What problem does it solve?
Large, strategic categories usually get a proper sourcing event. The tail does not: maintenance parts, facilities services, one off services and repeat buys that each look small. A buyer who sources one of them by hand writes the requirements, finds vendors, sends the request by email, chases answers, copies bids into a spreadsheet and writes the award. When buyers are short of time, a small request can end up with a known vendor without being competed.
The case studies on this page describe the same starting point. Fairmarkit's account of Snowflake says its procurement team of about 12 people ran sourcing largely by email, with an average cycle of about 25 days. Its account of G42 describes an average cycle of 5 to 7 days for a team of two managers and four buyers, with requesters unsure how to source what they needed.
A bp director describes sourcing and contract negotiations as a manual process that takes a great deal of time and people, and Fairmarkit's case study adds that bp's organization had shrunk by 25% over 18 months while the same volume of work remained. Public buyers add a rule on top: the MBTA case study describes bidding regulations under which two bids are received before an order is finalized.
How does it work?
- Take the request in. A requester describes what they need in a form or a chat, or a requisition arrives from the ERP. The agent turns it into a structured request: item or scope, quantity, delivery date and the attachments a vendor needs.
- Check it against the rules. Policy decides whether the request may be sourced automatically: value limit, category, approved or required supplier lists, number of bids needed. Anything outside the rules goes to a buyer as a normal sourcing task.
- Pick vendors and send the RFQ. The agent selects vendors from the supplier master and the history of past bids, and sends the request to several of them.
- Collect and compare the bids. Quotes come back by email or portal and are extracted into one comparison on price, delivery and stated terms. The agent chases vendors that have not answered.
- Award or recommend. Within set limits the agent awards to the best bid and notifies the winner and the others. Above the limits, or when bids are unclear, it shows the comparison and a recommendation to a buyer, who decides.
- Record and learn. The award, the bids and the savings against the earlier price are stored with the request, and the data feeds the next vendor selection and the spend reports.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Email, Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cycle time reduction | Too few to pool | 50% to 75% | 2 | 1 organization, 1 vendor |
| Automation rate | Too few to pool | 31.8% | 1 | 1 vendor |
| Cycle time | Not pooled | 6.4 days | 1 | 1 vendor |
Value drivers: Speed and cycle time, Employee productivity, Lower cost to serve, Compliance quality.
Indicative value
A manufacturer with 6,000 small sourcing requests a year
USD 60,000 to USD 1.1 million
Buyer time and price savings from automated RFQs per year
How this is calculated
Formula: requests * automatedShare * (hoursSaved * hourlyCost + averageSpend * savingsRate). The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Small sourcing requests per year requests, requests per year | 6,000 | 6,000 | The reference organization. Replace with your own count of requisitions below your sourcing threshold. |
| Share of requests the agent sources without a buyer automatedShare, fraction of requests | 0.1 | 0.3 | At or below the one share reported on this page (Fairmarkit reports 31.8% of Snowflake's sourcing events running autonomously). That is a single vendor reported result, and the share you reach depends on the rules you set. |
| Buyer hours saved per automated request hoursSaved, hours per request | 1 | 3 | Editorial assumption for writing, sending, chasing and comparing a small RFQ by hand. Replace with a time study from your own buyers. |
| Fully loaded buyer hour hourlyCost, USD per hour | 40 | 70 | Editorial assumption. |
| Average spend per request averageSpend, USD per request | 3,000 | 8,000 | Editorial assumption for tail spend, replace with the average from your own requisitions. |
| Price improvement from competing the request savingsRate, fraction of spend | 0.02 | 0.05 | Editorial assumption, deliberately low. Fairmarkit reports an additional 5% of awarded savings at Snowflake, but gives no baseline for it, so the range here is set lower. Replace with the savings you can show on your own awarded events. |
What it leaves out: Counts buyer time and a flat price improvement on automated requests only. It leaves out the cost of the platform and its set up, the effort of onboarding vendors, savings that would have happened anyway through an ordinary quote, compliance benefits, and the value of buyers moving to strategic categories.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
G42
United Arab Emirates · Technology and software · 2025
G42's procurement team of two managers and four buyers had an average sourcing cycle of 5 to 7 days, with requesters unsure how to source what they needed. It adopted Fairmarkit, whose intake tool lets a requester photograph an item on a phone. Fairmarkit's case study says an AI agent then creates the requisition, sources suppliers, gathers bids and awards contracts, and reports that the average cycle fell to 1 to 2 days.
