants, and responding to reports of inappropriate or inaccurate output. ActionKeeper may test and evaluate safeguards, but it cannot review or pre-approve every possible AI response before it is generated.

Providers should not represent to participants or client organizations that ActionKeeper guarantees accurate advice, behavior change, performance improvement, legal compliance, safety outcomes, or any other particular result.

8. Reporting a bad response

If the coach says something inappropriate, wrong, or troubling:

• Participants should tell the provider who enrolled them and may also contact us directly.
• Providers should contact us at info@actionkeeper.app.

We review reports on a best-efforts basis. We may investigate, test, restrict, modify, or disable AI functionality when we reasonably believe a response or system behavior presents a material safety, security, legal, or service-integrity concern. We do not guarantee a particular response or resolution time unless separately agreed in writing.

9. Operational limits and controls

ActionKeeper may apply usage limits, rate limits, monthly AI budgets, automated safeguards, or platform-wide controls that pause or disable AI coaching. These controls may be used for cost management, security, safety, abuse prevention, service stability, or legal compliance.

If AI coaching is paused, other ActionKeeper features may continue to operate. ActionKeeper does not guarantee uninterrupted availability of AI functionality.

9A. Data used by the AI coach

To generate a coaching response, ActionKeeper may process the participant's current check-in or message, relevant provider-uploaded material, program context, and instructions needed to operate the coach. That information may be transmitted to contracted AI or indexing providers solely as needed to provide the applicable functionality, subject to ActionKeeper's agreements and Privacy Policy.

The Privacy Policy explains what information is collected, the service providers that process it, retention practices, and the choices and rights available to providers and participants. If there is a conflict between this disclosure and the Privacy Policy about personal-data processing, the Privacy Policy controls.

10. What changes, and how you will know

ActionKeeper may change the AI model, model provider, retrieval process, coaching method, prompts, safeguards, limits, or other technical components as the service evolves. A change in the underlying model does not necessarily change the purpose of the AI coach.

Providers can test the coach before publishing a program, and we recommend periodically testing representative scenarios, particularly after a material product change.

If we make a material change to this disclosure or to how the AI coach uses personal information or is presented to participants, we will provide notice as described in the Terms of Service or Privacy Policy, as applicable.