* Topics
* [Guide](/academy/guide/)
* [Video course](/academy/marketplace-video-course/)
* [The podcast](/academy/marketplace-academy-podcast/)

[Subscribe](#subscribe-academy)

[Starting a marketplace](/academy/start-a-marketplace/)

# How to run your marketplace operations with AI agents

A practical framework for automating your marketplace operations with AI agents, proven by how Drive lah cut its team from 88 to 35 while growing the business.

![Photo of Juho Makkonen](https://images.prismic.io/sharetribe/aWTjxgIvOtkhBToB_juho_serious.png?auto=format%2Ccompress&fit=max&w=3840)

[Juho Makkonen](/author/juho-makkonen/), CEO & Co-Founder

Sep 22, 2026Last update Sep 22, 2026

![Woman pushing sphere on dominoes.](https://images.prismic.io/sharetribe/OFbqd8dJRP3btN_i_dominoes-chain-reaction.jpg?auto=format%2Ccompress&fit=max&w=3840)

Table of contents

1. [What is an AI agent, for a marketplace operator](#what-is-an-ai-agent-for-a-marketplace-operator)
2. [What AI agents already handle for marketplace operators](#what-ai-agents-already-handle-for-marketplace-operators)
3. [When AI agents aren't for you yet](#when-ai-agents-arent-for-you-yet)
4. [How to automate your marketplace operations with AI agents](#how-to-automate-your-marketplace-operations-with-ai-agents)
  1. [1\. Give agents case-specific context, not generic FAQ information](#1-give-agents-case-specific-context-not-generic-faq-information)
  2. [2\. Start with your single worst pain point, not everything at once](#2-start-with-your-single-worst-pain-point-not-everything-at-once)
  3. [3\. Ship something imperfect, then improve weekly, with a review layer watching it](#3-ship-something-imperfect-then-improve-weekly-with-a-review-layer-watching-it)
  4. [4\. Let agents score confidence and escalate uncertainty, instead of forcing a binary decision](#4-let-agents-score-confidence-and-escalate-uncertainty-instead-of-forcing-a-binary-decision)
  5. [5\. Treat response speed as a growth lever, not just a cost saver](#5-treat-response-speed-as-a-growth-lever-not-just-a-cost-saver)
  6. [6\. Pick the harness over the model](#6-pick-the-harness-over-the-model)
  7. [7\. Redesign roles around agents, don't just cut headcount](#7-redesign-roles-around-agents-dont-just-cut-headcount)
5. [Results at a glance](#results-at-a-glance)
6. [Key takeaway](#key-takeaway)
7. [FAQ: AI agents for marketplace operations](#faq-ai-agents-for-marketplace-operations)

Most marketplace advice is about getting to launch: finding supply, matching it with demand, taking your first cut of revenue. Less gets said about what happens after, once the transactions start coming in and the operational load behind them keeps growing too. That means verifying providers, resolving disputes, chasing late payments, and answering support tickets, each one tied to a specific booking. For most of the last decade, handling that load meant hiring, and hiring meant needing either enough revenue or enough funding to pay for it.

That's changed. AI agents can now read a marketplace's data and act on it the same way a person in an ops role would: check a document, flag a mismatch, answer a ticket, approve a payout. A small founding team can [run a business that used to need a much bigger one](/academy/how-to-scale-your-marketplace/), with a person reviewing exceptions instead of handling every case by hand.

We at Sharetribe have been [building toward enabling this for years](/balanced/marketplace-engine-for-the-ai-age), and we are now seeing our customers live in this new reality.

Here's what that looks like in practice, based on a recent conversation with Gaurav Singhal, co-founder of Drive lah: he walked through, in detail, how his team used agents to collect more of what customers owed, cut support and finance headcount, and turn the business profitable.

You can [watch the full conversation on Sharetribe's YouTube channel](https://youtu.be/OBQYUjGDrLw); we'll point back to specific parts of it as we go.

![Drive lah marketplace's main page screenshot](https://images.prismic.io/sharetribe/189d3618-feb3-4c47-b955-508dfad9a1da_DriveLah.png?auto=format%2Ccompress&fit=max&w=3840)

#### About Drive lah

[Drive lah](/customers/drive-lah/) (Drive mate in Australia) is a [peer-to-peer](/how-to-build/peer-to-peer-marketplace/) car-sharing marketplace built on Sharetribe since 2019, with more than 4,500 cars and 400,000 registered users across 12 cities in Singapore and Australia, and about $15M in revenue.

