02 Research
What I know, and what we still need to test.
What the evidence says, what I still bet on, and how we test the 3 bets that matter most before an agent orders alone.
How the research works
Evidence first. Then the bets it cannot settle. Then the most important tests, until the pilot proves it.
What I know
From published research, so far on the buyer side only. Everything else is still a bet, listed below.
The buyer side
61% prefer an overall rep-free experience. Buyers still want a rep’s input when checking whether a product fits their company’s needs.
Source: Gartner sales survey, published June 2025, 632 buyers
A sales rep is 1 tap away on every order, and decides every part the rules cannot confirm
As reported by the source. I have not re-run the study, and it may not hold for every merchant. Checked 22 Oct 2026.
No formal interviews yet. I made 3 informal calls to workshop managers in Mohali, India. Some still order parts from a mobile, mostly by calling. That is a different market. The 5 calls in week 1 and the visits check it in the pilot market, the Netherlands. Everything else on this page is published research or a labelled assumption.
What I still assume
Not yet tested in the pilot market. Confidence is my read of the evidence so far.
Most important. If wrong, the design changes, so we check these first.
| Assumption | Confidence | Tested by |
|---|---|---|
| Most important: Most urgent orders still come by phone or WhatsApp, not the webshop | High | Workshop visits, 2-week call count |
| Most important: The fit score’s weights, and its threshold of 85 to order alone, are right | Medium | Threshold replay |
| Most important: Merchants will let an agent sell after a replay of their own past orders | Medium | Merchant interviews, threshold replay |
| Buyers will talk to a voice agent in a noisy workshop | Low | Voice test |
| Buyers will let their own agent order inside their company’s rules | Medium | Workshop visits, pilot |
| A workshop has an owner who approves orders over a limit the company sets | Medium | Workshop visits with owners |
| The approver trusts a 1-tap approval on WhatsApp | Medium | Workshop visits with owners, pilot |
| Buyers accept an agent once it says what it is | Medium | Voice test, pilot |
| A buyer with an urgent part orders from whoever answers first | Medium | Workshop visits |
| A plate or VIN photo is enough to find the car | Medium | Workshop visits |
| Buyers will trust an equivalent the agent picks | Medium | Workshop visits, pilot |
| 10 minutes is long enough to undo | Medium | Pilot, undo rate |
| Merchants’ data is clean enough to replay and match | Medium | Threshold replay with 3 merchants |
| Sales reps can answer within 5 minutes | Medium | Rep queue pilot |
What we test first
The threshold replay. It needs past orders, not participants, so no recruiting holds it up. Saga already reads them from each merchant’s ERP. The assumption the whole design rests on becomes each merchant’s own evidence.
Why first
The threshold, a fit score of 85 out of 100 to start, carries the whole design. Too low, and the Golf in bay 2 gets the wrong water pump. Too high, and it asks the buyer about parts it could have ordered alone.
- 85/100The thresholdSet per category, checked on every partThe fit score a part needs, out of 100, for the agent to order it alone.
- 90%Rep agreementPer category, in the replay, then liveHow often the agent picked the part the sales rep picked. A category goes live only from 90%.
- From week 1ReplayRun a year of past orders through the rules. Does the agent pick the part the rep picked?Moves on At 90% agreement
- Before go-liveShadowOn live calls, the agent suggests and the rep decides. We compare the 2.Moves on If 90% holds live
- Category by categoryGo liveOnly what the results support. Each category gets its own threshold, and the ones that pass go first.
Stop ruleUnder 90% agreement with the reps, the agent only asks the buyer to confirm in that category. It never orders alone.
Before launch
The first 12 weeks: what each study decides, and what stops it. A star marks the 3 that test the most important bets. Show all adds who takes part and how we find them.
- Tests a most important bet: Visits to 8 to 10 workshopsWeeks 1 to 4How urgent parts are ordered today. Shapes the journey.Workshop leads and owners. Watch an urgent order on the workshop floor, then ask why. Recruited through the pilot merchant’s workshop accounts, picked from its order data: phone-heavy, webshop-heavy and lapsed.Go or stop For 2 weeks at 1 branch, reps tag every order needed the same day. If fewer than half come by phone or WhatsApp, the journey starts in the webshop. The visits explain why.
