How to Write an RFP for AI Agents
A good AI-agent RFP scores vendors on what a buyer can actually verify — per-outcome pricing, an independent quality score with a confidence interval, an SLA, and data-handling terms — instead of demos and logos.
A request for proposals exists to make vendors comparable on terms you can check — not to collect marketing decks. AI agents make this harder than traditional software, because the thing you are buying is work getting done, and most vendors price and pitch around that fact. This post is a buyer’s framework for an AI-agent RFP: how to define the unit of work, what pricing to require, how to demand verifiable quality, and what to put in the SLA and data-handling sections. It contains no benchmark numbers, because the numbers belong in the vendors’ responses and on the live index.
Define the outcome you are buying
Start the RFP by naming the unit of work in plain terms: a resolved support ticket, a reviewed pull request, a processed invoice, a booked qualified meeting. This single decision drives everything downstream — it makes vendor quotes comparable, defines what "quality" means, and gives the SLA something concrete to measure. An RFP that asks for "an AI support solution" gets back incomparable decks; one that asks for "price and quality per resolved ticket" gets back numbers.
Require per-outcome pricing
Ask every vendor to express price as dollars-per-completed-outcome, and to show their work converting from whatever they bill on (per seat, per token, per action, per month). This is the only way a "$2,000/month" plan and a "$0.99 per resolution" plan land on the same axis. Require them to state assumptions — expected volume, what counts as a completed outcome — so you can compare like for like and re-run the math on your own numbers.
Anchor the responses against a baseline. The public human-labor cost for the role (US BLS OEWS medians for the relevant SOC code) is a defensible floor; as the WorkForce Labor Index publishes a verified per-task rate for the category, that becomes the market anchor. A vendor charging well above the human baseline should be made to justify it with quality, not adjectives.
Demand an independent quality score
The hardest part of an AI-agent RFP is verifying quality, because self-reported metrics — run on tasks the vendor chose, graded by criteria the vendor set — are unverifiable. Require a third-party quality score produced on a sealed, independent task bank, reported with a confidence interval and sample size. An AQO score is one such credential. A score of 84 with a tight interval is a real claim; a bare number with no interval is a red flag. Make "no independent score, no shortlist" a stated rule.
Specify the SLA in outcome terms
Tie the service-level agreement to the unit of work, not to uptime alone. Useful clauses: a floor on resolution or acceptance quality, an escalation/handoff path when the agent cannot complete the outcome, latency targets per outcome, and remedies (credits) when the floor is missed. Because you defined the outcome up front, each of these is measurable rather than aspirational.
Pin down data handling
AI agents touch your customer data, so the data-handling section is not boilerplate. Require explicit answers on: where data is processed and stored, whether your data (or prompts and outputs) is used to train models, retention and deletion terms, sub-processors and model providers in the chain, and applicable compliance posture (SOC 2, GDPR, sector rules). Vague answers here are themselves a signal.
Score it on a single sheet
Reduce every response to one comparison sheet: vendor, dollars-per-outcome, independent quality score with its confidence interval, SLA floor, and data-handling posture. With price anchored and quality measured, you choose on quality-per-dollar against your defined outcome — not on which deck was glossiest. The best vendor is rarely the cheapest or the most expensive; it is the one whose quality premium is worth its price premium.
Methodology note
This article is a procurement framework and deliberately states no benchmark rate, score, or confidence interval — those belong in vendor responses and on the live index. For the human-labor floor, cite the relevant US BLS OEWS median by SOC code (public domain). WLI category rates publish with their confidence intervals and sample sizes once transaction coverage exists; see /methodology for how the index is calculated and governed.
Next steps
Require a free, independent quality score from every vendor at /eval, and anchor pricing against the live per-task rates at /wli as categories cross the publication threshold.