Artefact CTO Victor Coimbra: Agents Do 60% of Work in Mature AI Teams

A senior AI executive says mature tech teams already split 15% of work to humans alone, 25% to augmented employees and 60% to AI agents — and he puts a number on the upside: 1,000x when humans and agents run as one workforce.

Victor Coimbra, CTO and Managing Partner at Artefact, a global data and AI consulting firm, laid out the figures ahead of the company's Adopt AI Summit, scheduled for December 3–4, 2026, at the Grand Palais in Paris. Coimbra leads AI, data science and engineering across the Americas, after earlier strategy roles at Bain and Accenture.

Why do AI pilots stall?

The technology usually works, Coimbra said. The problem is the organization around it. "Many pilots stall because AI is added to an organization that was designed for humans only," he said. What's missing is the division of work: which tasks go to agents, which stay with humans, who reviews, and who owns the result.

He named three needs that slow scale-up: token costs that explode under usage-based pricing, security and governance, and adoption that never converts access into real usage. Companies that scale treat AI as organizational redesign. They run two loops — agents execute end to end in the inner loop (code, build, test, push, review), while humans in the outer loop set intent, verify results and monitor cost and quality.

What architecture does an AI-native company need?

Coimbra described the platform as four layers: Harness, Context, Model and Infra. The harness is where humans and agents work; context defines what agents know and their authority; the model layer handles routing and governance at the right cost; infra is where models run and who holds the keys.

Most attention goes to the visible 20% — copilots and chat tools. The hidden 80% is where engineering, cost, security and control live. His advice: buy generalist harnesses, build industry-specific ones, rent and experiment on models and infra because they change quarterly, and own your context completely. "It is the layer that is truly yours, and where I see the biggest opportunity to invest more," he said.

Who does what — agents or humans?

"Agents are very good at work that looks to the past, humans at work that looks to the future," Coimbra said. Agents handle analytics, coding, testing, documentation and reconciliation. Humans keep taste, vision, planning and judgment.

Safeguards, he stressed, are engineered, not prompted. Artefact uses six harness layers: task capture, memory, a secure sandboxed workbench, a loop with checkpoints, model routing, and guardrails covering human approvals, policy checks and PII controls. Autonomy grows with track record, the same way it does for a new hire.

The gains compound in steps: augmented employees deliver around 5x, agents working end-to-end around 10x, and orchestrating both as one workforce is where he sees the 1,000x potential. In his example hybrid team, a Head of IT owns outcome and cost, a senior developer verifies agent deliveries, an orchestrator agent routes tasks and a review agent checks every delivery.

How should data platforms change?

Design them for a new consumer: agents. "A dashboard with a wrong number gets questioned by a human. An agent with a wrong number acts on it, at scale," Coimbra said. He prioritized four capabilities: shared business definitions, an integration gateway with standard protocols like MCP, observable quality and lineage, and unstructured data treated as a first-class asset. Multi-agent migration pipelines, he added, now modernize hundreds of thousands of lines of legacy code in months.

Context carries risks. Agents without it follow the written rule and miss the unwritten one, retrieve wrong information and answer with total confidence, or work from last quarter's facts — and a context layer without strict permissions can expose sensitive data.

How do you measure ROI?

Measure the hybrid workforce, not the tool, across four levels against a pre-rollout baseline: adoption, productivity (lead time, throughput, resolution time), quality (escaped defects, change failure rate, rework), and value versus full cost per task. Coimbra also urged treating AI spend like cloud spend — visible, attributed to a team, tied to value — with model routing and FinOps in the platform from day one.

Sector lessons differ: retail moves fastest because mistakes are cheap; manufacturing is paced by reliability and system integration; financial services and healthcare keep humans in verification longer. Compliance, he said, works best as a platform property, with tiered oversight that loosens as agents earn trust.

On Europe, Coimbra said sovereignty is becoming a global design principle — multinationals design for the strictest regulation: own your context, rent your models, control where they run. "I believe the companies that solve the division of work first will lead the way," he said. "That is the conversation I want to have in Paris to define the next years."