Duvo Series A Raises $32M, but the Hard Problem Is Hidden Work

Duvo raised a $32 million Series A to automate enterprise operations, but its central challenge is not simply connecting AI agents to old software. The Duvo Series A backs an attempt to capture the undocumented decisions that employees make between ERP screens, spreadsheets, supplier portals, and email.

Balderton Capital led the round, while existing investors Index Ventures, Northzone, and Credo Ventures participated. The financing brings Duvo's total funding to $48 million and supports a United States expansion focused initially on retail and consumer packaged goods companies.

The tension is easy to miss. Enterprise automation vendors have spent years promising to connect fragmented systems. Duvo argues that the larger obstacle sits outside those systems because crucial rules, exceptions, and workarounds remain in employees' heads.

That thesis puts Duvo against a familiar implementation model. Companies usually document processes through internal teams, consultants, or forward-deployed engineers before software can automate them. Duvo wants its product to perform much of that discovery itself, then run the resulting process with human approvals.

The Duvo automation platform therefore combines two difficult jobs. It must understand how work really happens, including the exceptions, and execute that work reliably across software it does not control.

The Duvo Series A Funds a Push Beyond AI Pilots

The new capital gives Duvo room to test whether one automation layer can support many operational processes without becoming another collection of isolated projects.

Duvo announced the financing on October 14, 2026. Its funding announcement says the company will use the round to expand in the United States, beginning with retailers and consumer goods businesses.

That market choice reflects Duvo's origins. Co-founder and CEO Tomáš Čupr also founded Rohlik Group, a European online grocer. Retail operations exposed him to workflows spread across inventory systems, supplier communications, financial tools, and locally maintained spreadsheets.

These workflows rarely behave like clean software diagrams. A buyer might compare an email with an ERP record, review a contract term, contact a supplier, and escalate an exception. Another employee with the same job title may follow different rules because that person's suppliers operate differently.

Duvo describes its product as an Enterprise Execution OS. The term refers to a shared layer where people and AI agents can discover, run, approve, and monitor processes across existing enterprise applications.

The company is not proposing that customers immediately replace their ERP or warehouse management systems. Instead, Duvo aims to sit above them. Its agents use APIs when available and interact with existing user interfaces when a system lacks a suitable connection.

This distinction matters because most large companies cannot treat core software replacement as a casual experiment. Their systems hold financial records, supplier data, inventory positions, and approval controls. Even an outdated application can remain essential because many other processes depend on it.

Duvo says its platform can establish one set of connections, approval rules, and operational records for multiple processes. If that claim holds, a second automation should reuse more infrastructure than the first.

That is a more consequential goal than delivering a successful demonstration. A pilot can succeed with intensive engineering support and carefully selected cases. A platform must continue working as interfaces, personnel, policies, and business conditions change.

The Duvo Series A also funds a more direct commercial push. Michael Green, previously an executive at Databricks and Salesforce, joined to lead commercial expansion. Spencer Insolia, a former Kmart president and Accenture managing director, is leading the United States business.

Those appointments suggest that enterprise distribution is now as important as product development. Duvo must enter complex organizations, identify valuable processes, satisfy security reviews, and show measurable returns without allowing deployment work to consume its economics.

The funding is therefore not just a vote on agent capabilities. It is a bet that process discovery, implementation, governance, and measurement can become repeatable parts of one product.

Why Undocumented Knowledge Blocks Enterprise Automation

Duvo's most important argument is that enterprise processes are incomplete until the people performing them explain the exceptions.

Čupr reportedly reached that conclusion while trying to introduce agents at Rohlik. Employees could identify repetitive work, but documenting the entire process proved harder. Many decisions relied on experience that had never entered a manual, database, or workflow tool.

In an interview about the human API problem, Čupr said people often function as interfaces between disconnected systems. They read one screen, remember an unwritten rule, interpret a message, and enter the result somewhere else.

That pattern helps explain why conventional automation frequently handles the predictable center of a process but leaves its messy edges to employees. A script can copy a known field. It struggles when a supplier partly confirms an order while disputing the price in the same email.

