The Content Valet Playbook

The Content Valet is only as useful as the information it can access and the controls governing how it uses that information. A general-purpose AI model may generate fluent answers without knowing whether a programme's latest figures have changed, whether a project has ended or whether a statement is approved for public use. For a development organisation, these are not minor errors. They can affect donor confidence, public understanding and the reputation of programme partners. The solution is to build a managed knowledge base and use a retrieval-augmented generation (RAG) approach where appropriate.

Why is it important?

RAG combines information retrieval with generative AI. When a visitor asks a question, the system retrieves relevant material from an approved information source and uses that material to help formulate an answer. For example, if someone asks, “What has this climate resilience programme achieved?”, the Content Valet could retrieve the programme's latest results report, a relevant project update and a published evaluation. It could then summarise the findings and link to those original documents.

RAG can improve the relevance and traceability of responses, but it does not guarantee accuracy or eliminate hallucinations. The system can retrieve the wrong passage, misunderstand a document or generate a claim that the evidence does not support. AI systems also face security risks, including prompt injection, which must be addressed through broader application controls.

Build the knowledge base in stages

1. Audit existing content. Gather approved website pages, annual reports, FAQs, policy briefs, programme documents, videos, impact stories and relevant datasets.
2. Assign ownership. Identify the team or individual responsible for approving each content category and confirming whether information remains current.
3. Organise and label the material. Tag assets by topic, country, programme, audience, format, publication date, approval status and review date.
4. Prioritise authoritative sources. Establish which documents take precedence when figures or descriptions conflict. A current approved results report, for example, should normally take precedence over an older campaign page.
5. Connect the approved sources. Configure retrieval so that the AI can find relevant passages, retain source links and distinguish published evidence from missing or uncertain information.
6. Test before launch. Check whether the system retrieves the correct material, uses figures in context, acknowledges uncertainty and refuses to invent unsupported claims.

A vector database may be useful for semantic retrieval, but it is not a requirement for every project. Smaller organisations may be able to begin with a well-maintained document repository, website search and a carefully scoped AI service. The appropriate architecture depends on content volume, security requirements, budget and the complexity of the use case.

Make every important answer traceable

For development communications, the ability to verify a claim is as important as the ability to summarise it. Where possible, the Content Valet should provide:

For example:

“The programme's latest published report records progress against its agricultural resilience targets. You can review the reported results and reporting period in the programme report. I cannot confirm a more recent figure from the information currently available.”

This example is illustrative. A live system should link to the specific report supporting its answer, not to a generic publications page.

Establish response guardrails

The Content Valet should have explicit rules for questions involving sensitive information, unverified claims, political issues, financial transparency and personal circumstances.

It should:

These measures should be reinforced by technical safeguards, usage permissions, monitoring and human review, rather than relying on a prompt alone.

A visitor who finds a relevant story may want to learn more, share it, attend an event, contact a programme team or contribute financially. The Content Valet should make those options easier to discover without assuming that everyone has the same motivation or capacity to act.

Match the next step to the visitor's needs

Consider a visitor reading about a programme supporting pastoralist communities in East Africa. The Content Valet might offer:

Each recommendation should have a clear connection to the content the visitor has chosen to explore. Avoid generic appeals that interrupt the visitor's journey or imply that a donation guarantees a specific result unless the organisation can substantiate that claim.

Make donation journeys secure and transparent

In-chat giving may be technically possible through a supported payment integration, but it is not automatically the safest or simplest option.For many organisations, a better starting point is to direct visitors to an established, secure donation page. If a payment provider supports an embedded flow, the organisation should assess the security, accessibility, privacy, transaction fees and maintenance implications before deployment. The Content Valet should never request card details, mobile-money PINs, passwords or one-time authentication codes in ordinary chat messages. It should not claim that a payment has succeeded until the payment system confirms the transaction.

For a Kenyan audience, for example, a donation journey might direct visitors to an organisation's verified M-Pesa payment instructions or an approved and secure card-payment service. The payment method, fees, currency and any applicable restrictions should be clearly communicated.

