In 2026 alone, monday.com's AI agents handled tens of thousands of leads, booked thousands of meetings, and generated millions of dollars in pipeline. Demo request response times dropped from 24 hours to under 2 minutes. This is real AI GTM in production at a $1B+ ARR company.

I've been following AI GTM closely — I wrote a 60-page report on it earlier this year — but what doesn't get talked about is what happens when these plays collide with intricate routing logic, global rollouts, legacy tech stacks, and real compliance requirements.

Then I got connected to Oran Akron, the VP of AI for GTM at monday. Oran built monday’s RevOps org from zero to 80+ people over nearly 8 years. Last year he was tasked with creating an internal startup to rethink GTM workflows using AI agents (they call it RevAI). The scale here is wild: a publicly traded company growing 24% year-on-year with 1,000+ GTM employees.

I’ll show you three of monday’s most impactful AI GTM workflows, then unpack the messy work of building (and rebuilding) AI agents that work at enterprise scale.

Oran needed a wedge case where AI agents could meaningfully improve the existing GTM approach. He quickly honed in on speed-to-lead; the faster the response, the higher the conversion to qualified meeting.

The AI GTM team built “Amanda” – an inbound qualifying and hand-off agent. Their philosophy: sales reps shouldn't handle leads; they should only receive qualified meetings with the context already prepared.

Amanda is set up as a dedicated user in monday's CRM that gets assigned leads as they come in. Lead assignment triggers orchestration across 20+ preparation workflows including insight gathering, timezone resolution, agent versioning, accent matching, and local phone number selection — all within 60 seconds of a 'contact sales' CTA submission.

From there, Amanda executes voice calls that last about 5 minutes on average. These are deployed in English-speaking markets along with places like Brazil, France, and Germany. I tested a call from Amanda and was pretty impressed by the experience. Some of the nuances of the call execution:

Oran’s team was quite conservative in the initial deployment and A/B tested multiple iterations compared against manual outreach. Amanda has now organically spread across the org and well beyond the intended user base.

monday has a PLG motion with a large volume of free trial signups. “Jax” — their trial activation agent — appears as an avatar in the bottom right corner of the app, acts as a personal sales engineer, available to talk live at any point during the free trial, and can take actions on the product frontend to configure custom instances.

They’ve seen promising early results with over 3,000 monthly calls, 50% of users coming back for a second call, and an overall increase in conversion to paid. Jax users convert at 2.5x the rate of the control group, according to Oran.

The team wasn't sure how users would respond. But they're seeing high engagement, and users miss Jax when the trial ends. That's not something you hear often about a sales automation tool.

monday has an enterprise sales motion, too, and counts 60% of the Fortune 500 as customers. “Oscar” helps reps and territory managers go deeper on account strategy. Think: podcast listening, financial report analysis, internal platform usage data, combined into account plans with target departments, personas, and suggested outreach cadences.

Tasks that previously took reps 1–2 weeks now take ~5 minutes. Oscar is presented to reps as a single AI employee but is actually a set of orchestrated agents connected to Clay, LinkedIn (despite some frustrating API limitations), custom agentic solutions, and pre-AI enrichment tools the team had already licensed. It helps reps decide where the right opportunities are, how to stay current on those accounts, and how not to miss signals inside a company that can be turned into an opportunity.

It’s worth reiterating that monday has had an internal RevAI team (and, presumably, a very large token budget) devoted to AI GTM workflows since July 2025. This team has built and rebuilt their agent workflows at least four times. (Four times in under a year!)

Amanda launched as a single-prompt agent. It worked initially, then started hallucinating as complexity scaled. Latency constraints with the voice agent limited use of the most capable models.

The RevAI team moved to multi-agent orchestration with specialist sub-agents. These sub-agents handled discrete tasks like scheduling, CRM updates, lead ingestion, and the conversations themselves. Performance improved but consistency was still off.

The team made their most important architectural decision — stripping out AI reasoning wherever possible. The new concept: "smart tools and naive agents.” Deterministic flows were added everywhere they could get away with it (via n8n); AI only where reasoning was truly required.

Oran’s team added in-house microservices to stop hallucinations when scheduling meetings live on calls. This converted time zones based on where the prospect was based, updated available time slots accordingly, and booked meetings with 0% scheduling failures. The microservices layer simplifies agent responses, too – instead of “Error 153” the agents receive “That time isn’t available, here are three alternatives.”

This was critical for building organizational trust and operational stability. It was a lot faster, too. This shift allowed Amanda to scale to the ANZ and EMEA regions which brought new time zone and localization challenges.

monday then moved from UI-based prompt writing to agents-as-code using Cursor and Claude. This allowed for monitoring and evaluating Amanda’s calls in a templatized way (see below), then having AI agents autonomously iterate accordingly.

The RevAI team became worried about becoming married to their models, and couldn’t route to different LLMs in an auditable way. They shifted from model-specific, rigid prompts to simpler, objective-based prompting. This had the added benefit of enabling a more natural conversation flow with prospects.

The next design evolution, which Oran admits is still in early discovery, is going back to the original vision of one agent that could plan, act, and grow smarter over time. The technology is now catching up with a new generation of agent harness capabilities (Claude Code SDK, deep agents, modular skills).

This lets a single agent do what monday previously had to engineer around. It can plan ahead, spin up specialist sub-agents on the fly, keep its own working memory across long jobs, and reach into a library of skills the way that senior ICs pull a playbook off the shelf. The RevAI team is going back to the original vision with a deep agent backbone that keeps upgrading existing and future agents.

Three things stand out from monday's experience:

Don't get married to your stack. monday's current stack includes n8n (orchestration), Salesforce (CRM), ElevenLabs (voice), Twilio (telephony), SendGrid (email), ChiliPiper (scheduling), Clay (data), Claude (LLM), and Cursor (AI coding). That's a lot of vendors — several of those weren't in the original build and the vendors probably keep changing as I write this. Build versus buy becomes a constant decision; vendor lock-in can hold your whole effort back.

Deploy at 60% readiness. monday faced real challenges with reliability and hallucinations. The unknown variables in an AI implementation are endless. Continuous iteration is the only path forward. Optimize simultaneously for results and learning.

Over-communicate internally. In monday’s playbook were gradual rollouts by cohort, cultivating internal AI champions, and being transparent about the performance data of all AI agents. They added smaller touches, too, like programming Amanda to send funny voice notes to reps summarizing the meetings that were booked. The change management work was as hard as the technical work.

In Oran’s view, AI should eliminate admin work, qualification, research, and meeting preparation – freeing up reps to focus on relationship-building, strategy, and enterprise selling. If your first (or second) attempt at AI GTM workflows didn’t pan out, consider trying again.