Duvo Clarity, the company's discovery component, is designed to capture that missing context. According to Duvo, it combines AI-led employee interviews, screen recordings, and existing documents to map how work happens in practice.

The process map must include more than a sequence of clicks. It needs to identify decision owners, approval requirements, common exceptions, and the evidence employees use when choosing an action.

That turns process discovery into a knowledge problem. A company first needs to find what employees know, distinguish policy from habit, and determine which judgments can safely become rules.

The same challenge appears in personal and team knowledge systems. Documents provide only part of the answer when important context remains scattered across conversations and individual memory. A searchable knowledge base becomes useful when it preserves context rather than merely collecting files.

Duvo applies a similar principle to operational execution. It seeks to turn tacit knowledge into a controlled process that an agent can follow, with people retained at defined decision points.

The mechanism has three broad stages. First, the system observes and documents the current workflow. Second, it identifies steps suitable for automation and those requiring approval. Third, agents perform approved actions across connected systems while recording what happened.

That last stage separates process documentation from process execution. Consulting engagements and process-mining tools can reveal inefficiencies, but another project may still be required to implement the changes.

Duvo wants discovery and execution to share one system. That connection could shorten the distance between finding a problem and changing the operation.

It also introduces risk. Employee explanations can be incomplete, inconsistent, or shaped by local workarounds. Screen recordings may show what someone did without revealing why that choice was correct.

An automated process can therefore formalize a bad habit as easily as a good one. Management must decide whether the discovered workflow represents company policy, a necessary exception, or technical debt that should disappear.

Duvo says human owners approve decisions that require judgment. Operating dashboards show pending decisions and completed outcomes, while system changes are recorded for audit purposes.

This human-in-the-loop design, meaning people approve selected agent decisions, is essential for sensitive operations. It also creates a practical question about scale. If too many cases require review, the agent merely relocates the bottleneck.

The strongest version of Duvo's thesis is not that employees should disappear from the process. It is that software should capture repeatable knowledge while directing genuinely ambiguous decisions to the right person.

Success depends on that boundary staying accurate as the business changes. A promotion rule, supplier agreement, or risk policy that worked last quarter may not remain valid. Duvo must keep the process map current rather than treating discovery as a one-time event.

Duvo Enterprise AI Challenges the Services-Heavy Model

Duvo is betting that enterprise AI deployment can become product-led before implementation labor overwhelms the value of the software.

Large organizations often use consultants and implementation specialists to understand operations before changing them. Those teams interview employees, map systems, document approvals, and coordinate new software with existing controls.

Forward-deployed engineers follow a related model. They work closely with customers to adapt software, integrate data, and build reliable workflows inside complicated environments.

That approach can deliver results because humans absorb ambiguity. It can also make deployments slow, expensive, and difficult to reproduce across customers.

Duvo's primary challenge to this model is straightforward: the product should perform more of the discovery and configuration. AI interviews can gather process knowledge, agents can build workflow maps, and reusable controls can reduce custom integration.

Čupr told Pathfounders that Duvo aims to compress process discovery from months into weeks. He also argued that well-designed enterprise AI should become increasingly self-deployable.

Duvo does not frame every consulting firm as an enemy. It has an alliance with Deloitte Central Europe, which combines consulting expertise with Duvo's process mapping and execution software.

That partnership reveals the likely near-term reality. Consultancies can become distribution and implementation partners even when software reduces the labor required for certain engagements. Their strategic and organizational work can remain valuable while the tooling changes how process discovery occurs.

The pressure falls most directly on repetitive implementation labor. If Duvo can capture workflows faster, customers will question long projects devoted mainly to interviews, diagrams, and manual configuration.

The company must still prove that the compressed approach produces an accurate result. Speed has limited value if a missed exception later disrupts inventory, supplier payments, or financial reporting.

Competition will make that proof harder. The enterprise automation market already includes robotic process automation vendors, process-mining platforms, workflow products, and newer agent companies.

Pathfounders identified Orby AI, Skan AI, Mimica, Sola, and Scribe among relevant competitors. These companies do not all offer identical products, but several address workflow observation, process intelligence, or agent-driven execution.