- Cycle time reduction: 75%
"75% Reduction in Cycle Times: The average procurement cycle went from 5-7 days to just 1-2 days."
Claimed by: vendor
Snowflake
United States · Technology and software · 2025
Snowflake's procurement team of about 12 people sourced largely by email, with an average sourcing cycle of about 25 days. A new procurement leadership team switched on Fairmarkit, including its auto send and auto award features, an intake assistant and a scope of work generator. Fairmarkit reports the average cycle fell to 6.4 days and that 31.8% of sourcing events now run autonomously.
- Cycle time: 6.4 days
"Cycle times dropped dramatically, from 25 days to 6.4 days."
Claimed by: vendor - Automation rate: 31.8%
"31.8% of sourcing events now run autonomously through Fairmarkit, up nearly 15x from the previous year."
Claimed by: vendor
bp
United Kingdom · Energy and utilities · 2021
bp's procurement team replaced a manual sourcing process of sending requirements to suppliers, collecting bids and selecting a vendor with Fairmarkit, after the organization shrank by 25% and still had to deliver the same volume of work. According to Fairmarkit's case study, bp put $1.2B of spend through the platform in 18 months and sources $100M through automated processes with zero human intervention. A bp director says cycle time has come down by about half.
- Cycle time reduction: about 50%
"So for example, we have taken our cycle time down by about half and what that means in real terms is our stakeholders get their work and their product and services about 50% quicker."
Claimed by: organization
Massachusetts Bay Transportation Authority (MBTA)
United States · Logistics and transportation · 2019
The Massachusetts Bay Transportation Authority moved its small purchases from a paper process to Fairmarkit in 2017. Buyers upload a request, the platform sends the RFQ to several vendors in seconds and lays the bids side by side, and the buyer chooses the award. Fairmarkit reports more than 14,000 RFQs and 36,000 bids since the start and $10M of cost avoidance. The MBTA's Chief Transformation Officer says he paid for the system in the first month.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Requisitions with item descriptions, quantities and delivery dates
- A supplier master with categories, contacts and approval status
- History of past quotes and awards per category
- Written sourcing policy with value limits, bid counts and approved supplier rules
Systems to integrate
- ERP or purchasing system for requisitions, purchase orders and supplier master
- Procurement suite or intake tool where requesters submit needs
- Email or supplier portal for RFQs and bids
- Contract repository, to check for existing agreements before going to market
Complexity: Medium
The agent needs requisitions and supplier data from the ERP or procurement suite, a written rule set for which purchases may be sourced and awarded without a buyer, and vendors who answer through a channel the system can read. Awarding is a commercial commitment, so approval thresholds and a purchase order handoff have to be agreed with finance before anything runs autonomously.
- 1
Choose the first categories
Take a spend cube and pick two or three tail categories with many repeat requests, clear specifications and several known vendors, such as maintenance parts or facilities services. Leave engineered items and sole source purchases out of the first wave.
- 2
Write the award rules
Decide per category the value limit for automatic award, the number of bids required, the approved and excluded vendors and the tie break between price and delivery. Have finance and legal sign the rules before the agent can use them.
- 3
Start with recommend only
Run the agent so that it sends RFQs and compares bids but a buyer clicks the award. Compare its recommendation with what buyers would have chosen, on a sample of at least a few hundred events.
- 4
Switch on auto send, then auto award
Allow the agent to send requests without review once the requests it drafts are accepted almost every time. Allow automatic award only for the categories and limits where the comparison has matched the buyers' choice.
- 5
Open the supplier base up
Add vendors to each category deliberately, including local and diverse vendors where policy calls for it, and measure how many bids each request receives.
- 6
Report savings against a baseline
Fix how savings are calculated before launch, for example against the last paid price or the median bid, and report them per category so finance accepts the number.
