Elma Consulting
ELMA Consulting is an accounting, legal and tax firm. Its leads come in through Instagram Direct from Meta ads and land in the firm's CRM.
An AI sales-desk agent: it scores every Instagram lead with a quote as evidence for each point, sorts target from noise, and puts a priority queue in front of the sales team every morning.
Where it stood
Leads were reaching the CRM. The problem was that every lead created the same manual work, whatever its quality.
Every lead cost the same work
Each new chat got the same cycle of tasks and touches, whether it was a real prospect or a one-word message from an ad button.
Busy managers on top of other work
The sales managers handled Instagram on top of their other work, with a planned 3–5 touches per lead that in practice became 5–7.
Noise in the pipeline
Silent and non-target chats went through the same cycle as real prospects, and some of them leaked further into the sales funnel.
Rules in one person's head
What makes a good lead was known to the head of sales, but written down nowhere.
What we did
Rules written down first
We turned the head of sales' judgement into a versioned set of written rules. They change without touching the workflow.
Scoring with evidence
The model scores each lead on a few criteria and has to quote the message that justifies each point. It may answer 'not enough evidence'. Code adds up the points and applies the thresholds; the model never sets the total.
Noise only on proof
A lead is marked as noise only on positive evidence. When evidence is thin it goes to 'clarify', never to noise, so a real prospect is not thrown away.
A morning priority queue
Every workday the sales team gets the leads in order in Telegram, each linked to its CRM card; noise shows only as a count.
A strict human gate
The agent writes only two fields on the lead card and its own log. It never sends anything to a client.
Data kept tight
Profile facts only from Meta's official APIs, no full dialogues stored, the model provider approved in writing by the client, and a move to the client's own server.
// what the agents did
Read new leads every 15 minutes, scored them with a quote for each point, sorted target from noise and built the morning queue.
// what people decided
The head of sales owns the rules and graded the test set; managers decide who to call and what to say; the agent's two fields went live only after calibration.
What changed
- Every new Instagram lead scored about 15 minutes after the client's last message, around the clock
- The sales team starts each workday with the leads already in priority order — no inbox sweep first
- Noise no longer gets the manual 5–7-touch cycle: it shows only as a count in the morning queue
- 90% agreement with the managers' own marks on 51 live leads; 0 target leads marked as noise
How we run it today. Sprint 1 closes with the move to the client's server. Next sprints are proposed: the historical lead base and automatic replies.
Get a proposal
Tell us what you want to grow. We reply within 24 hours with what we'd look at first. Every engagement starts with a free audit: SEO, ad account or analytics (GA4, GTM). You keep the findings either way.