Visma · Holded Technologies

Making AI Work

The Operational Foundations
Behind Real Impact

Constantinos Yenis

VP of Revenue Operations
Holded Technologies SL

2025

Don't think outside the box.

destroy

Two Ways to Use AI

Adding AI to the box is not the same as
removing the box.

AI-assisted

Improve the existing workflow.
Add a copilot. Add a chatbot.
Same stack. Faster.

Thinking outside the box.

AI-native

Redesign the workflow from first principles.
Question the stack itself.
Fewer tools. More control.

Thinking like there is no box.

Both are valid. But the bigger unlock comes from rethinking, not just improving.

A Real Example

Our website. Two approaches.

AI-assisted

Webflow
Figma
OneTrust
Google Tag Manager
AirOps
AI personalisation
AI chatbot
7tools · 7 vendors · 7 contracts
→

AI-native

Claude Code

1 toolFull controlFull ownership

Same outcome. Fundamentally different operating model.

The Central Idea

AI is not magic.
It is an outcome multiplier.

×3

24

PRs reviewed(was 8)

135

Chats resolved(was 45)

9

Demos booked(was 3)

36

Reports generated(was 12)

75

Leads enriched(was 25)

The question is: what is your multiplier?

The Leverage Question

The curious get more leverage.

Better questions lead to better prompts.
Better prompts lead to better outcomes.
AI rewards the people who ask more, test more, and iterate faster.

Curiosity is the new seniority.

Plot Twist

The most senior person in the room
might not win anymore.

PRs reviewed / week

Senior · struggles with AI

8× 2
=16

Junior · AI-fluent

3× 10
=30

Same people. Same company.
Completely different outcomes.

Try this today.

> “I would like to multiply my outcomes with AI as a Your Title working in Your Team at Your Company. Help me be world-class at my job and be relevant in 2027.”

ChatGPT · Claude · Gemini · Perplexity

ClaudeClaude
PerplexityPerplexity

Then actually read what it says.

But experimentation alone
is not enough.

Cool demos are not the same as
repeatable business impact.

Most companies are hype-ready.
Very few are production-ready.

Alpha

Experimentation. Trying things. Learning.

Beta

Controlled pilot. Measured. Owned.

Live

Production. Trusted. Accountable.

“Be ready for something before you need it.”

The Real Problem

AI does not fail because the model is weak.

It fails because the workflow is.

Broken process in → broken output out

Poor context → hallucinations

Disconnected systems

Unclear ownership

No feedback loop

The Title of This Talk

Real AI impact sits on

operational foundations.

01Process
02Data
03Context
04Systems
05Ownership

The model is the engine.

Context and memory are the fuel.

Model

The reasoner

Powerful out of the box.
Blind without a briefing.

Context

The briefing

What you give it right now:
data, notes, docs, history.

Memory

The compound

What it learns over time.
This is where AI compounds.

Context in Practice

Your second brain is the context layer.

Obsidian vault as shared memory. Two tools, one brain.

Claude Code

Claude Code

The operator. Interactive, in the loop.
Conversations, decisions, creative work.

OpenClaw

OpenClaw

The cron job. Async, automated.
Sweeps, alerts, vault hygiene.

ObsidianObsidian
Claude CodeClaude Code
OpenClawOpenClaw

The vault is the shared brain. Claude Code does the thinking. OpenClaw does the housekeeping.

The Context Port

MCP is the USB-C port for AI.

One standard way for Claude to plug into the systems where work already happens. CRM in HubSpot. Tickets in Intercom. Product signals in Amplitude.

HubSpotHubSpot

CRM context

AmplitudeAmplitude

Product analytics

→→

Model Context Protocol

MCP

Claude plugs in once

←←
IntercomIntercom

Tickets and conversations

ClaudeClaude

AI reasoning engine

Same conversation. Same model. But now it can read, search, and act inside real business systems instead of guessing from a blank chat window.

Claude demo next

MCP in Practice

Sales without opening the CRM.

