Case Study · Account-Based Marketing · AI Research Platform
How We Built An AI Account Research Platform That Found A Hidden 98-Store Chain
Snapshot
The Build At A Glance
The brief
Replace A Spreadsheet And 34-Page Docs With One Ranked List
Our client sells into Ontario’s licensed cannabis stores, from 100-store chains down to single shops. Product reaches the shelf through the provincial wholesaler, so the sale is really won with the retail buyers who decide what to stock. That is a textbook account-based marketing problem: a known, finite list of companies, and a small number of people inside each one who matter.
Before this build, the target list lived in a Google Sheet of about 237 accounts. Every key account had its own long Google Doc, some running to around 34 pages. Nobody could see at a glance which accounts were well understood, which were out of date, who the buyer was, or what was still unknown. Research was manual and uneven, and facts weren’t tied to where they came from.
The brief was to replace all of it with one platform: every possible account in the market, ranked, with a one-page brief per key account and every fact backed by a source. And one rule from the start: research first, no outreach, so the first time anyone picks up the phone it is to the right person, about the right gap, inside the rules.
The market
Why The Big ABM Tools Don’t Work Here
Platforms like 6sense, Demandbase, Apollo and Clay are built for B2B markets where every buyer has a domain, a LinkedIn page and a record in a data vendor’s database. Ontario cannabis retail is none of those things.
The buyers aren’t in the databases
The market is roughly 1,900 licensed stores, most of them small private companies. The source of truth is a government register, and the name of the real buyer sits in places like licence footers on store websites, trade press and court filings. Generic B2B data tools are thin here.
The account is the licensee, not the sign
Stores trade under brand names, but the buying decision sits with the licensed company behind them, which often runs several banners. Off-the-shelf tools key on website domains, so they count one buyer as four.
It is a regulated industry
Ontario’s retail rules ban inducements: no rebates, paid shelf placement, exclusivity or paid travel. Every recommended play had to be built around education, training and samples instead, and checked against those rules.
The research has to be provable
A sales team acting on a wrong buyer name or a guessed email burns the relationship before it starts. The platform needed to show where every fact came from and how confident it was, which the big tools don’t do for this kind of research.
The challenge
Three Problems We Had To Solve First
Building a dashboard is the easy part. Making the research behind it accurate, honest and affordable to run was the actual job.
Find the real buyer
Nothing in the regulator’s register links stores to their parent company. The link hides in website footers, licence numbers that turn up on another chain’s site, shared loyalty programs and acquisition news.
Make AI research trustworthy
Left alone, AI researchers will guess an email, invent a name, mark a task done when they found nothing, or overwrite a person’s work. Any one of those makes the whole list untrustworthy.
Run it unattended, safely
Deep research on one key account takes around 22 minutes. It had to run on its own, one account after another, without improvising risky commands or burning through the AI plan’s usage limits.
The hidden #2 chain
In the original spreadsheet, four store banners sat as four separate mid-sized accounts of 12 to 46 stores each. The platform’s parent-company check found the same numbered company in two of the banners’ website footers, and a loyalty program shared across all four. Merged, they are one 98-store company: the second-largest cannabis chain in Ontario.
That is the difference between pitching four mid-sized accounts separately and pitching one buyer who controls 98 shelves. Later, when we rebuilt the list automatically from the public register alone, the platform found the same link on its own. In total it found and merged eight duplicate accounts, including five “single stores” that were really part of a 55-store chain.
How it works
From A Government Register To A One-Page Brief
1. Start from the regulator’s list, not a data vendor’s
The whole market, not a sample of itThe platform downloads the AGCO’s public register of every authorized cannabis store in Ontario: 1,903 stores. Nothing is missing because a vendor never indexed it.
2. Group stores by who really buys
625 accounts, sorted into four tiersStores are grouped by licensed company, then checked for hidden parents through shared websites, legal names in footers, licence numbers and loyalty programs. Accounts are tiered by footprint: 11 key accounts with 20+ stores get a custom one-to-one plan, 61 with 4–19 stores get a regional play, 162 with 2–3 stores get a broader one, and 391 single stores wait for phase two. When we tested it, rebuilding the list from the brief alone put 70 of the 72 hand-built key accounts in the same tier.
3. Research each account with a team of AI agents
Four researchers, a fact-checker and a writerFour AI researchers work in parallel on each account: the company, the buying centre, the shelf, and “why now” signals like store openings, acquisitions and leadership changes. A separate fact-checker grades every finding and rejects anything without a source. A writer turns what survives into a one-page brief, with a recommended play and the routes to reach each buyer.
