Victor Coimbra is CTO and Managing Partner at Artefact, where he leads AI, data science, and engineering work across the Americas. His career combines business strategy experience at Bain and Accenture with hands-on leadership delivering AI and machine learning systems in complex enterprise environments. He focuses on helping organizations connect technical capabilities with measurable business outcomes, including the architecture, governance, and operating models needed to scale AI.
Artefact is a global data and AI consulting company that helps enterprises develop strategies, build data platforms, deploy AI solutions, and manage organizational change. Its work spans industries including financial services, healthcare, manufacturing, and retail. Ahead of its Adopt AI Summit, taking place December 3–4, 2026, at the Grand Palais in Paris, Coimbra discusses how enterprises can move beyond pilots, build teams that combine people and AI agents, and measure the value of AI in production.
Your career has taken you from business strategy at Accenture to leading AI, data science, and engineering teams at Artefact across the Americas. How has this combination of strategic consulting and hands-on technical leadership shaped your approach to helping enterprises turn AI investments into measurable business outcomes?
It taught me to always connect AI strategy to delivery. Strategy at Bain and Accenture showed me that executives buy a change in a P&L line. Engineering taught me that a lot of that value can get lost between the slide and production.
Today AI is less about technology and more about building a moat around your core business operations. The key question is how to turn individual productivity into collective enterprise value, with a true shared cognitive capacity across the company.
That takes a mix of strategy and delivery, and it is something I learned a lot from doing both.
Despite significant investment in artificial intelligence, many organizations still struggle to move beyond successful pilot projects. In your experience, what are the most common obstacles preventing enterprise AI initiatives from reaching production, and what distinguishes companies that successfully scale their deployments?
In my experience, many pilots stall because AI is added to an organization that was designed for humans only. The technology usually works. What is often still open is the division of work: which tasks belong to agents, which stay with humans, who reviews, and who owns the result.
On top of that, 3 needs tend to slow the scale-up: token costs that explode with usage-based pricing, security and governance, and adoption that never turns access into real usage.
The companies I see scaling well treat AI as an organizational redesign. They split delivery into 2 loops. In the inner loop, agents execute end to end: code, build, test, push, review. In the outer loop, humans set the intent, verify the result and monitor cost and quality. Then they build one platform underneath and reuse it for every use case.
You’ve written about the importance of building an AI-native organization rather than simply adding chatbots and copilots to existing systems. What fundamental changes to enterprise technology architecture are necessary to support this transition, and where are businesses currently underinvesting?
An AI-native company runs on a platform, and that platform is the sum of 4 layers: Harness, Context, Model and Infra. The Harness is where humans and agents meet and work, the full application around the model. Context defines what agents know and what authority they have. Model is how models are accessed, routed and governed at the right cost. Infra is where they actually run, and who holds the keys.
Naturally, most of the attention goes to the visible 20%: copilots and chat tools. The hidden 80% is where most of the engineering lives, and where cost, security and control are decided.
My suggestion is simple. Buy generalist harnesses and build the industry-specific ones. Rent and experiment on models and infra, because they change every quarter. And own your context completely. It is the layer that is truly yours, and where I see the biggest opportunity to invest more.
As AI evolves from assisting employees to autonomously executing complex workflows, how should organizations determine which processes are appropriate for agentic automation? What technical and operational safeguards must be established before granting AI agents greater decision-making authority?
The way I see it, agents are very good at work that looks to the past, humans at work that looks to the future. Agents excel at anything built on existing information, like analytics, coding, testing, documentation and reconciliation. Humans own what needs creativity, and creativity by definition looks forward: taste, vision, planning, deciding, and judging whether a result is actually good.
In practice that maps to the 2 loops. Agents run the inner loop and execute. Humans keep the outer loop: intent, verification and monitoring.
The safeguards are engineered, not prompted. We use 6 harness layers: task capture, memory, a secure workbench with sandboxed runtime and permissions, a loop with checkpoints and clarifications, model routing, and guardrails that define what an agent cannot do alone, like human approvals, policy checks and PII controls. Autonomy then grows with the track record, the same way it does for a new hire.
You’ve described AI agents as teammates rather than conventional software tools. How does this perspective change the way organizations should structure teams, assign responsibilities, and train employees to collaborate with increasingly autonomous AI systems?
We are already in the hybrid organization. The question I find most important now is how to divide the workforce between humans and agents. I believe it is the key to scaling.
In a hybrid team, the org chart allocates cognitive capacity, not headcount. A typical split I see in a mature tech team is 15% human-only work, 25% augmented work and 60% done by agents. The Head of IT sets intent, allocates capacity and owns outcome and cost. A senior developer verifies agent deliveries and decides architecture. An orchestrator agent routes tasks and reports cost, and a review agent checks every delivery.
The gains compound in 3 steps. Augmented employees deliver around 5x. Agents as synthetic employees doing end-to-end work deliver around 10x. Orchestrating both as one workforce is where I believe the 1000x potential lies.
