Artificial intelligence is creating new opportunities for charities to reduce administration, support employees and improve how they respond to supporters.
A fundraising team might spend hours answering routine enquiries. A volunteer coordinator may regularly search for information across different systems. Employees delivering services may need quicker access to approved procedures and guidance.
Salesforce Agentforce offers ways to support these activities through AI agents that can retrieve information, respond to questions and carry out configured actions.
However, charities cannot approach AI adoption in exactly the same way as every commercial organisation.
They may hold sensitive information about beneficiaries, vulnerable people, donors and volunteers. They have responsibilities to supporters, trustees, regulators and the communities they serve. An inaccurate response or inappropriate disclosure can have consequences beyond an inconvenient customer experience.
At Sweet Potato Tec, we believe charities should explore Agentforce with clear objectives, appropriate safeguards and realistic expectations.
The starting point is not how much work AI can automate. It is identifying where AI can genuinely help without compromising trust, privacy or human accountability.
How can charities use Agentforce safely and responsibly?
Charities can use Salesforce Agentforce responsibly by starting with low-risk, clearly defined tasks, restricting access to appropriate data and establishing human oversight. Each AI agent should be tested for accuracy, privacy and unexpected behaviour before deployment. Charities also need escalation procedures, monitoring, audit records, staff training and a documented AI policy. Sensitive decisions and safeguarding responsibilities must remain under human control.
What is Agentforce, and which charity tasks are suitable for AI?
Agentforce is Salesforce’s platform for building AI agents that can respond to requests, retrieve relevant information and perform configured actions within defined processes.
Unlike a conventional chatbot that primarily provides predefined answers, an Agentforce agent can use relevant information and tools to support a more complex workflow.
For example, an appropriately configured agent might retrieve information from an approved knowledge base, categorise an enquiry or initiate a permitted Salesforce process.
However, the fact that an AI agent can perform an action does not mean it should be authorised to perform it independently.
For charities, the appropriate level of autonomy depends on the nature of the task, the information involved and the potential consequences of an error.
Which use cases are suitable for an initial rollout?
A sensible starting point is a task that is repetitive, well understood and relatively low risk.
Potential examples include:
- Helping employees find approved internal procedures or guidance.
- Answering straightforward public enquiries using verified charity information.
- Categorising routine administrative requests for staff review.
- Summarising non-sensitive operational information for authorised employees.
- Helping staff locate relevant Salesforce records within their existing permissions.
- Drafting routine communications for a person to review before sending.
These are potential use cases rather than functions that every charity can activate immediately.
Each requires appropriate configuration, access controls, testing and licensing.
An internal knowledge assistant may be a particularly useful starting point because its scope can be limited to a defined collection of approved material.
For example, a charity could explore an agent that helps employees find information about internal procedures.
The agent should be instructed to use the approved source material, indicate when it cannot find an answer and direct the employee to the appropriate person when further judgement is required.
Which tasks should charities approach with greater caution?
Some activities carry substantially greater risks.
These include assessing safeguarding concerns, determining eligibility for essential services, interpreting sensitive case information, making decisions about beneficiaries or providing individualised medical or legal guidance.
Such activities may involve vulnerable people, complex circumstances and consequences that require professional judgement.
An AI agent might support an authorised employee by retrieving an approved procedure or helping organise information.
It should not be treated as an independent safeguarding officer, caseworker or decision-maker.
At Sweet Potato Tec, our Agentforce approach begins with identifying the right use case and designing the agent around a defined business need.
Our Salesforce services for charities and nonprofits also recognise that technology must support the organisation’s mission rather than introduce unnecessary complexity.
For charities, this means considering risk alongside potential efficiency.
How should charities protect sensitive information when using Agentforce?
Charities should restrict AI agents to the information and actions required for their approved tasks, applying appropriate Salesforce permissions and data-protection controls.
This is one of the most important aspects of responsible Agentforce implementation.
A charity may use Salesforce to manage supporter relationships, donations, volunteering and service delivery.
Not all that information should be available to every employee.
The same principle applies to AI agents.