Guardrails
- Automatic award only inside written value limits, categories and vendor rules, with everything else routed to a buyer
- Every RFQ, bid, comparison and award stored with the version of the rules that applied
- Vendor communication from approved templates, with no price or volume commitment made by the agent beyond the award rule
- Sanctions, conflict of interest and vendor onboarding checks run before a vendor can win an award
- A named buyer who owns each category and can switch the autonomous mode off
KPIs to instrument
- Share of sourcing events that run without a buyer touch, by category and value band
- Days from request to award, split between automated and manual events
- Bids received per RFQ and the share of RFQs with fewer than the required number
- Price improvement against the baseline, with the calculation method stated
- Awards a buyer reverses or a requester disputes
Human in the loop
Buyers own the rules, review every recommendation above the award limit and sample automated awards each month. They handle vendor disputes, sole source requests and any request where the requirements are unclear, and decide when a category moves between recommend only and autonomous.
Common failure modes
- Savings that are really a lower baseline
- Savings are measured against an inflated reference price, so the number looks large while the paid price barely moves. Fix the baseline method before launch and audit a sample.
- Cheapest bid wins a poor specification
- The request was vague, vendors quoted different things, and the lowest price won. Check specification completeness at intake and send unclear requests to a buyer.
- Vendors stop answering
- Automated requests arrive in volume and vendors learn that most are ignored or low value. Track answer rates, feedback to losing bidders and the number of repeat requests per vendor.
- Split purchases to dodge the limit
- Requesters divide a large need into several small ones to stay under the automatic award limit. Monitor repeat requests for the same item or vendor and aggregate them.
What are the risks and rules?
EU AI Act
Minimal risk
Sourcing indirect goods and services for the organization's own use is not listed in Annex III, and the system has no direct effect on individuals, so only AI literacy duties apply. Check the design if bidders are natural persons, for example sole traders, or if the agent produces any assessment of individual workers. Public bodies also remain bound by national public procurement rules on equal treatment and transparency, whichever tool runs the event.
Rules that apply
Guidance
- AI Risk Management Framework (National Institute of Standards and Technology, North America). Voluntary framework, with a generative AI profile, for governing and measuring the risks of an AI system such as an agent that commits the organization to purchases.
- PPN 017: Improving transparency of AI use in procurement (Cabinet Office, Europe). UK note for central government departments, their executive agencies and non departmental public bodies (other contracting authorities may apply it), mainly on suppliers' use of AI in tenders. It says decisions should be made with the support of AI systems and without reliance on them, and that authorities must not discriminate against suppliers when asking about AI use.
Controls to put in place
- Written award rules with an owner, a version and a change log
- Immutable record of each RFQ, bid, comparison and award
- Separation of duties, so the agent cannot create a vendor, change bank details or approve its own limit
- Monthly sample review of automated awards by a buyer who did not set the rules
- Vendor communication logged and available for audit
Frequently asked questions
- How much of the sourcing workload can be automated?
It depends on the rules you allow. Fairmarkit reports that 31.8% of Snowflake's sourcing events now run autonomously, up nearly 15 times from the year before, and that bp sources $100M through automated processes with zero human intervention. The bp page describes automation and does not mention AI. Snowflake's page says it is expanding to more categories, so start with a recommend only phase and measure your own share.
- Is this the same as electronic sourcing software?
Electronic sourcing software gives buyers a place to run events. The autonomous version also writes and sends the request, compares the bids and, within limits, makes the award, so the buyer sets rules instead of running each event. Fairmarkit describes auto send and auto award as separate features at Snowflake, and Fairmarkit's case study on G42 describes an AI agent that creates the requisition, sources suppliers, gathers bids and awards contracts.
- Who is accountable when the agent awards to the wrong vendor?
The organization is, the same as for a buyer's decision. That is why automatic award sits inside written limits, why every award is logged with the rule version, and why a named buyer owns each category and a monthly sample of automated awards is reviewed by someone who did not set the rules.
- What do these programs report on cycle time?
Fairmarkit reports that Snowflake's average cycle fell from about 25 days to 6.4 days and that G42's fell from 5 to 7 days to 1 to 2 days. A bp director says cycle time is down by about half. All three come from Fairmarkit's own case study pages, and each organization counts cycle time its own way.
How to cite this page
Blits.ai AI Use Case Library, "AI agent for autonomous RFQ and tail spend sourcing", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/autonomous-rfq-and-tail-spend-sourcing. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published