## What is an AI agent, for a marketplace operator

Let's get one thing out of the way first: an AI agent isn't just a chatbot with a better name. A chatbot answers questions.

**An agent is different**: it's a system that can look at your marketplace's own data and take an action on it, the kind of action a person in an ops role would otherwise have to do by hand, without you asking it to every time.

Think of it as a spectrum. At one end, you've got a tool you can simply ask things, like "_how many bookings did we get last week from Sydney_," and it pulls the answer straight from your data and hands it back in plain English. That's genuinely useful, and often the easiest place to start.

At the other end is something closer to what [Drive lah](/customers/drive-lah/) built: named agents with real jobs.

* _Stella_, Drive lah's collections agent, chases down unpaid invoices before they turn into bad debt.
* _Fiona_, Drive lah's sales agent, qualifies leads in Facebook Messenger, quotes prices, negotiates, and collects a deposit before a staff member even joins the conversation.

Most marketplace operators start at one end of that spectrum and move toward the other as they get comfortable trusting an agent with real decisions. Nothing says you have to jump straight to the deep end.

## What AI agents already handle for marketplace operators

Here's where this shows up in a live marketplace today.

Four places it's already doing real work:

* Customer service
* Verification and onboarding
* Collections and accounts receivable
* Sales and lead qualification

#### Customer service is the most obvious one.

Most marketplace support tickets aren't generic FAQ questions, they're tied to one specific booking: why is this refund the amount it is, why hasn't this door code arrived. 

An agent that can pull the actual booking from Sharetribe, the payment from Stripe, and the message history from Intercom, in real time, can resolve a real share of those without a human touching them.

At Drive lah, that's about 40% of roughly 5,000 monthly tickets, resolved with zero human involvement, which lets the team run support with 22 people instead of 36.

#### Verification and onboarding is another.

Checking an ID or approving a listing used to mean a team working 9-to-5, so anything that came in outside office hours just sat in a queue.

Running the checks in parallel, across identity databases, fraud signals, and [internal risk scores](/academy/most-common-marketplace-attacks/), means approvals can happen in seconds instead of hours, any time of day. Drive lah brought **verification time down from 3-4 hours to 40 seconds** this way, most of it without a person touching it.

#### Collections is where a lot of marketplaces feel the most pain.

Collections is a good place to start precisely because the return is so measurable: chasing overdue payments and resolving disputes is repetitive, rules-based work an agent can take on end to end, in chat, with no human involved.

Drive lah's collections agent, Stella, works from trip and user records in Sharetribe and payment data in Stripe to work out exactly who owes what. It took their **collection rate from 70% to 89%** and added more than half a million dollars in recovered payments in eight months.

#### Sales and lead qualification too.

Leads that arrive outside business hours are usually the ones that go cold, simply because nobody's there to answer in time. An agent that lives inside a messaging app can qualify a lead, quote a price, negotiate, and collect a deposit before a person ever joins the conversation.

Drive lah's sales agent, Fiona, pulls vehicle pricing and photos straight from [Sharetribe's API ](/features/headless-marketplace-solution/)to answer every lead within minutes, any time of day. It **lifted conversion by 60%**.

These four aren't the whole list. They're just what happened to come up in the webinar. The process worth automating first is whichever one is quietly costing you the most, whatever that turns out to be for your marketplace.

## When AI agents aren't for you yet

You'll hear a version of this: start on day one, don't wait.

Gaurav said almost exactly that in the conversation: if you're not deploying AI in some part of your business, you're already behind, whatever stage you're at. It's a fair point, and also easy to misread.

The part that gets lost is what "starting" really means. It doesn't mean switching an agent on before you're ready. It means starting to get your data into one place, because an agent built on top of scattered spreadsheets and three disconnected tools won't fix the mess, it'll just make bad decisions faster than a person would.

Garbage in, garbage out still applies, agent or not.

People ask whether they're big enough for this. Wrong question. The real one is whether you know where your own data lives. If the honest answer is no, that's your day one, and it's exactly where we start below.