- Tests a most important bet: Threshold replay on past ordersWeeks 1 to 8At which fit score the rep would agree. Sets the threshold per category.3 merchants, the pilot merchant among them. Past orders run through the rules. Shadow mode on live calls follows when the pilot starts. Recruited through Saga’s customer success team, each merchant with a signed data agreement.Go or stop A category goes live only at 90% rep agreement. Every disagreement is read with the rep’s reason, because the rep can be wrong too. Under 90%, the agent only asks the buyer to confirm. Without clean order history, shadow mode becomes the gate.
- Tests a most important bet: Merchant interviewsWeeks 2 to 4 and 9What a merchant needs to see before letting an agent sell.5 merchant admins and 5 sales reps. 45 minutes each, the sales reps with a prototype of the case card they get with each handover. Recruited through Saga’s customer success team. The researcher picks from the full list, including merchants with low webshop use.Go or stop In week 9, each of the 3 replay merchants sees its own replay. If a merchant still refuses, its agent only asks, in every category.
- 5 calls to workshop leadsWeek 1The buyer’s own words for an urgent order. Shapes the visit guide.Workshop leads. 20 minute phone calls: how do you order an urgent part today. Recruited through the pilot merchant’s workshop accounts.Go or stop Not a study. If the calls contradict the journey, the visits test that first.
- Voice test in a real workshopWeeks 6 to 8Where calls break in real workshop noise, and what to fix first.6 buyers. The real speech service listens in real workshop noise, and a person plays the agent’s decisions. 6 buyers find failures, not rates. Recruited through workshops from the visits.Go or stop If any buyer gets a wrong part, or 2 fail the same way, we fix it or test another speech service. Voice notes are the backup meanwhile.
- Rep queue pilotWeeks 9 to 12The case card a sales rep gets with each handover, the 5 minute answer target, and the lock: opening a case locks it, so 2 reps never take the same one.1 branch. The case card in the sales reps’ real queue. Recruited through the pilot merchant.Go or stop If reps miss the 5 minute target on more than 1 case in 5, the target or the staffing changes before launch.
After launch
What the pilot measures. It passes only when the pass marks hold, read after 300 agent orders and their return window.
- North starAgent speed on urgent orders, without more wrong parts.
- Minutes from the first message to the order placedPass mark5 minutes or lessthe median, beside the sales rep’s time on orders with no approvalFor orders with no approval, so the agent is judged on what it controls. The sales rep’s time is measured the same way, approval orders apart, on 30 to 50 of the pilot merchant’s phone orders, before day 1. Orders that need the approver are timed with the wait, as their own group, because the car waits too.
- Wrong-part returns and undo, by fit score bandPass markReturns do not doublethe goal: level with a sales rep’s, within 1 percentage pointAgainst sales rep orders in the same band. If they double, the pilot stops at once. At 300 orders only a doubling can be read, so the 1 point goal is read later, on more orders.
- Share of urgent orders through the agentshould grow
- Handover rate and sales rep answer timeanswer within 5 minutes
- Override rate per category: a sales rep changes the agent’s pickshould fall over time
- Days from signup to live, categories live per merchantshould fall
- Threshold reviewed per categoryevery month
- 20 agent conversations read in fullevery week
- The reason behind each override, codedevery week
- Interviews with 5 buyers who undid or returned an order, and 5 who did notevery month
- Workshops that order again within 30 daysshould hold or grow
- Saga’s margin per conversation, once out of shadow modeshould stay above 0
- Approver answer timeapprovals in seconds, every time, are a warning
- Buyers who check the part before fitting itasked in the monthly buyer interviews
I assume the pilot’s first 2 categories, filters and cooling, give about 280 agent orders a month, so the pilot runs 6 to 10 weeks. 300 orders is a harm check, not proof: it catches a real failure, and cannot prove the agent as good as a sales rep. Published wrong-order rates vary, so it measures against the merchant’s own baseline. 66% of German workshops put theirs under 5%: AUTO SERVICE PRAXIS survey, 2024, 500+ workshops.
AI drafts the synthesis. A researcher checks every quote against the recording before it becomes a finding.
The buyer’s agent and the merchant’s 4