An agent designed to answer routine fundraising questions should not automatically have access to confidential beneficiary records.
An internal knowledge assistant should not retrieve sensitive case notes simply because those records exist in Salesforce.
Apply the principle of least privilege
The principle of least privilege means providing only the access necessary to complete an authorised task.
Before deploying an agent, the charity should establish:
- Which Salesforce records and fields the agent needs.
- Which external data sources it can access.
- Which actions it is permitted to perform.
- Whether it can create, update or delete information.
- Who can interact with the agent.
- Which information must remain inaccessible.
These decisions should be documented and tested.
Salesforce provides security and trust capabilities for Agentforce, including controls relating to permissions, data access and AI interactions.
However, Salesforce’s security guidance describes a shared responsibility model.
The platform provides technical controls, while organisations remain responsible for configuring access and agent-specific safeguards appropriately.
A charity should therefore never assume that using Agentforce automatically makes every possible interaction secure.
What about beneficiary and safeguarding information?
Beneficiary records may contain particularly sensitive personal information.
Depending on the charity’s activities, this could include health information, details of personal circumstances or safeguarding-related records.
Before allowing an agent to access such information, the organisation should establish whether that access is genuinely necessary.
In many cases, the safer decision will be to exclude sensitive records from an initial AI deployment.
If a later use case requires access, the charity should conduct a more detailed assessment involving relevant operational, technical, safeguarding and data-protection personnel.
The same caution applies to connected systems.
An integration that makes information available to Salesforce should not automatically make it available to Agentforce.
Access needs to be considered throughout the full data flow.
Our Salesforce for Human Services Charities work emphasises role-based security, appropriate case management and safeguarding-aware processes.
Those principles remain important when AI becomes part of the technology environment.
Does Agentforce automatically ensure UK GDPR compliance?
No.
Salesforce provides technical security and privacy capabilities, but charities remain responsible for complying with applicable data-protection requirements.
The organisation needs to consider the purposes for which personal information is used, the relevant lawful basis, data minimisation, transparency, security and retention.
Depending on the proposed processing and its risks, a data protection impact assessment may be required.
Charities should also consider whether people need to be informed that their information is being processed through an AI-enabled system.
The Information Commissioner’s Office provides guidance on transparency and AI-related personal-data processing.
The precise requirements depend on the charity’s use case.
A technical implementation should therefore be informed by the organisation’s data-protection responsibilities rather than treated as a substitute for them.
How can charities prevent inaccurate or misleading Agentforce responses?
Charities can reduce the risk of inaccurate AI responses by using reliable information sources, defining clear boundaries, testing realistic scenarios and requiring human review where mistakes could cause harm.
Generative AI can produce responses that appear convincing but are incorrect.
These errors are often described as hallucinations.
An agent may misunderstand a question, use outdated information or generate a statement that is not supported by its available sources.
For a charity, this could mean giving someone incorrect information about a fundraising campaign, volunteering requirements or an available service.
In a sensitive context, the consequences could be more serious.
Use approved and reliable information
An agent should have access to the information necessary for its task.
For a public-facing enquiry assistant, that might mean approved website content and an appropriately maintained knowledge base.
For an internal employee assistant, it could involve verified policies and procedures.
However, connecting an agent to more information does not automatically improve accuracy.
Outdated, contradictory or poorly organised documents can make responses less reliable.
The charity should establish who owns the source material and how it will be reviewed.
If a policy changes, the information available to the agent should be updated.
Define what the agent must not do
A responsible agent needs boundaries.
For example, a charity might configure an agent to answer general questions about volunteering but prevent it from determining whether an individual is suitable for a particular safeguarding-sensitive role.
An agent answering public enquiries could be instructed not to provide personalised medical advice.
A fundraising assistant might be restricted from making promises about the use of donations that are not supported by approved information.
These boundaries need to be reflected in the agent’s configuration, permissions and available actions.
Instructions alone should not be treated as a complete security control.
What should happen when the agent does not know the answer?
The agent should have an appropriate way to acknowledge uncertainty and refer the enquiry to a person.
An incorrect answer delivered confidently may be more damaging than an honest explanation that further assistance is needed.