## How to automate your marketplace operations with AI agents

Once you've decided this is worth doing, here's the order that works, based on what Drive lah did, plus a couple of things worth knowing before you start.

### 1\. Give agents case-specific context, not generic FAQ information

Here's a mistake that's easy to make: treating this like installing a chatbot. A generic FAQ bot fails in a marketplace because almost nothing about a marketplace ticket is generic. "Why is my refund this amount" only has an answer once you know which booking, which host, which dispute it's attached to.

Drive lah's support agents work because they're wired into the actual booking, payment, and message history. That's the difference between an agent that resolves 40% of tickets with zero human involvement, and a generic chatbot that just frustrates people faster.

How to provide this context to the agents? In the early days, it might be enough to give the agent read access to your different systems – Sharetribe for core marketplace data (users, listings, transactions, messages, reviews, etc), Stripe for payments, Intercom for support tickets, and so on.

As you scale, this approach could become inefficient. A single customer ticket can take half a dozen calls across three systems before the agent knows enough to answer, and the same records get pulled again on the next ticket, and the one after that. Drive lah solved this by building their own data warehouse where they pull data from Sharetribe, Stripe, Intercom, and their other systems. This data warehouse serves as the context layer for their agents.

### 2\. Start with your single worst pain point, not everything at once

Don't try to automate everything at once.

Here's a real way to find where to start: look at where you've already added headcount and the problem keeps getting worse anyway. That's usually the tell.

Drive lah had exactly this signal with collections: they kept hiring more people to chase down late payments, and the collection rate kept slipping regardless. More people weren't fixing it, which is a good sign the problem is the process, not the staffing.

Now, Stella, Drive lah's collections agent, works from trip and user records pulled via Sharetribe's API and [payment data from Stripe](/academy/marketplace-payments%5F%5Fstripe-connect-overview/), to work out exactly who owes what and why.

So start where your business is already hurting, see if there is an AI solution, prove it works, then move to the next one.

### 3\. Ship something imperfect, then improve weekly, with a review layer watching it

Any agent you deploy will get things wrong at first. That might mean something needs fixing, or it might just be the normal rough patch any new agent goes through. You often can't tell which until it's running in production.

Drive lah's collections agent, Stella, wasn't good on day one either.

They shipped it anyway, then improved it on a weekly cadence, with a separate monitor-agent watching for mistakes and flagging them.

Waiting for an agent to be perfect before you deploy it just means waiting forever. Build the review layer instead, and let the agent get better in production.

### 4\. Let agents score confidence and escalate uncertainty, instead of forcing a binary decision

Not every decision an agent makes has to be certain. What matters is whether it knows when to act on its own and when to hand a case to a person.

[Identity verification ](/academy/build-trust-marketplace/)is the clearest example of this: checking a new driver's ID or a [car owner's listing](/create/how-to-build-marketplace-for-car-rentals/) before they can use the marketplace.

Drive lah's agent runs checks in parallel across several sources, government databases, ID verification,[ fraud signals](/academy/financial-crime-marketplaces/), and turns them into a single confidence score. Clear cases get auto-approved or auto-declined. Everything in the uncertain middle gets escalated to a person.

### 5\. Treat response speed as a growth lever, not just a cost saver

Most people treat an agent's speed as a cost question: does it work faster than a person, so you need fewer of them.

Drive lah found something else: speed moved numbers they weren't even trying to move.

Verifying a new driver's ID or a car owner's listing used to take hours; making that instant didn't just save time, it lifted first-time activation by 41%, because fewer people dropped off while waiting for approval.

Instant lead response through Fiona, Drive lah's sales agent, lifted conversion by 60%, because leads that used to go cold at 9pm on a Saturday now got answered right away. Fiona pulls vehicle pricing and photos straight from Sharetribe's API to make that possible.

If there's a slow, wait-heavy step anywhere in your own funnel, it's worth checking what that wait is actually costing you beyond labor hours. The answer might be conversions or signups, not just time.

### 6\. Pick the harness over the model

It's tempting to think the AI model you pick is the important decision.

Gaurav's view, which he said held true across every agent Drive lah built, was that the specific model matters far less than how you use it.