Traditional robotic process automation, commonly called RPA, provides another reference point. RPA tools often reproduce structured user actions, such as copying information between applications. They can be effective when inputs and interfaces remain predictable.

Duvo argues that AI agents can handle more ambiguity than rigid scripts. Its agents reportedly interpret emails, navigate screens, compare records, and escalate decisions rather than following one fixed click path.

The company has described a legacy systems deployment using Anthropic's Claude, browser interaction, tool connections, and human approvals. Duvo says computer-use capabilities allow agents to operate systems without modern APIs.

That flexibility expands the addressable work, but it does not remove the reliability problem. Graphical interfaces change. Access permissions expire. Data formats vary. A screen that looks obvious to a person can confuse an automated system after a minor redesign.

API-based integrations have their own costs, yet they usually provide clearer contracts than visual interaction. Duvo must decide when interface automation is acceptable and when a deeper integration remains necessary.

Model dependence presents another issue. An agent may behave differently after an underlying model update, especially on ambiguous inputs. Enterprises will expect Duvo to test those changes before they affect production decisions.

Duvo says its architecture is model-agnostic. Customers can reportedly use approved cloud endpoints, dedicated infrastructure, or open-weight models where security and sovereignty requirements demand more control.

That positioning also creates operational complexity. Different models vary in reasoning quality, latency, deployment requirements, and behavior. Supporting several options does not guarantee equal performance across them.

The Duvo enterprise AI proposition will therefore compete on more than intelligence. Buyers will compare implementation effort, auditability, security, exception handling, maintenance, and measurable business outcomes.

The decisive comparison is not simply Duvo versus consultants or Duvo versus RPA. It is productized deployment versus labor-intensive customization. Duvo needs to show that its product can absorb complexity without quietly recreating a services business behind the scenes.

The Customer Results Are Promising, but Mostly Company-Reported

Duvo has published concrete operating improvements, yet buyers still need independent evidence about durability, total cost, and exception rates.

Duvo says it has generated more than $100 million in value for clients across finance, supply chain, commercial operations, customer care, marketing, and human resources. Its named customers include Morrisons, Holland & Barrett, Boden, Notino, Allegro, and Rohlik.

The company reports that Notino's reconciliation of supplier bonuses, covering approximately $45 million annually, fell from 70 days to 15. Specialist work reportedly declined from about 120 hours per cycle to fewer than 10.

Duvo has also published detailed claims from Rohlik. It says one deployment generated €2.8 million in annualized savings within three months and entered production eight weeks after the first discussion.

According to the company, product availability in that case rose from 78 percent to 93 percent in two weeks. Inbound delivery confirmations reportedly increased from 52 percent to 90 percent.

The same case study says annual negotiations became about 80 percent automated, with more than 60 full-time employees redirected from data entry toward strategic negotiation. Duvo also reports faster margin reviews and a shorter month-end close.

These figures make the Duvo automation platform easier to assess than a vague claim about productivity. They connect the technology to inventory availability, supplier negotiations, reconciliation, and working time.

However, most of the detailed metrics come from Duvo or customer statements selected by Duvo. They have not been presented as results from an independent controlled evaluation.

That does not make them false. It means readers should treat them as reported case-study outcomes rather than universal performance benchmarks.

Several questions remain unanswered. The published material does not fully separate value created by Duvo from process redesign, management attention, or other operational changes made during deployment.

Annualized savings also differ from realized cash savings. A new process may identify an opportunity without guaranteeing that every identified amount reaches the income statement.

Buyers should ask how Duvo establishes a baseline, attributes gains, and accounts for the employees involved in implementation and review. They should also examine how results change after the initial process expands across countries, teams, and software configurations.

Reliability requires similarly specific measurement. Completion rates alone are insufficient if agents choose incorrect actions or send too many cases to human reviewers.

Useful operational metrics would include exception frequency, approval rates, error severity, rollback frequency, and the time employees spend supervising each process. Those figures would show whether the product removes work or changes its form.

Security and governance deserve equal attention. Duvo says it records approvals and changes written to customer systems. It also reports holding SOC 2 Type II, ISO 27001, and ISO 42001 certifications.