Charities should establish what happens when an enquiry falls outside the agent’s approved scope.
That could involve creating a case for an employee, directing the person to an appropriate contact or providing approved instructions for obtaining assistance.
The correct response will depend on the use case.
For urgent or safeguarding-related enquiries, the escalation process needs particular attention.
The agent should not create the impression that an urgent issue has been reviewed by a qualified person when it has not.
When should an Agentforce agent hand over to a human?
An Agentforce agent should hand over when an enquiry requires professional judgement, involves sensitive or unexpected circumstances, falls outside its authorised scope or presents a meaningful risk if handled incorrectly.
Human oversight should be designed into the process before the agent goes live.
It should not be added only after a serious problem occurs.
Identify escalation situations in advance
A charity may want an agent to answer routine questions about its services.
However, an enquiry might reveal that someone is distressed, requires urgent assistance or wants to make a complaint.
The agent should not continue treating that conversation as a standard administrative request.
The charity needs clear rules about which situations require human involvement.
Examples might include:
Safeguarding concerns: An enquiry suggests that a child or vulnerable adult may be at risk.
Sensitive service requests: A person describes circumstances requiring assessment by a trained employee.
Complaints: A supporter raises a serious concern about the organisation or its services.
Financial exceptions: A donation, refund or payment issue requires investigation or approval.
Uncertain information: The agent cannot find an adequately supported answer.
Requests outside the approved scope: Someone asks the agent to perform an action it has not been authorised to complete.
Not every situation requires the same response.
Some may need routine referral.
Others may require an urgent escalation route that reflects the charity’s existing safeguarding procedures.
Why must human accountability remain clear?
An AI agent can support a process, but it does not take responsibility for the charity’s decisions.
Trustees and relevant employees remain responsible for governance, operational decisions and compliance.
For example, an agent might help a service-delivery employee locate an approved policy.
The employee remains responsible for applying that policy appropriately.
An agent might prepare a draft response to a supporter.
The authorised employee remains responsible for reviewing and approving it when the workflow requires human approval.
This distinction is essential.
AI should help employees perform their responsibilities more effectively, not create uncertainty about who is accountable.
At Sweet Potato Tec, our Agentforce implementation approach includes routing, escalation, testing and user enablement.
The purpose is to ensure AI supports employees within clearly defined processes.
How should charities test Agentforce before allowing it to interact with supporters?
Charities should test Agentforce against realistic situations, including incorrect information, unexpected questions, access restrictions and scenarios requiring human escalation.
A demonstration that works under ideal conditions is not sufficient evidence that an agent is ready for everyday use.
Testing needs to reflect what could happen when employees or supporters interact with the system.
Test ordinary tasks
The charity should confirm that the agent performs its intended function accurately.
For example, an internal knowledge agent might be asked to retrieve approved volunteering procedures.
Testing should establish whether it finds the correct information, presents it clearly and avoids adding unsupported instructions.
Test unusual and difficult requests
People do not always ask questions in the way the implementation team expects.
They may use informal language, provide incomplete details or ask several questions at once.
The agent should be tested against these variations.
It should also be tested with misleading instructions and attempts to persuade it to ignore its boundaries.
These are important because AI systems can encounter malicious or unexpected input.
Test data access and permissions
The organisation should confirm that users cannot obtain information they are not authorised to see through the agent.
Testing should include attempts to access restricted records and requests involving information outside the agent’s permitted scope.
The charity should also check what actions the agent can perform.
For example, if an agent is intended only to retrieve information, it should not be able to change supporter records.
Test escalation
Escalation processes should be tested with realistic examples.
Does the agent recognise when a question needs human attention?
Does the correct employee or team receive the request?
Is the necessary context transferred appropriately?
Does the person making the enquiry understand what will happen next?
For sensitive situations, the charity should test the relevant escalation procedures with the employees responsible for managing them.
Start with a controlled rollout
Rather than introducing several agents across the organisation simultaneously, a charity may benefit from starting with one carefully defined use case.
This allows the organisation to assess how the agent performs and whether its safeguards are effective.