The clearest example of that is what happened on the dev side: his team shrank from 14 people to 6 while shipping 5 to 6 times more, mainly using Claude Code, and in his view, swapping in a different top model would have gotten similar results nine times out of ten.

What actually matters is the harness around it, the workflows, the context, the guardrails you build. The model is replaceable. The system you built around it usually isn't.

### 7\. Redesign roles around agents, don't just cut headcount

There's a version of this story that's just about headcount, and it's true: Drive lah runs a bigger business with 35 people instead of 88\. But that's an easy read.

The harder, more useful shift is redesigning what your team's roles are once agents take over the repetitive parts.

Gaurav's co-founder, Dirk-Jan ter Horst, put it well: design the business around agents and humans working together from the start, instead of bolting an agent onto an org chart that hasn't changed. That's the part that's easy to skip, and probably the part that matters most.

## Results at a glance

None of this is hypothetical.

Automate the right processes and the results tend to show up in the same few places: fewer people needed to run the same or a bigger business, revenue recovered that used to slip through as bad debt, and processes that used to take hours happening in seconds.

Here's what that looked like for Drive lah specifically. In about a year, their team went from 88 people to 35, running a bigger business across 12 cities. Their collection rate went from 70% to 89%, worth more than half a million dollars in eight months. Verification time dropped from 3-4 hours to 40 seconds. And the business, which was burning $2.5M a year and struggling to raise its next round, is now profitable.

## Key takeaway

None of this requires a big team, a big budget, or waiting for the perfect model. It requires knowing where your data lives, picking the processes that are hurting the most, and being willing to ship something imperfect and improve it in public.

Drive lah was already a large marketplace when they started working with AI. If Gaurav and Dirk-Jan were to launch a marketplace today, they would design their process around AI from the get-go, allowing them to scale their business without scaling headcount.

## FAQ: AI agents for marketplace operations

#### Do I need a developer to build AI agents for marketplace operations?

For simple data questions, no: a tool like [Lemonado](https://www.sharetribe.com/help/en/articles/8813081-talk-to-your-data-with-lemonado) can connect to your marketplace and answer questions in plain English without any code. For a custom agent that takes actions on its own, like Drive lah's, yes, at least early on. But the balance is shifting: development now is less about headcount and more about defining exactly what you want built, and increasingly, that work can be done by people who don't know how to code themselves, as AI agents help with that too.

#### What's the minimum scale for AI agents to make sense in your operations?

Smaller than you'd think, if you frame it as "what's my worst manual process" rather than "am I big enough." The pain point, not the revenue number, is what tells you you're ready.

#### Which marketplace operations should you automate with AI agents first?

Whichever one is hurting the most right now, not whichever one looks easiest. Drive lah started with collections because unpaid invoices were the sharpest pain, not because it was simple.

#### Does this replace my team?

Not in the way that sounds. Drive lah's team got smaller, but the business got bigger, and the people who stayed took on different work: reviewing exceptions, handling escalations, deciding what the agents should do next.

#### What's the difference between a tool like Lemonado and a custom agent like Drive lah's?

[Lemonado](https://www.sharetribe.com/help/en/articles/8813081-talk-to-your-data-with-lemonado) answers questions about your data. Drive lah's agents act on it: chasing payments, verifying IDs, replying to customers, without anyone asking them to first. One's a starting point. The other's where you end up once you trust the system enough to let it act on its own.

---

![Photo of Juho Makkonen](https://images.prismic.io/sharetribe/aWTjxgIvOtkhBToB_juho_serious.png?auto=format%2Ccompress&fit=max&w=3840)

[Juho Makkonen](/author/juho-makkonen/), CEO & Co-Founder

Sep 22, 2026·Last update Sep 22, 2026

Copy article link

Start your 14-day free trial

## Create a marketplace today!

* Launch quickly, without coding
* Extend infinitely
* Scale to any size
[Start free trial](https://console.sharetribe.com/new)

No credit card required

[Starting a marketplace](/academy/start-a-marketplace/)

[Business model](/academy/business-model/)

[Marketing & growth](/academy/marketing-growth/)

[Platform technology](/academy/platform-technology/)

[Insurance](/academy/insurance/)

[The Backstage](/academy/backstage/)