How to Brief an AI Agent Vendor: Eight Items Before the First Call
A vendor brief is a pack you can forward before the first call: scene, boundaries, samples, knowledge, APIs, entry, pilot gates, and handover constraints. It is not a rewrite of the 10 go/no-go questions (/blog/ai-agent-selection-checklist). Those decide whether to proceed; the brief decides what the vendor quotes against. Build vs buy is a separate fork (/blog/build-vs-buy-custom-agent). GeonAI estimates a first scene from attachments; /agents are capability references only (not a public trial of 363+ presets). Email [email protected].
One-sentence definition
A usable brief = all eight items either have an openable attachment or explicitly say “none / unknown, to be measured.” A persona prompt is not a brief. Without attachments the vendor fills gaps with their demo intents.
Eight items to write down (paste into the email)
| Item | What the brief must say | If missing, the vendor will |
|---|---|---|
| 1 Scene and Owner | One-sentence scene + business Owner / decision maker | Build a platform, or leave ownership vague |
| 2 Boundary tables | Do / do not / must confirm with a human | Let scope balloon after the demo |
| 3 Sample threads | Success path + hard cases, plus current handling | Treat generic presets as your intents |
| 4 Knowledge inventory | Sources, classification, updater, whether data may leave your env | Ingest blindly, or mis-quote cleanup |
| 5 System interfaces | Read/write, staging, auth | Underestimate integration |
| 6 Entry and peaks | Channels, languages, busy-hour volume or “unknown, to measure” | Design for ideal traffic |
| 7 Pilot acceptance | Harm metrics to sample in 30 days | Accept “can chat” |
| 8 Handover and constraints | Whether evals/rollback ship, data residency, whether you must hold the runtime | Lock-in or a compliance block after go-live |
How to put boundary tables in the contract: /blog/custom-agent-scope-boundary-in-contract. Who owns the work: /blog/custom-agent-project-owner-raci. What must ship: /blog/custom-agent-delivery-artifact-checklist.
Do not hand-wave items 2, 3, and 7
- Boundary tables: price changes, lead times, refunds, and access grants default to human-confirm or do-not—not “the model will judge”
- Sample threads: redaction is fine; if you have no logs, list 10 fictional hard cases as eval seeds and label them fictional
- Pilot acceptance: name missing citations, wrong promises, unauthorized writes, and missed handoff. Gates: /blog/custom-agent-30-day-pilot-metrics
Common omissions (quotes go wide)
Knowledge cleanup, ACL, integration, and canary releases are often missing from quotes (/blog/custom-agent-quote-underestimated-costs). If items 4 and 5 are vague, you will hit those lines later.
You do not need all eight on day one
- Minimum to meet: 1+2+3 (Owner, tables, samples)
- Enough to quote: add 4+5 (knowledge and APIs)
- Enough to sign a pilot: add 7+8 (acceptance and handover). If item 6 is unknown, write “to measure”—do not invent peak numbers
Anti-patterns (reschedule the call)
- Sending only a system prompt as the brief
- Treating 363+ preset pages as “we already have an intent table”
- Writing “hire a digital employee” without splitting steps (/blog/what-is-digital-employee-vs-custom-agent)
- Signing an NDA and still sending no samples—secrecy is not a substitute for hard cases
How to send this to GeonAI
Put the eight items in one email (mark gaps as “none / unknown, to measure”). [email protected], /pricing, or Live chat. We estimate whether a first scene can be accepted from attachments. We do not require all eight at once, and we do not treat site presets as your brief.
Frequently asked questions
How is this different from the 10 selection questions?
Those 10 decide whether to proceed. These eight are attachments for the first vendor call. Answering the 10 without a brief still leaves the vendor filling gaps from a demo.
Must all eight be ready before the first meeting?
No. 1+2+3 is enough to meet. 4+5 is enough to quote in good faith. 7+8 is enough to sign a pilot. Mark unknowns; do not leave blanks for the vendor to guess.
What if we have no real chat logs?
Use redacted fictional hard cases as eval seeds and label them fictional in the brief. With zero samples the vendor will map you onto their support demo.
Is a prompt enough?
No. A prompt sets tone, not boundary tables, API rights, or pilot harm metrics. See /blog/custom-agent-more-than-a-prompt.
Can an NDA replace samples?
No. An NDA covers secrecy, not the acceptance object. Send redacted samples. Your legal process is yours—do not skip item 3 because the NDA is unsigned.
Can the 363+ presets be brief attachments?
No. Presets are capability references—not your Owner, tables, samples, or APIs.