The data didn't exist. So we manufactured it.
Paris Gourmet, a $35M specialty food importer, sells gourmet ingredients to chefs, but only through distributors. It had never sold to a restaurant directly, and could not name one of its own end customers. The CMO wanted a direct channel. The obstacle was not strategy: it was that the data required to prospect restaurants does not exist as a commercial product. Nobody sells a list that tells you what a restaurant cooks, at what price point, what it therefore buys, and how to reach the chef rather than the host stand. So we built one: an AI agent swarm that assembles 30,000 verified, ICP-qualified restaurant records at about $0.08 each. In two quarters the channel produced 140 net-new direct accounts, worth an estimated $2.4M a year.
The problem
Paris Gourmet had spent decades selling ingredients to chefs, and could not name a single chef it sold to. Everything moved through distributors: the company's actual end customer was a stranger.
The CMO wanted to change that: a direct-to-restaurant channel running alongside the distributor model, not replacing it. The company's own plan valued it at roughly $5M a year, about 14% on top of the base, at a better margin. And it was blocked. Not on strategy. On data.
You cannot prospect restaurants with a purchased list. Every field that would make one a qualified prospect is a field no provider sells.
A stranger on the plate. Everything moved through distributors. That model is most of a $35M business, but it means no relationship, no data, no visibility, and no way to sell anything directly to the people who put the products on a plate.
A channel worth millions, unreachable. The CMO's direct-to-restaurant channel was valued at ~$5M a year at a better margin. It was blocked not on strategy but on data: the list required to reach it did not exist, at any price.
Commercial data is useless here. You can buy a restaurant's name, address and phone. You cannot buy what cuisine it serves, what price tier it runs at, what that implies it buys, or who the chef is as opposed to who answers the phone at the host stand.
And whether any of it is still true. Even the fields you could assemble decay fast: restaurants close, chefs move, menus change. A restaurant list is 20-30% wrong within a year, which is exactly why no vendor sells a good one.
The diagnosis
I was hired to build an awareness campaign. There was nothing to aim it at.
This is the reframe the whole engagement turned on, and it's not a marketing insight. It's an engineering one. The brief was a campaign. The problem was a data-manufacturing problem. Any agency handed this would have bought the best restaurant list available, run a campaign against it, produced a 1-2% reply rate, and concluded that direct-to-restaurant doesn't work for specialty food. It would have been the wrong conclusion, reached competently.
If the data doesn't exist for sale, that's the reason nobody else is in the channel
The information wasn't missing. It was scattered across Google Maps, Yelp, delivery platforms, social media, public records and each restaurant's own website. Assembling it by hand for 30,000 restaurants is uneconomic, which is precisely why no vendor does it and the market sat empty. Agents change that arithmetic: what was uneconomic at $50 an hour is trivial at $0.08 a record.
The menu is the buying signal
A restaurant publishes, in public, a detailed declaration of exactly what raw materials it buys, in what style, at what price point. It's called a menu. Nobody was reading them, because reading 30,000 menus wasn't a thing a person could do. It's now a thing an agent can do in an afternoon. Your buyers are almost certainly publishing their buying signal somewhere public too.
Verify the human, not just the record
A restaurant's listed phone number reaches a host stand, not a chef, and the chef is the only person who decides what gets bought. So the agents call and establish whether the line reaches a decision-maker or a podium. A verified number that reaches the wrong human is a verified wrong number, and that single distinction is most of why the response rate came out at 24% instead of 2%.
What I built
A two-stage agent swarm that manufactures the data
Collection agents pull from Google Maps, Yelp, delivery platforms, aggregators, social media and public records. Enrichment agents clean, deduplicate, verify and ICP-filter what comes back, including automated menu and price-tier analysis, and agentic phone verification that distinguishes a decision-maker from a front desk.
A proprietary database that doesn't exist anywhere else
About 30,000 enriched, ICP-qualified restaurant records: cuisine, inferred raw-material demand, chef and management profiles, verified email and phone, geography, and price tier. National coverage, weighted to the east coast. Every record re-enriched weekly; roughly 1,500 net-new records added every fortnight.
The outbound engine on top of it
Sequences across email and phone through Instantly and ActiveCampaign, orchestrated in n8n, on 12 sending domains and around 30 rotating calling lines. AI agents write each message from that specific restaurant's menu, inferred requirements, price point and location, because the data is good enough to make genuine personalisation cheap.
The CRM motion to catch what it produced
HubSpot pipeline stages built for direct-restaurant deal flow, lead routing across the 18-rep sales team, and lifecycle attribution tying new-channel activity to the accounts it won.
And the reporting to prove the channel works
Prospecting and ABM dashboards (outbound volume, response and meeting rates, account coverage, and channel-attributed pipeline and revenue) so the CMO can defend the investment with numbers rather than belief.
The outcome
Live since September 2025 and still running; figures over the first two quarters. The ~$5M channel value and ~40% attainment are Paris Gourmet's own plan numbers, not my forecast.
Shashank didn't come back with a campaign. He came back and said the data we needed didn't exist and he was going to build it. In two quarters we had a hundred and forty restaurants buying from us directly.
What this is, and isn't
This did not replace the distributor model, and was never meant to. Paris Gourmet still sells through distributors and always will. This is an additional channel on top of an existing business, not disintermediation, and not a rescue. Anyone reading this as “consultant helps company cut out its distributors” has read it wrong.
The margin story is real but unmeasured. Direct sales carry a better margin than distributor sales. That's arithmetic, not a claim. But I can't yet tell you by how much, because the company doesn't measure it cleanly enough to publish. I'd rather say that than give you a number I made up.
And this system has an ongoing cost, permanently. Restaurants close, chefs move, menus change: a restaurant database is 20-30% wrong within a year of being built, which is exactly why nobody sells a good one. Weekly re-enrichment isn't a nice-to-have; it's the price of the asset continuing to exist.
Next: the channel is at ~40% of its plan and growing. The obvious extension is applying the same menu-reading logic to the distributor's customers: the accounts Paris Gourmet already reaches indirectly and still cannot see.
Why this matters for you
This is your problem if:
You sell through distributors, resellers or a channel, and you cannot name a single one of your end customers
The list you'd need to run outbound properly doesn't exist for sale. You've checked, and what's available is names and phone numbers
You have bought a contact database and the fields that would actually qualify an account aren't in it
Your outbound reply rate is 1-3% and you've concluded the channel doesn't work in your market
Your ICP is defined by something no database has a column for
If the data you need doesn't exist for sale, that's not a reason to abandon the channel
It's the reason nobody else is competing in it. The absence of a list is a moat, and it belongs to whoever builds one first.
Your buyers are probably publishing their buying signal in public
A restaurant's menu is a purchase declaration written by the buyer, for free, updated whenever they change their mind. Find your industry's equivalent (the job posting, the tech stack, the planning application, the product page) and read it at scale.
Verify the human, not just the record
Data providers verify that a phone number connects. They don't verify that it connects to a person who can say yes. Those are completely different products, and only one is worth paying for.