4. Humans keep every judgement call
The AI proposes, a person acceptsAnything the AI isn’t allowed to decide on its own lands in a review queue called “Needs your eyes”, each item with a recommendation: merging two companies, conflicting sources, a proposed score, or a change to something a person entered. The AI can never contact anyone, delete anything, guess an email address, or collect anyone’s personal details.
Every fact is stored with its source link, its date and a confidence grade: verified by two independent sources, single-source, or unverified. A monthly check lowers an account’s research score as facts age past their refresh window, so the list never quietly goes stale.
The scoring
Three Numbers, Never Blended
Most ABM tools roll everything into one opaque score. This platform keeps three separate ones, because they answer three different questions.
Fit: should we want them?
A 0–10 score built from ideal-customer fit, current need, opportunity size and buying history. It only shows when every part has been scored, so a half-known account never looks like a sure thing.
Research: do we know enough?
The share of a tier-specific checklist that is actually filled with a sourced fact. “Not found” counts as zero and is retried after 30 days, so no account can fake 100%.
Warmth: do they know us?
A 0–100 score from pipeline stage and logged touches, banded Cold, Aware, Warming, Warm and Hot. Touches fade to zero over 90 days, so it measures the relationship now, not its history. Replies, meetings, store visits and training sessions count the most.
The order matters. Research comes first, then warming. An account is only worth warming once we know who the buyer is, what is missing from their shelf, and which rules shape the conversation.
What we delivered
One Platform, Built To Take More Clients
Built for this client first, with a client switcher so the same platform can run account-based programs for other clients, each with its own list, checklist and scoring.
The account platform
An overview with the tier ladder, a spreadsheet-style list of every account with its warmth, research and fit side by side, a one-page brief per account, a board for the key accounts, and a store view of all 1,903 locations for field routing. CSV import and export throughout.
The AI research engine
Nine specialised AI agents covering company research, the buying centre, the shelf, signals, parent-company checks, fact-checking and writing. It works through accounts on its own, one at a time, and every run passes a quality gate before anything is posted.
The list builder
Builds a new client’s account list from a public register in one step, including the parent-company check, so the next market starts from a complete, grouped and tiered list instead of a blank sheet.
Guardrails you can audit
A review queue for every judgement call, compliance checks on every recommended play, nightly integrity-checked backups and 110 automated tests that run against copies of the data, never the live database.
It runs on our own setup with no per-seat licences and no paid data subscriptions. Every source it uses is public: the register, company websites and their public menus, trade press, filings and conference pages. No logins, and no age gates crossed.
The results
What The First Week Produced
Before
A 237-row spreadsheet grouped by store name, long research docs for the biggest chains, and research progress that was self-reported. Four banners of one company counted as four accounts.
After
One ranked list of 625 accounts covering all 1,903 stores, a sourced one-page brief per key account, and the hidden 98-store chain treated as the single buyer it is. 10 of the 11 key accounts were researched by the engine in its first week.
A stricter definition of done
Under the old, looser rules the key accounts looked about 78% researched. Against the platform’s strict checklist, where only a sourced fact counts, they averaged 51%. That drop is the point. The team now knows exactly what is still unknown for each account, instead of a number that looked finished and wasn’t.
A typical research summary for one key account read: research coverage up from 5% to 42%, the chain’s licensing model confirmed, 37 stores matched against the regulator’s register, and a shelf check across a sample of its stores showing that none carried a product in our client’s format. That is a reason to call, found before anyone calls.
Figures are as recorded in the platform’s database and run logs on , covering the build on September 24–25, 2026 and the unattended research runs through September 29, 2026. This was a research-first build: no outreach had been sent through the platform in this period, by design, so there are no meeting or pipeline figures to report yet.
What’s next
From Research To Warming
The research half of the program is built. The warming half is where it goes next.
Hot-account alerts
When an account’s warmth crosses into Hot, the right person is told straight away, with the brief and the play attached.
Consent-aware outreach
Email outreach built with CASL consent handling from the start, so warming never gets ahead of the rules.
CRM sync
Accounts, contacts and touches flowing both ways with the sales team’s CRM, so activity logs itself.
Outcomes feeding the scores
Won and lost deals re-weighting the fit score, so the ranking learns which accounts actually convert.