TAMLY ai salesplatform

From zero to AI-ready in 3 steps

Before the AI can do anything useful, it needs data. The onboarding had to bridge a critical gap - getting users to connect their CRM without it feeling like an IT integration task.

From noise to organized, living lists

The concept wasn't novel - segmenting accounts into lists is standard practice. What I redesigned was the maintenance burden. Audiences in TAM are criteria-based: you define the filters once, and any new entrant is automatically evaluated and placed. Lists stop being static snapshots and start reflecting reality.

Key design decisions:

•Audience name = human label, not a filter slug. Users name their audiences like they think about their customers: "Core ICP", "Strategic ICP", "Marketing ICP"

•Dynamic badge on save: makes the always-updating behavior visible and reassuring, not hidden

•Filter → count update: as each filter is added or removed, the matching company count animates to its new value - communicating that the audience is computed in real time, not frozen

Configuring AI without writing code

Research is where deals stall. The enrichment agent addresses this directly: a single prompt triggers a search across public sources and Warmly's internal databases, returning results as a column attached to any audience. Build it once, reuse it everywhere.

The trust question mattered here. I chose source attribution over confidence scores or tooltips - showing where the data came from, not how sure the system is. Provenance over abstraction.

From a company to the right people inside it

B2B deals don't close with companies - they close with specific people inside companies. But most CRMs have account data and contact data in separate silos, with no intelligence about which contacts matter for a given deal. The buying committee finder had to solve two problems in sequence: classify existing contacts by persona, then surface who you're missing.

Buyer personas: name your persona (Champion, Economic Buyer, Technical Evaluator), define include/exclude titles, add special instructions. Drag to rank by priority. As many personas as needed.

Test panel: pick any company → instantly see which existing CRM contacts match each persona and which would be net new.

The TAM column view shows the buying committee inline per account so users can quickly scan available data and focus on the most complete one.

Run log records every enrichment: company, matched contact, persona, confidence %, last run timestamp. Confidence range: 98% for strong title matches, 75% for adjacent roles, 12% for weak matches.

Row states in the table: Normal → Processing → Reprocessing (user-triggered) → Cancelling → Failed. Each state has distinct visual treatment so users know what the system is doing per account.

A machine learning model that a sales rep can set up in 3 clicks

The next logical step was to enhance the system with a score - one that could take all available signals and data per account and contact and answer a single question: who do I prioritize? The design challenge was making a machine learning model feel owned and configured by the user, not handed to them as a black box. At the same time, the configuration UX couldn't require any data science knowledge.

We launched the TAM Agent platform in beta and spent the next six months refining it. The core retention signal we tracked: how many users came back versus how many checked it once and left. 71% returned within the first week - a strong signal that the platform was delivering enough value to pull people back.

One pattern stood out in the usage data - prompt quality was low. Users weren't spending time crafting good prompts, which directly limited what each agent could return. That became a fast follow-up: a lightweight prompt improvement flow that made it easy to write better inputs. Users who engaged with it got noticeably richer outputs across every agent they configured - enrichment, buying committee, scoring.