Certifications provide evidence that defined controls exist, but they do not guarantee that every automated decision is correct. Customers still need process-specific permissions, monitoring, testing, and escalation rules.

The broader agent market gives buyers a reason for caution. Gartner has predicted that more than 40 percent of agentic AI projects will be canceled by the end of 2027. Its agent project forecast cites escalating costs, unclear business value, and inadequate risk controls.

Duvo explicitly positions itself against those failure modes. It emphasizes measurable results, governed approvals, reusable connections, and deployment across systems customers already operate.

The company must nevertheless face the same economic pressures. Model calls, browser sessions, support, monitoring, and human review all contribute to the real cost of an automated process.

A workflow that creates meaningful value can justify those costs. A low-value process with frequent exceptions may never become economical, regardless of how capable the agent appears in a demonstration.

Customer retention will provide a stronger test than early case studies. Expansions into additional departments would suggest that the shared platform produces compounding value. Narrow renewals would imply that performance depends on unusually favorable processes.

The Duvo Series A gives the company time to collect that evidence. It does not settle the question of whether the results remain stable across customers, industries, and changing operating conditions.

Three Signals Will Show Whether Duvo Can Scale

Duvo's next test is whether its process-discovery advantage produces repeatable deployments, broader customer adoption, and sustainable operating economics.

The first signal is expansion within existing customers. Duvo says one platform can support processes across finance, supply chain, customer care, marketing, and human resources.

Buyers should watch whether customers move from one successful workflow into several departments. That expansion would support the argument that connections, controls, and process knowledge can be reused.

The quality of those expansions matters as much as the number. Deploying several similar supplier workflows is easier than crossing into financial approvals, workforce processes, or regulated customer decisions.

Broader deployments would strengthen Duvo's platform claim. If most customers remain limited to one narrow use case, the product may behave more like a specialized automation project.

The second signal is evidence from the United States rollout. Duvo has hired experienced commercial leaders and is targeting retail and consumer goods companies, where fragmented supplier and inventory systems create obvious automation opportunities.

United States customers may move faster during evaluation, but enterprise contracts still require security, legal, integration, and procurement reviews. Duvo must show that faster process discovery also shortens the entire path to production.

Case studies should disclose deployment time, implementation staffing, exception rates, and realized outcomes. Independent customer accounts would carry more weight than additional aggregated value claims.

The third signal is how Duvo's model holds up against competitors and consulting partners. Rivals in workflow discovery and agent automation will continue improving, while established automation vendors can add generative AI to their installed platforms.

Consultancies also have options. They can distribute Duvo, adopt competing products, build internal tools, or retain more manual discovery where clients value human analysis.

Duvo's alliance strategy will be tested when partners decide whether the platform expands their business or compresses profitable work. A repeatable partner channel would support Duvo's claim that software can improve consulting delivery without eliminating strategic services.

Technical evidence will matter within all three signals. Duvo should demonstrate how its agents respond when a screen changes, an employee gives conflicting instructions, or an underlying model behaves differently.

Buyers should also watch the ratio between automated decisions and human approvals. Rising automation with stable error rates would strengthen the thesis. Growing review queues would weaken it.

The most revealing metric may be time to the second process. Duvo says the first deployment can go live within weeks. If later processes become materially faster, the shared execution layer is creating the promised leverage.

If every new workflow requires another extensive discovery and engineering cycle, the economics will look closer to conventional enterprise implementation.

Duvo has chosen a difficult but important target. The work between enterprise systems is expensive precisely because it combines fragmented data, human judgment, and organizational history.

The Duvo Series A does not prove that agents can absorb all that complexity. It funds a serious attempt to turn hidden operational knowledge into governed software without replacing the systems beneath it.

Enterprise buyers should now ask a practical question: does Duvo reduce the total effort needed to understand, automate, and maintain a process after the demonstration ends? Track repeat deployments, independent outcome data, and human review rates over the coming months. Those signals will show whether Duvo enterprise AI is becoming durable infrastructure or another promising pilot that depends on exceptional support.