Encourage sharing without making it compulsory

Not everyone can donate, but some visitors may be willing to share useful information with their networks. After someone finishes reading a story, the Content Valet could ask:

“Would you like a short summary of this story to share with your network?”

The assistant can draft a caption or suggest a visual using approved content. It should preserve the original meaning, avoid overstating impact and provide a link to the source story. If the organisation uses photographs, video or personal stories from communities, confirm that permissions cover the intended reuse. Do not create AI-generated depictions of real beneficiaries that could mislead audiences into believing they are genuine documentary images.

Personalise responsibly

Personalisation can improve content discovery, but it should not come at the expense of privacy. If someone explores climate adaptation resources, the assistant can recommend related reports during that session. Automatically attaching an inferred interest to a named donor's CRM profile is a different decision and requires an appropriate legal basis, transparent notice and suitable controls.

Before syncing conversational data with a customer relationship management system, decide what information is genuinely needed, why it is being collected, who can access it and how long it will be retained. For Kenyan organisations for examples, these decisions should be reviewed against the Data Protection Act, 2019 and guidance from the Office of the Data Protection Commissioner. Where other countries are involved, applicable local laws must also be considered.

A Content Valet operates at the intersection of information, technology and human relationships. Its design should therefore reflect the standards expected of a responsible development organisation.

Be transparent about AI

The welcome message should make it clear that visitors are interacting with an AI assistant. A suggested opening is:

“Hello! I'm Julia, your personal AI guide to our organisation's work. I can help you find stories, programme information and published results. I'll link to original sources where possible, and you can contact our team if you need further assistance. What would you like to explore?”

The assistant should also explain when it cannot answer a question confidently. Transparency helps visitors make informed decisions about whether to rely on its response.

Protect personal information

Before launch, establish a written data-handling policy that covers:

Collect the minimum information necessary. Do not ask visitors to disclose personal circumstances simply to access public information. Sensitive conversations should not automatically be copied into a marketing database. The ODPC's published guidance provides additional resources on matters including consent, data protection policies, children's data and data protection impact assessments.

Design for accessibility

The chatbot should work for people using keyboards, screen readers, mobile devices and assistive technologies. It should not become the only route to essential information. Use the Web Content Accessibility Guidelines (WCAG) 2.2 as a technical reference.

Practical measures include:

Test with people who have different access needs rather than relying exclusively on automated accessibility tools.

Protect children and vulnerable people

If the organisation works with children, displaced people, survivors of violence or communities affected by emergencies, additional safeguards are essential. Do not encourage people to share identifying details, precise locations or sensitive experiences through an open-ended AI conversation unless there is a clearly justified, appropriately protected process. Avoid generating case-specific advice about health, legal status, safety or eligibility for assistance.

The UNICEF guidance on AI and children's rights is a useful reference for organisations whose work involves children. A Content Valet should not be presented as an emergency-response service or a substitute for trained safeguarding personnel. Establish clear referral procedures and verified contact details before deployment.

Keep human support always available

Some questions should never be handled solely by an automated assistant. Examples include safeguarding disclosures, complaints about staff or partners, allegations of fraud, requests involving confidential information and questions about individual eligibility for assistance.

Create a clear escalation process. Where the organisation cannot provide immediate human support, the assistant should state this honestly and offer a verified contact route or the next appropriate step.

A successful implementation does not need to begin with an expensive, fully integrated AI platform. A smaller, well-defined pilot can reveal whether conversational content discovery solves a genuine problem before the organisation commits more resources.

Step 1: Establish the baseline

Review the existing website before introducing the Content Valet. Record which pages attract visitors, which resources are frequently accessed, where visitors leave, and which actions people complete. Examine search queries, common questions received by communications teams and feedback from programme staff.

Set baseline measures before launch so that you can assess whether the new experience improves on what already exists. Website analytics can help organisations understand traffic and completed actions, but they should be interpreted alongside qualitative feedback and the wider communications context (Google for Nonprofits).