HubSpotHubSpot
SalesforceSalesforce
ClayClay
IntercomIntercom
ClaudeClaude

Before a call

▎ Pull up everything on Acme Corp: last 3 activities, deal stage, decision maker, and any open support tickets

Claude reads CRM plus Intercom via MCP and returns a five-second briefing. No tab-switching. No timeline archaeology.

After a call

▎ Log this: spoke with Maria, CFO. They are evaluating Q3. Budget approved, need security review. Move to stage 3 and set a follow-up for April 10th

Claude writes back via MCP: updates the deal, adds the note, creates the task. The rep never opens HubSpot.

Enrichment + outbound

▎ Find 20 companies in DACH, 50-200 employees, using Intercom. Draft a personalised first touch for the top 5 based on their latest funding round

Claude queries enrichment tools, drafts outreach, and keeps it in one workflow instead of five disconnected ones.

Pipeline review

▎ Show me all deals stuck in stage 2 for more than 14 days with no activity in the last week

Claude returns the table. Manager replies: send the owners a nudge in Slack. Done through MCP.

Before

CRM-centric

After

AI + MCP

Rep opens 5 tabs

Rep opens Claude

Manually logs notes

Dictates, Claude writes to CRM

Searches CRM for context

Asks Claude, gets a briefing

Manager builds reports in dashboards

Asks Claude, gets answers

Data entry = compliance burden

Data entry = byproduct of conversation

What actually changes

The CRM becomes the database, not the interface. Reps stop typing into forms after the fact. Data gets captured in context, while the conversation is still alive.

Live Demo

AI Account Executive in action

HubSpotHubSpot
IntercomIntercom
ClaudeClaude

Take these home.

Templates from what we built.

Real projects. Copy the prompt. Adapt to your stack.

Copy / Paste Template

Omnichannel Support AI Agent

Scalable customer support across chat, voice, and forms

~80%

of support interactions
handled by AI.

91%

CSAT maintained

~€2M

Annual savings

6

Channels covered

Workflow

1.

Customer asks on chat, voice, or form

2.

AI classifies intent and sentiment

3.

Pulls context from KB or CRM

4.

Answers, routes, or escalates

5.

Logs transcript and feeds QA loop

Intercom Fin AI · ElevenLabs Voice · n8n · Help Center RAG

> “Help me build this for my company: an omnichannel support AI agent that handles inbound messages from Intercom. Classify intent, draft responses from a KB, pull CRM context from HubSpot, escalate complex issues, and offer ElevenLabs voice handoff for premium accounts. Include a QA feedback loop.”

Copy / Paste Template

AI QA Assistant for Support

Every DSAT reviewed by AI in real time \u2014 coaching delivered to Slack

IntercomIntercom
→
n8nn8n
→
SlackSlack

Workflow

1.

DSAT rating triggers Intercom webhook

2.

n8n validates, deduplicates, enriches data

3.

Routes by rating: bot DSAT vs human DSAT

4.

AI reads full conversation, scores agent

5.

Coaching + better approach posted to Slack

6.

All results logged to Google Sheets

What the AI Outputs

Summary of what happened

What the agent did well

Communication / Empathy / Efficiency scores

Specific coaching points

Better approach with exact suggested wording

Training recommendations + pattern risk flag

Copy / Paste Template

AI SDR Workflow

Automated prospecting from lead to booked meeting

Clay Enrich

Firmographic data

→

Claude Qualify

AI lead scoring

→

AI Outreach

Personalized sequences

→

ChiliPiper Route

Auto-book meetings

→

HubSpot Log

CRM sync

Clay · Claude Code · n8n · ChiliPiper · HubSpot

Zero manual researchPersonalized at scaleInstant booking

> “Help me build this for my company: an AI SDR workflow using n8n for orchestration. When a new lead enters Salesforce, auto-enrich with firmographic data, use AI to score and generate personalized outreach, adjust messaging based on engagement signals, and auto-book qualified leads via Chili Piper.”

