The referee for deals between agents.
When your agent negotiates or coordinates with another agent, PrefSift finds what works, and says why.
Settlesa late-night quote, a damage claim, a table for eight, a refund request
Your price list and policies, enforced on every customer’s AI, without showing your minimums.Your price list and policies, enforced on every customer’s AI: one reply, day or night, on the record, without ever showing your minimums.
First decision set up free · a person replies within 2 business days
Customer’s assistant · 11:40pm
“Saturday clean, nothing over $200.”
✓ Offered Saturday, $195
$160 is under your $180 minimum
Customer’s assistant · 7:05am
“Cracked table. Wants the full $800 back.”
✓ Offered a $500 refund or a replacement
Refunds capped at $500 by policy 7.2
Organizer’s assistant · 9:12am
“Table for 8 this weekend. One vegetarian, nobody over $60 a head.”
✓ Offered Fri 8pm, set menu $48
Saturday groups need $75 a head
Behind the scenes
- ⚡Efficient.Each request is checked against your rules in code, with no AI bargaining: milliseconds when your rules are settings.
- 🔒Confidential.Customers never see your $180 minimum, or how far you’d go on a claim.
- 🧾Auditable.Every answer is recorded with the rule that applied.
Already happening
Customers’ AI assistants are starting to call, chat and book.
AI callers: Sierra, which runs customer-service AI for big brands, saw AI callers start phoning its healthcare clients’ AI in 2025. In October 2026 it shipped a tool to spot callers who are AI. Sierra, Oct 2026 (opens in a new tab)
Every one of those requests needs an answer inside your rules. PrefSift checks each one.
Which sounds like you?
What it changes for your business.
For agents in one system or across companies.
Every customer gets the same remedy for the same claim.
Today
Bargaining agents ended the same damage claim two different ways in five runs, two of them with no deal at all.
With PrefSift
Your refund cap is a rule. PrefSift returns the same remedies every time, and records which rule applied, for when a customer disputes it.
How we know: PrefSift gave the same answer in five of five runs.
The stronger agent wins
Don’t lose to a customer with a stronger AI.
In Anthropic’s Project Deal, AI agents bought and sold for real people for a week. Agents on the stronger model got better prices, and the people on the losing side rated their deals just as fair.
Today
A customer’s assistant on a stronger model can talk your agent below your own minimum.
With PrefSift
Already answering customers with Muse for Small Business, Sierra, Intercom’s Fin or your own agent? Add PrefSift, and your minimums and policies are checked in code before it answers, so a stronger AI on the customer’s side can’t talk it below your own rules.
Already answering customers with Muse for Small Business, Sierra, Intercom’s Fin or your own agent? Add PrefSift.
- Sellers on Claude Opus 4.5 vs Claude Haiku 4.5
- +$2.68 for the same item
- Buyers on Claude Opus 4.5 vs Claude Haiku 4.5
- paid $2.45 less
- How fair people rated their deals (4 = fair to both)
- 4.05 with the stronger model and 4.06 with the weaker. The losing side didn’t notice.
In our 270 test haggles, 14 deals broke a limit a side had set, and in 11 of them that side’s agent ran the smaller model.
How PrefSift decides
Every side’s context in. One table of options. Every option checked.
1Each side sends its context
Messages, rules, listings, policies, as they are.
Seller’s agent
“Don’t go under $450. I’d rather they pick it up.”
Buyer’s agent
“Up to $480, and I really want the lock and lights.”
The listing
price: 420 · 460 · 500 handoff: pickup · shipped extras: none · lock, lights
2A model reads it
Into limits and preferences, each tied to what a side said.
Read by a model
LimitSeller: at least $450
LimitBuyer: at most $480
PreferenceSeller: pickup over shipping
PreferenceBuyer: lock and lights over bike only
3The solver decides
Checks every option against every limit and proves which are beaten.
Decision table · 12 options
4Each side gets its answer
With its own reasons. Never the other side’s limits.
The seller sees
✓ $460, pickup, with lock and lights
Offers under $450: under your $450.
The buyer sees
✓ $460, pickup, with lock and lights
$420 offers: don’t work for the seller.
The same steps run whether the sides send messages or rules as data; data skips the reading. The example is a recorded bike sale: the options and reasons are the ones PrefSift returned.
Head to head
Your agent bargaining, or your rules checked.
Your agent bargaining
…PrefSift
…- 2 sides sent their context
- A model read it into a decision table: 4 options, every limit and preference
- A solver checked every option
What you get
Auditable. Efficient. Confidential.
Every answer
On the record
Each decision is stored with the rule that applied, ready for a dispute or an audit. A bargaining agent leaves a transcript. PrefSift leaves a reason for every option.
47 ms
No AI cost when rules are settings
A cleaning quote decided from your settings: no model, no tokens. Bargaining agents took about 15 seconds for the same quote. Messages in words are read once, in seconds.
0
Times your minimum was shown
In 23 of 60 recorded runs, a bargaining agent said its own side’s limit out loud. With PrefSift, customers’ assistants see “doesn’t fit the business,” never your number.
How it reaches the answer
Agents talk through a few options. PrefSift checks every one.
Every option on the table, checked
Bargaining agents only ever raised 48% of the options on the table in our recorded runs. PrefSift checks all of them against every side’s limits, every time, in milliseconds once the rules are read.
Nobody worse off for nothing
It keeps only the options no one can improve on without costing someone else, the economist’s standard for an efficient deal. In our 270 test haggles, 8% of the deals agents reached failed it.
Fair by your policy, not by who pushed hardest
When several options fit, your policy picks, the same way for every customer, whatever their assistant says or how long it argues.
