# AI instructions **Where to find it:** Settings → AI → Instructions **Who can access it:** Admins *** **What it is** Startdeliver's AI model covers the fundamentals of Customer Success out of the box — the most common assessment patterns, health signals, risk indicators, and recommended actions that apply across most CS teams and most customer bases. For the vast majority of your customers, the AI will assess and act correctly without any additional input. AI instructions cover the remaining edge cases. The things that are specific to your business, your product, your customer segments, and your way of working that the AI couldn't know from the data alone. Think of it like briefing a highly experienced new CSM. You don't teach them what churn risk looks like or how to run a Business Review — they already know that. You tell them the things unique to your company: which customer segment has unusual behaviour, which product configuration signals something important, which stakeholder change should trigger an immediate response. That's what AI instructions are for. *** **What to include** Instructions should cover the specific, the niche, and the contextual. If the AI would handle it correctly without any guidance, don't include it. If it requires knowledge of your specific business to get right, add it. **Industry or segment-specific exceptions** When a certain type of customer behaves differently from the norm — in how they use your product, what they need from your team, or what signals mean something different for them. *Example: "Customers in the public sector do not use Product C. Low usage of Product C for these accounts is not a negative signal — do not flag it as a health risk."* *Example: "Customers in the healthcare segment have a slower onboarding cycle due to compliance requirements. Treat 90-day onboarding as standard for this segment rather than flagging it as delayed."* **Commercial thresholds that change the CS motion** When ARR, seat count, or contract value changes how your team operates — who gets involved, what level of attention is warranted, what the escalation path looks like. *Example: "For customers with ARR above 3M, there is a dedicated support person assigned. Always reference this in the context of support health — delayed tickets for these accounts should be escalated immediately, not treated the same as standard accounts."* *Example: "Customers on our Enterprise contract include professional services hours. If a customer hasn't used their PS hours 60 days before renewal, flag this as a risk and create a task for the CSM."* **Stakeholder changes that matter disproportionately** When a specific type of user change on the customer side has an outsized impact on the relationship — a key role changing hands, a champion leaving, a new executive arriving. *Example: "A change in user type for a Super Admin is high priority. This often signals an internal restructure or a change in ownership of the product. Always flag this immediately and create a task for the assigned CSM regardless of current health status."* *Example: "When the primary billing contact changes, notify the CSM and create a follow-up task. This sometimes precedes a contract renegotiation."* **Product-specific signals** When usage of a specific feature or product tells you something important about the customer's trajectory — positive or negative — that the standard health model wouldn't surface on its own. *Example: "Activation of the API integration is a strong positive signal — customers who integrate via API have 40% higher retention rates. When this happens, flag it as an expansion signal and suggest the CSM schedule a technical deep-dive."* *Example: "Customers who have never activated the reporting module are at higher churn risk at renewal. Flag this as a risk if it's still inactive 90 days before renewal."* **Your terminology and naming** If your internal language differs from standard CS terminology, or if you use specific names for products, tiers, or lifecycle stages that the AI should recognize and use correctly. *Example: "We refer to our top-tier customers as Platinum accounts, not Enterprise. When referencing these accounts in assessments and actions, use 'Platinum' not 'Enterprise'."* *** **What not to include** Don't replicate what the AI already handles well. Instructions about general churn risk, onboarding best practice, how to run a Business Review, or what good adoption looks like are already part of the AI's model. Adding them to your instructions won't improve the output and makes the instructions harder to maintain. The test: if this instruction would apply equally to any CS team using any SaaS product, you probably don't need it. If it's specific to your business, your customers, or your product, add it. *** **How instructions are applied** Instructions apply globally across your entire workspace — to every customer assessment, every Jecta action, and every AI-generated recommendation. They run on top of the standard model, adding your specific context to every decision the AI makes. The AI reads your instructions alongside all the customer data it has access to. When a situation matches one of your instructions, that instruction takes precedence over the default model behaviour. Everything else continues to run as normal. *** **Version history and change tracking** Every change to your AI instructions is tracked and versioned. When you save an update, the previous version is preserved — you can see exactly what changed, when it changed, and who changed it. This matters because AI instructions apply globally. A single change affects how the AI assesses and acts across your entire customer base. Version history gives you a safety net — if an instruction change produces unexpected results in assessments or Jecta actions, you can see precisely what changed and restore a previous version. Before making significant changes to your instructions, it's worth reviewing a few customer assessments first so you have a baseline. After saving, check the same customers again to see how the change affected the output. If something looks wrong, version history lets you identify and reverse the specific change quickly. *** **Access and permissions** AI instructions are restricted to admins only. Only team members with Admin access can view, edit, or save instructions. This is intentional. Because instructions apply globally and affect every AI assessment and Jecta action across the workspace, changes should be deliberate and controlled. If you need to update instructions based on input from your CS team, an admin should make the change — not individual CSMs. If you need to request a change to AI instructions, contact your Startdeliver admin. *** **Tips for writing good instructions** **Be specific and concrete** Vague instructions produce vague results. "Public sector customers are different" tells the AI nothing. "Public sector customers do not use Product C — low usage is not a risk signal for this segment" tells it exactly what to do. **Use if/then framing where helpful** Instructions written as conditions are easiest for the AI to apply consistently. "When X happens, do Y" or "If a customer is in segment Z, treat this signal differently." **Keep it focused** A few precise instructions outperform a long document of general guidance. Prioritize the situations that come up regularly and have the highest impact when the AI gets them wrong. **Review periodically** As your business evolves — new products, new segments, new commercial thresholds — your instructions should evolve too. A quarterly review keeps them accurate. **Test after updating** After adding an instruction, open a customer it applies to and check whether the assessment and recommended actions reflect the new context correctly. If not, refine the wording. *** **Where to go next** → [AI Assessments](https://docs.startdeliver.io/jecta-agent-and-ai/ai-assessments) — how the AI assesses customers \ → [AI data & context](https://docs.startdeliver.io/jecta-agent-and-ai/ai-data-and-context) — what signals the AI reads \ → [Jecta: AI Agent](https://docs.startdeliver.io/jecta-agent-and-ai/jecta-ai-agent) — how Jecta acts on assessments and instructions --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://docs.startdeliver.io/jecta-agent-and-ai/ai-instructions.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.