Training follows the same logic. People learn to brief an agent like a junior engineer, with structured context and clear outcomes, and to spend their time on taste, planning and decisions.
Enterprise AI systems depend heavily on reliable data infrastructure, yet many organizations continue to struggle with fragmented data, inconsistent governance, and legacy technology. What foundational capabilities should businesses prioritize when preparing their data platforms for agentic AI, and how can they modernize without disrupting existing operations?
Design your data platform for a new kind of consumer: agents. A dashboard with a wrong number gets questioned by a human. An agent with a wrong number acts on it, at scale.
Prioritize 4 capabilities. Shared business definitions, so every agent and team means the same thing by “active customer.” An integration gateway with authentication and standard protocols like MCP, so agents access data with the permissions of the role they act for. Observable quality and lineage. And unstructured data treated as a first-class asset, because contracts, tickets and meeting notes hold most of the real knowledge.
To modernize without disruption, wrap legacy systems first and migrate use case by use case. Agents also change the economics here. Multi-agent migration pipelines now modernize hundreds of thousands of lines of legacy code in months, with humans deciding the architecture.
Beyond access to structured data, AI agents increasingly require an understanding of institutional knowledge, business processes, and organizational context. How can enterprises effectively capture and integrate this knowledge into their AI systems, and what risks arise when agents operate without sufficient business context?
Context is what turns a model into a colleague. Tools understand tasks, agents understand context. You don’t ask a junior engineer to work without explaining the goal, the constraints and what good looks like, and an agent is no different.
Capture knowledge from the work itself, rather than from large documentation projects that are hard to keep up to date. Every meeting, ticket, decision and correction feeds a context layer that combines files, tables, vectors and a graph of relationships. Every finished task writes back to memory, so the organization learns with each delivery. And context needs its own governance: evals, permissions and cost control.
Without context, agents tend to struggle in a few predictable ways. They follow the written rule and miss the unwritten one everyone knows. They retrieve the wrong information and answer with total confidence. They work on last quarter’s facts. And a context layer without strict permissions can put sensitive data at risk.
Many companies measure AI success through productivity improvements, adoption rates, or the number of pilots launched. What metrics should CIOs and business leaders use to evaluate genuine AI return on investment, particularly when accounting for infrastructure costs, ongoing maintenance, and the operational complexity of autonomous agents?
Measure the hybrid workforce, not the tool. Adoption is only the first of 4 levels, always against a baseline taken before rollout. Adoption asks if people use it. Productivity asks if we deliver faster: lead time, throughput, resolution time. Quality asks if the output is better: escaped defects, change failure rate, rework. Value and cost ask if it is worth it: effort freed, time to market, full cost per task.
One metric that is easy to overlook is how cognitive capacity moves between humans and agents, team by team. Freed human capacity has to be redeployed to future-facing work with intent. Otherwise part of the gain gets lost.
On cost, pricing is moving from flat fees to usage, which can make bills harder to predict, so model routing and FinOps belong in the platform from day 1. Treat AI spend like cloud spend: visible, attributed to a team, and tied to value.
Artefact works with enterprises across financial services, healthcare, manufacturing, retail, and other industries where AI adoption presents different challenges. What lessons are emerging from real-world deployments in these sectors, and how should organizations balance innovation with regulatory compliance, reliability, and human oversight?
Every sector faces the same question: how to divide the work between humans and agents. What changes is how much of the outer loop must stay human. In retail and consumer goods, mistakes are cheap, so agents take more of the work faster. In manufacturing, reliability and integration with operational systems set the pace. In financial services and healthcare, every decision must be explainable, so humans keep more of verification and decision for longer.
The lesson I see everywhere is that compliance works best as a property of the platform, built in from the start. When guardrails, audit logs and human approvals live in the harness, every new use case inherits them. When compliance comes only at the end, each new project has to start that conversation again.
Oversight should be tiered. Agents act alone on low-risk tasks, get reviewed by exception on medium risk, and only prepare the decision on high risk. The tiers move as trust is earned.
With Artefact’s Adopt AI Summit taking place in Paris this December, discussions around AI sovereignty, governance, and industrial competitiveness are becoming increasingly important. How do you see Europe’s approach to AI adoption influencing global enterprise strategies, and what developments do you believe will define the next phase of enterprise AI over the coming years?
Europe was early in asking the sovereignty question, and I think it is becoming a global design principle. Multinationals don’t want one architecture per regulation, so they design for the strictest one: own your context, rent your models, and keep control of where they run and who holds the keys. The next opportunity for Europe is to bring the same energy to adoption, because competitiveness will come from usage at scale.
The next phase is the hybrid organization, and I believe it is the future of work. Companies will move from counting headcount to managing a shared cognitive capacity between humans and agents. Agents will take most of the work that looks to the past: analysis, code, reporting. Humans will move to the work that looks to the future: taste, vision, planning and decisions.
I believe the companies that solve the division of work first will lead the way. That is the conversation I want to have in Paris to define the next years.
Thank you for the great interview, readers who wish to learn more about enterprise AI adoption, agentic workflows, and data transformation should visit Artefact.