Sweet Potato Tec’s Agentforce Quickstart reflects this approach.
We help organisations identify practical opportunities, assess readiness and develop a sensible roadmap rather than assuming they need to automate everything immediately.
Our published Agentforce implementation guidance also emphasises focused use cases, structured testing and scaling only when the initial solution has demonstrated value.
For charities, that staged approach is particularly useful because it creates opportunities to review risks before expanding access or functionality.
What governance and AI policies should charities establish?
Charities should establish an AI policy that defines approved uses, responsibilities, data-access requirements, human oversight, testing, monitoring and procedures for responding to problems.
The policy should reflect the organisation’s actual activities.
A charity providing sensitive support services may need different controls from an organisation using Agentforce only for internal administrative assistance.
However, both need clear accountability.
Who should be involved in AI governance?
The relevant people will depend on the charity’s structure.
They may include senior leadership, trustees, the Salesforce administrator, information-governance personnel, safeguarding leads and employees responsible for the processes being automated.
The organisation should agree who can approve a new AI use case and who is responsible for its operation.
It should also establish who can suspend an agent if a significant problem is identified.
What should an AI policy cover?
A practical policy should address the following areas.
| Governance area | Questions the charity should answer |
| Approved use cases | What is the agent permitted to do, and what is outside its scope? |
| Data access | Which records, documents and systems may the agent access? |
| Human oversight | Which actions require review or approval? |
| Escalation | When and how must the agent refer a matter to an employee? |
| Testing | What checks must be completed before deployment? |
| Transparency | How will employees and relevant users understand when AI is involved? |
| Monitoring | Who reviews agent performance and investigates errors? |
| Audit records | What evidence of agent activity must be retained and for how long? |
| Incident response | How can the agent be restricted, corrected or disabled? |
| Training | What must employees understand before using or supervising the agent? |
The policy should also be reviewed as the charity introduces new use cases.
An agent that only retrieves internal guidance may have relatively limited permissions.
A later agent that updates records or communicates with supporters will require additional consideration.
How can charities maintain auditability?
Organisations need sufficient visibility to understand what their agents have done.
Depending on the Salesforce products and configuration, monitoring and audit capabilities may help review agent activity and investigate problems.
Salesforce’s official Agentforce guidance discusses monitoring, audit trails and the importance of reviewing agent behaviour.
However, charities should confirm which capabilities are available in their particular environment and configure them appropriately.
They should also consider the sensitivity and retention of logs.
An audit trail can itself contain personal information, so it needs appropriate access controls and retention arrangements.
The objective is to make AI activity accountable without unnecessarily retaining sensitive information.
How can charities monitor Agentforce and improve it after launch?
Charities should monitor Agentforce through a combination of response quality, escalation performance, employee feedback, operational results and incident reporting.
Successful deployment is not the end of responsible AI management.
The organisation needs to understand how the agent performs in everyday use.
Measure accuracy, not just speed
An agent may respond quickly while providing information that is incomplete or misleading.
Response time is therefore only one possible measure.
The charity should also review whether answers are accurate, whether the agent follows its instructions and whether it recognises situations requiring human involvement.
For example, a knowledge agent might be assessed on whether employees can find approved information more easily.
But the organisation should also check whether the information is correct and whether employees understand its limitations.
Review escalation performance
Escalation is not necessarily a failure.
In many circumstances, referring an enquiry to a person is the correct outcome.
The charity should therefore avoid setting performance targets that encourage the agent to resolve every interaction independently.
Instead, it should assess whether enquiries are handled at the appropriate level.
Are routine requests completed effectively?
Are sensitive situations escalated?
Are employees receiving enough information to continue the conversation?
Are important issues being missed?
These measures provide a more useful picture of responsible performance.
Monitor unexpected behaviour
The charity should establish a way for employees to report incorrect responses, inappropriate actions or other concerns.
Reports should be investigated and used to improve the configuration, source information or processes.
Where a serious risk is identified, the organisation should be able to restrict or suspend the relevant agent.
Monitoring should also continue after updates.