Copy / Paste Template

AI Migration Coworker

Migrate data from previous accounting software into Holded

Claude CodeClaude Code
→
Holded MCPHolded MCP

Workflow

1.

Team defines migration task with source export (Excel) and Holded schema

2.

AI interprets fields, edge cases, and mapping logic via Holded MCP

3.

Suggests field mappings and validation checks automatically

4.

Human validates final decisions and approves the migration

5.

Executes migration with error logging and rollback capability

Why It Works

Team of 10 doing 2-3 migrations per day

AI handles field mapping — humans validate decisions

Custom Holded MCP connects Claude directly to the platform

> “Help me build this for my company: an AI migration coworker that helps teams migrate data from previous accounting or invoicing software into a new system. Accept source and target schemas, interpret field mappings, flag edge cases, suggest validation checks, let the human approve before execution, and log errors with rollback. Build it as an MCP server so it connects directly to the target platform.”

Copy / Paste Template

Intelligent Demo Routing

From broken routing to the right calendar in real time

Claude CodeClaude Code
→
ClayClay
→
n8nn8n
→
Chili PiperChili Piper
→
LovableLovable
→
HubSpotHubSpot

The Problem

×

High inbound volume, no proper qualification

×

Freelancers routed to live demos, wasting AE time

×

Qualified leads waiting 1u20132 weeks for availability

×

Partners lost in generic routing

×

No fallback when demo slots are full

The New Flow

1.

Lead fills form → Clay enriches in real time

2.

Qualification score determines routing path

3.

High-qualified → live demo via Chili Piper

4.

Lower-qualified → next webinar on Livestorm

5.

No availability? Automatic webinar fallback

6.

Partners → routed to partnership team directly

+

Now testing: AI Demo Agent— an AI that runs the first product demo live. I'll show you right after this.

Copy / Paste Template

AI Call Intelligence

Turn every sales call into structured, actionable data

Workflow

1.

Call transcribed in Gong

2.

Webhook triggers n8n automation

3.

AI agent extracts structured insights

4.

Results stored in tables and CRM

5.

Feeds dashboards, coaching, and strategy

What AI Extracts

Objections raised by the buyer

Key selling points that landed

Buyer persona and decision stage

Next steps and commitments made

Competitive mentions

Gong · n8n · AI Agent · HubSpot / Airtable

> “Build an AI call intelligence pipeline. When a call is transcribed in Gong, send it via webhook to n8n. An AI agent extracts objections, key selling points, buyer persona, decision stage, next steps, and competitive mentions. Store structured output in tables and feed dashboards for coaching and strategy.”

And More We Shipped

mdOS & Openclaw

Too many models, no routing → Right LLM for the right task

OpenClawOpenClaw

New Website

Webflow + 5 vendor dependencies → One Claude Code build

Claude CodeClaude Code

AI Account Executive

Static demo scripts → Live AI sales conversations

HubSpotHubSpot
IntercomIntercom

The Pattern

What all of these have in common.

✓Clear workflow
✓Real use case
✓Structured context
✓Measurable output
✓Human oversight where needed

AI will not replace operational excellence.

It will reward it.

AI multiplies what is already there

Curiosity creates leverage

Operations make it repeatable

Small wins compound fast

The winners will not be the teams with the most AI tools.
They will be the teams that build the best multiplier.

Thank You

Constantinos Yenis

8 years at Holded · Visma

Leaving as an employee this month.
Continuing as external consultant.

WhatsApp: +34 600 204 608

“Be ready for something before you need it.”

LinkedIn QR code — Constantinos Yenis
Scan or click to connect on LinkedIn
01 / 28Constantinos Yenis

Speaker Notes

I'm not here to sell you on AI. You've already heard that pitch. I'm here to talk about what actually happens after you decide to use it — the operational reality, the wins, the failures, and the multiplier effect that separates teams that talk about AI from teams that ship with it.

← → navigate  ·  N notes  ·  F fullscreen