Step 2: Choose a focused pilot

Select one programme or topic with a clear audience need and a reasonably well-maintained collection of public information. For example, an organisation working on climate resilience might start with a Content Valet that helps visitors explore programme objectives, verified results, community stories and published reports.

Avoid starting with every programme, country office, donor and stakeholder group at once. A focused pilot is easier to govern, test and improve.

Step 3: Define the technical scope

Work with the communications, IT, data protection and programme teams to decide:

Use the OWASP GenAI Security Project as a reference for identifying risks in applications built with large language models.

Step 4: Design the conversation

Keep the experience simple. Give visitors a few clear starting points, such as:

Allow free-text questions as well, but do not force visitors to use a conversational interface if the information is already available through a straightforward page or search function.

Step 5: Connect systems selectively

Integrate only the systems needed for the pilot. These may include the website content repository, an analytics platform, a contact form or a secure donation service. Each integration introduces additional technical, privacy and maintenance considerations. CRM synchronisation, for example, should not be enabled simply because the technology makes it possible. Establish a clear purpose and appropriate safeguards first.

Step 6: Test, approve and launch

Before public launch, have staff from communications, programmes, IT and relevant safeguarding or data protection functions test the assistant. Include real-world questions, ambiguous requests, conflicting figures, outdated documents, unsupported claims, attempts to obtain restricted information and situations requiring human assistance. Document the results, correct identified weaknesses and agree on who has authority to approve the launch.

The stages are a practical application of the Manjano Path, not a rigid sequence. Audience research, content development, testing, and evaluation should always inform one another throughout implementation. This approach also reinforces a fundamental distinction: AI is the delivery mechanism, while the Manjano Path is the communications approach. You should always evaluate audience needs, content priorities, and desired outcomes before deciding which technology to adopt.

A Content Valet should be evaluated as a communications intervention, not simply as a technology deployment. A high number of chatbot conversations does not necessarily mean people understand the organisation better. Equally, a visitor who finds the information they need quickly may leave without browsing multiple pages. That can be a successful outcome. Use a balanced set of measures, for example:

Define each measure before launch. For example, decide what counts as a successful content discovery, how a donation will be attributed and how frequently response accuracy will be sampled. Where feasible, compare the pilot with a pre-launch baseline or a suitable alternative journey. Consider seasonal campaigns, changes in traffic and other factors that could influence results. Avoid attributing every change to the chatbot without supporting evidence. Review the findings with programme and communications teams, and use them to decide whether to improve, expand or discontinue the pilot.

The following schedule is an indicative starting point. Actual timelines depend on procurement, technical capacity, content quality and organisational approvals.

Days 1 to 30: Discover and map: Research audience needs, audit the website, select a priority programme, establish baseline measures and define the pilot's scope.

Days 31 to 60: Clarify and build: Clean and organise source content, configure the assistant, develop conversation flows, review privacy and security, and test with staff.

Days 61 to 90: Connect, learn and improve: Launch a controlled pilot, monitor performance, collect feedback, fix problems and prepare an evidence-based recommendation on next steps.

If content is unreliable, privacy safeguards are incomplete, or critical errors remain unresolved, delay the launch.

Check out our Content Valet prompt sequence, which is a test guide for configuring an AI assistant. It should always be adapted to the chosen platform and supported by technical controls, approved knowledge sources and testing. A prompt alone cannot enforce every requirement. Access controls, retrieval permissions, output validation, secure integrations, logging policies and human oversight must support the instructions.

The Content Valet offers non-profit and development organisations an opportunity to rethink how audiences encounter their work. Instead of asking visitors to make sense of a complex website on their own, organisations can offer a guided experience that connects questions with credible information and meaningful opportunities for engagement.
The Manjano Path keeps this contained in communications practice where the technology should serve that process, not dictate it.
A successful Content Valet will not necessarily be the one with the most sophisticated model, the longest conversations or the highest number of donation prompts. It will be the one that helps people find what they need, understand what the evidence says, make informed choices and engage with the organisation on terms that respect their time, privacy and agency.
For a development communications team, that is a more useful measure of progress than a non-functional chatbot installation.