Conditions stay conditions
“Weekends only for deep cleans” or “if Saturday, then morning” is kept as a rule, not flattened into a score or a sentence another agent can argue with.
A language model reads what people say. The decision is math, and it shows its work.
Every number here comes from the recorded head-to-heads on this page (October 2026), every run counted.
How it works
Each side sees its own reasons. The other side never sees your limits.
PrefSift receives every side’s limits and shows them to no other side.
The same result, seen by each side of a cleaning quote:
The customer’s assistant sees
$230: crossed out, over your $200.
$195: kept.
Your booking agent sees
$230: crossed out, doesn’t fit the customer.
$195: kept. $15 above your $180 minimum.
Calls, sessions, permissions and what it won’t do
One call, or a session
One call: your agent already has every side’s limits and sends them together. If it’s missing something, PrefSift says what to ask the other side’s agent, even one that doesn’t use PrefSift.
A session, when each side should send its own:
- 1Your agent starts a decision and gets a code for each other side.
- 2It invites them. Each agent reads the docs, then joins or declines.
- 3Each side sends only its own limits, straight to PrefSift.
- 4Everyone gets the result at once, with the reasons they may see.
“Hollis & Pine invites you to settle ‘a damaged table claim’ with PrefSift, a neutral referee. Your limits go to PrefSift, not to the other participants, and it costs you nothing. To join, POST /v1/sessions/ses_8Kq2…/join with header X-Session-Code: ps_Rw4…; to say no, POST …/decline.”
Who sees what
Each decision has its own small permission list: who sees whose limits, whether reasons name them, and who gets the record.
- Limits
- Each side sees its own, by default. Or named people, or everyone.
- Reasons
- Name only the viewer’s own limits, or all of them.
- Raw inputs
- Only the side that sent them, unless you allow more.
- The record
- Whoever starts the decision and any observers, or everyone.
Where the options come from
You choose, per decision, and can mix them.
You send them
Your catalog, your calendar, your quotes
Found in the parties’ documents
Offer emails, PDFs, a listing
PrefSift searches
Open tables, flights, venues, with a spend cap
Your agent searches
With its own tools, then sends what it found
It won’t
- Pick a fairness rule. It applies the one your app declared.
- Show one side’s limits to another unless the decision allows it.
- Treat silence as yes. A missing fact stays open, with the question to ask.
- Book, pay, or message anyone.
Control
You set the rules, approve the close calls, and keep the record.
PrefSift never books, pays or messages anyone. It decides what fits, and your systems act on it.
Approvals
Set a margin, say within 3% of your minimum. A result inside it waits; your system gets a signed notice, and the person you named approves or rejects it.
Who sees what
Each decision has a short permission list. By default the customer sees which options work, never your numbers.
Each location’s own prices
Your system sends each store’s minimums and price list with its requests.
Observers
Finance, legal or a manager can read a decision and its record without taking part.
A record per decision
Your system can fetch it: which rule applied, what each side sent, who approved.
Any customer’s assistant
It can join with a code, or simply answer the questions PrefSift says to ask.
Questions
Who sees our floors and rules?
PrefSift and your own team. The customer’s assistant sees which options work, and “doesn’t fit the business” for the rest, never the number behind it.
What if it gets something wrong?
Set an approval margin, say within 3% of your minimum, and close calls wait for a person you name. Every decision keeps a record of which rule applied and the quote it came from, so you can see exactly why, and override it.
What do my customers see?
Their assistant gets the options that work for both of you, and “doesn’t fit the business” for the rest, never your numbers. A refund policy you publish can be quoted as usual.
Can it settle conflicts between our own agents?
Yes. If you run more than one agent, for sales, support or scheduling, and their answers clash with your finance or staffing rules, PrefSift checks every proposal against all of your rules the same way every time, and records which rule decided it.
How is this different from an AI negotiator for my store?
An AI negotiator works for you and bargains with the customer’s assistant, so the answer depends on how the conversation goes, and it has to know your minimum to defend it. PrefSift works for neither side. It checks the customer’s request against your rules and returns the options that fit, with a reason for each, the same way for every customer. Rules you send as data never reach an AI model, and nothing gets haggled down.
How does it connect to our systems, and who sets it up?
Your agent, or your booking system, calls one web address with what it already has: the open slots or prices, your rules, and what the customer’s assistant asked for. We set up your first decision with you, free, as a short consulting engagement, so you don’t need a developer on staff to try it.
What does it cost?
You pay per decision: nothing when your rules and the customer’s request arrive as settings, and typically 1 to 6¢ when PrefSift reads a few messages. Each account has a monthly spend cap, so the bill can’t run away. Setting up your first decision is free; help beyond that is quoted per business. Put another way: at 1 to 6¢ a decision, one quote $15 under your minimum pays for 250 or more.
How fast is it?
When your rules and the customer’s request arrive as settings, milliseconds, with no AI model involved. When PrefSift has to read a customer’s message, usually 5 to 17 seconds the first time; a message it has read before is reused in milliseconds.
What happens to our data?
It’s used only to make the decision and keep its record. Reading messages uses OpenAI’s API, which doesn’t train on API data by default; rules sent as settings never reach a model. Retention, storage location and a data processing agreement are set per deployment. We don’t have a SOC 2 report or a HIPAA business associate agreement yet, so don’t send health details as free text.
We set up your first decision free.
- Tell us one decisionA quote, a refund, a group booking. A person replies within two business days.
- We set up your rules, freeWe map that decision with you and turn your prices and policies into settings.
- Your agent checks every replyAt 1 to 6¢ a decision, one quote $15 under your minimum pays for 250 or more.