Michael Jaszczyk is CEO of NEWWORK Software, which provides a digital workforce platform that makes AI usable for enterprise organizations.getty​Picture an automotive manufacturer building an engine: The transmission team is world-class, and so is the engine block team. Each group designs the best version of its component and meets every specification it was given.​Then on the assembly line, a parts assembler tries to bolt the two together, but the mounting holes don’t line up. As it turns out, nobody established the proper interface mechanism before either team began building.​An auto manufacturer can recover from this. It’s one company with one engineering organization, and someone has the authority to send both teams back to redesign the mounting. Enterprises are experiencing the same problem with AI right now, but they do not have this option.​Their engine block came from one software vendor, and the transmission from another. Those vendors follow different roadmaps, ship on different update cycles and answer to many other customers. Neither is going to redesign its AI around the way one business happens to operate.​For enterprises, the interface between systems cannot be retrofitted into components they did not build. It must exist above them. Only then can it understand the whole business. Otherwise, it doesn’t really exist at all.​Capable Agents With Partial Pictures​Gartner predicts that 40% of enterprise applications will be integrated with task-specific AI agents by 2026, up from less than 5% in 2025. That is an extraordinary amount of automation arriving inside systems that vendors already own, which is a good thing. After all, vendors understand their own domain better than anyone, and the agents they ship reflect that.​The cross-system limitations show up at the edges: An agent inside a CRM has a detailed model of the sales pipeline yet no awareness of the finance policy that governs what can be promised to a customer, or an agent inside an HR system knows employment records but has never seen the service commitment that a sales team member made last quarter.​Each agent operates with total confidence but exists inside a partial picture of the organization.​Very little enterprise work fits within one system. Onboarding a new customer, closing a quarter or resolving a service escalation are processes that often move across four or five platforms and several departments before they are considered finished. Specialized intelligence at each stop produces a series of confident, locally correct handoffs, but a human is still required to stitch it all together properly. ​Why Connection Is Not Coordination​A natural counterargument here is that businesses and vendors are already solving this. Open protocols for agent interoperability are maturing quickly, and Forrester predicts that 30% of enterprise application vendors will launch their own Model Context Protocol (MCP) servers, enabling outside agents to work with their data.​These protocols can deliver access to data and functions in neighboring systems, subject to existing permissions. But what access leaves untouched is everything that must be decided by a human when two systems disagree.​If a CRM says a customer qualifies for a discount and the finance system says the account is on credit hold, no MCP protocol arbitrates this. A human has to step in and determine which record has the ultimate authority. This critical judgment layer, sitting above fragmented system applications, is where true promise lies for enterprise AI.​Four Things No Single System Can Orchestrate​The coordination interface that enterprises need ultimately comes down to four decisions, none of which any individual vendor’s AI can make on its own:​• Authority: Which system is the source of truth for certain data, and what happens when two conflict? Most organizations haven’t determined this, which means work is stopped until someone intervenes and decides which system takes precedence.• Context: Understand the organization’s specific policies, thresholds and regional requirements. These provide the roadmap that shapes how work should move throughout the enterprise. This knowledge lives in company documents, approval chains and institutional memory rather than any single application’s data.• Escalation: When a cross-system process requires a human decision, the coordination layer must know who that person is. For example, an agent operating inside one application can raise a flag, but it may not know that a $50,000 exception requires approval from a regional manager who doesn’t use the application.• Evidence: To understand how work moved across the enterprise, an audit record must exist that can reconstruct what happened after the fact. Partial logs in existing applications cannot reconstruct the whole process or help automate the manual work that AI promises to eliminate.​Each item here represents a cross-application boundary that agents struggle to cross on their own, highlighting why the actual system controlling how work progresses must live above existing applications. Waiting for vendors to solve the interface problem is a losing position.​Identifying What Is Already Defined​Gartner also projects that over 40% of agentic AI projects will be canceled by the end of 2027. Some of this is ordinary attrition in a hyped AI market, but a meaningful share will be projects that should work as designed yet are abandoned because nothing could carry the outcome forward.​So how should enterprises take the first step at solving the interface problem?​Before approving the next embedded agent, start with one process that already crosses departments or systems. Try to document those four decisions for that specific workflow: name the system of record for each disputed fact; gather the company policies that govern the process; identify the human who owns the exceptions; and specify where the audit trail should live.​Most organizations that attempt this realize few, if any, of the four have been formally defined. Instead, employees are informally compensating for the gaps.​Automated coordination and decision-making can then be added to the mix once these decisions have been made, formalized and documented. Until these deliberate decisions are defined and enforced in a coordination layer above individual systems, every new agent is just one more excellent component with mounting holes that don’t line up.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?