Changes to knowledge sources, Salesforce configuration, permissions or connected applications may affect agent behaviour.
Provide ongoing staff training
Employees need to understand what Agentforce can and cannot do.
They should know when to rely on its assistance, when to check information and when to take over.
Training should also explain how to report concerns and why sensitive information must be handled appropriately.
At Sweet Potato Tec, our Managed Services include Agentforce guidance, ongoing optimisation, user support and regular platform reviews.
This reflects our wider approach to Salesforce: a system should continue improving after it goes live.
For charities, that ongoing responsibility is particularly important when AI is involved.
What should charities do before expanding Agentforce?
Before introducing additional agents, the organisation should review whether the initial use case has delivered the intended benefit without creating unacceptable risks.
It should consider whether the agent is accurate, appropriately supervised and trusted by the employees using it.
The charity should also review whether its data, permissions and governance arrangements can support a more complex use case.
An agent that successfully answers routine internal questions does not automatically establish that the organisation is ready to automate beneficiary-facing processes.
Each new use case requires its own assessment.
FAQs
Is Salesforce Agentforce safe for charities?
Agentforce provides security and governance capabilities that can support responsible use, but safety depends on how the charity designs, configures, tests and monitors its agents. Organisations need appropriate permissions, reliable information, defined escalation procedures and human oversight. Using Agentforce does not remove the charity’s responsibilities for privacy, safeguarding or compliance.
Can Agentforce access confidential donor or beneficiary information?
Agentforce can be configured to work with permitted information, depending on the agent’s identity, permissions, connected data sources and available actions. Charities should restrict access to what is necessary for the approved task. Sensitive beneficiary and safeguarding information should not be made available simply because it exists in Salesforce.
Can charities use Agentforce for safeguarding decisions?
Charities should not rely on an AI agent to make independent safeguarding decisions. Agentforce may potentially support authorised employees with limited administrative tasks or retrieval of approved guidance, but safeguarding assessments, escalation and accountability must remain with appropriately trained people and established organisational procedures.
Can Agentforce provide incorrect information?
Yes. Like other generative AI systems, Agentforce can produce inaccurate or unsupported responses. Reliable source information, clear boundaries, testing, monitoring and human review can reduce the risk, but they cannot justify assuming that every response will be correct.
Do charities need an AI policy before using Agentforce?
A documented AI policy is strongly advisable. It helps establish which uses are approved, who is accountable, how information is protected and what happens when something goes wrong. The charity should also consider existing data-protection, safeguarding and information-security policies, rather than treating AI governance as a completely separate activity.
Should a small charity use Agentforce?
Possibly, if it has a clear use case, suitable Salesforce foundations and the resources to manage the agent responsibly. A smaller charity may benefit from an internal knowledge assistant or another limited administrative use case. However, a simpler Salesforce automation or process improvement may sometimes be more appropriate. The decision should be based on practical value, risk and ongoing capacity.
Introduce Agentforce with confidence, not unnecessary risk
Agentforce offers charities opportunities to reduce repetitive work, help employees find information and improve how routine enquiries are managed.
But the potential benefits need to be considered alongside the responsibilities charities have towards supporters, beneficiaries, volunteers and employees.
Safe adoption starts with choosing the right use case.
It depends on reliable data, restricted access, appropriate testing and clear rules about when a person must take over.
It also requires ongoing monitoring and accountability.
At Sweet Potato Tec, we approach Agentforce as part of a wider Salesforce and organisational improvement strategy.
Our Agentforce specialists help organisations identify appropriate opportunities, define the scope of AI agents, configure safeguards, test their behaviour and support employees during adoption.
For charities, our nonprofit Salesforce expertise ensures those decisions are considered in the context of fundraising, service delivery, sensitive information and limited resources.
If your charity is exploring Agentforce but is unsure where to begin, our Agentforce Quickstart provides a practical starting point for assessing readiness and identifying suitable opportunities.
We can help you understand what AI could achieve, which foundations need attention and where human oversight must remain central.
The aim is not to introduce AI everywhere.
It is to use it where it can genuinely support your people and your mission, without compromising the trust placed in your organisation.