Why AI agents break at department handoffs - and the fix
The agent runs clean inside one team and stalls the moment the work has to cross into another team's lane.
Short answer: AI agents break at department handoffs because the process was drawn as department-owned swimlanes with a human passing the work across every boundary, and the agents inherited that same seam. Celonis's 2026 survey of roughly 1,600 leaders found 54% blame siloed teams and 44% blame poor coordination between departments. The fix is to redraw the process so one agent owns it end to end.
We have rebuilt cross-department workflows for operators, and the agent stalls in the same place almost every time: the border between two teams, where the demo never went.
Why do AI agents fail exactly where one department hands off to another?
Agents fail at the border between two departments because that border is where the old process handed the work to a person, and nobody taught the agent what the person quietly did there.
Think about how that work got built. It predates AI. Each department owns a lane, and at every boundary a human picks up the work, reconciles it, translates it into the next team's system, and passes it along. That translation was invisible labor. It never made it into a spec. VentureBeat, reporting on the Celonis data, describes the root cause: knowledge sits trapped inside departments that "developed their own languages and systems over time" and never shared a common understanding. So an agent reaching across the line is guessing at a step a person used to fill in by hand.
The survey data points at the same seam. Celonis's 2026 Process Optimization Report, based on roughly 1,600 business and operations leaders, found that 85% of businesses want to become an "agentic enterprise" inside three years while 76% admit their current processes cannot support one. The two blockers leaders named most were siloed teams, at 54%, and a lack of coordination between departments, at 44%. The wall stands between the teams. The model was never the bottleneck.
Is this an agent-quality problem or a process-design problem?
The process was broken before the agent arrived; the agent only made the break visible.
Salesforce put this bluntly when it launched Agentforce Operations. SVP of Product Sanjna Parulekar told VentureBeat that "the brokenness in a process is probably in your product requirements document". Her point: before an agent can run a process, you have to break that process into explicit, deterministic steps. Where a human used to improvise across a handoff, there was no step written down at all. The agent hits that gap and stops.
So the diagnostic question answers itself. If your agent performs inside a function and fails across a boundary, the process was never fully specified, because a person was covering the unspecified part by hand.
What is a swimlane, and why does it break when an agent inherits it?
A swimlane is a process map where each horizontal lane is one department and the arrows that cross lanes are handoffs. It is how nearly every enterprise workflow was drawn, and it is exactly the shape that fails an agent.
Here is the mechanic. Inside a lane, the steps are clear and the agent runs fine. At the arrow between lanes, a human used to do three unwritten things: confirm the work was actually done, restate it in terms the next department understands, and carry the context that lived in their head. Drop an agent into the lane and it inherits the visible steps. It does not inherit the person standing at the crossing. The context that traveled in someone's head now travels nowhere.
That is why a rollout looks perfect in the demo and falls apart in production. The demo runs one lane. Production has to cross four.
Will an orchestration platform fix the handoff, or do you redraw the process first?
Redraw the process first. An orchestration layer coordinates agents; it does not decide what the process should be, and buying one to sit on top of the old seam repeats the mistake that broke the workflow in the first place.
The tooling-first pitch is everywhere. Vendors like Moxo argue that shared workflow state across agents removes the queues where work waits for a human to reconcile it, and Dataiku describes a governed orchestration layer of agent hubs and shared memory to coordinate agents across departments. Shared state genuinely helps. It does not close the gap on its own. The gap shows up in the trust numbers: Harvard Business Review Analytic Services, in a survey of 603 leaders sponsored by Workato and AWS, found that 74% are pursuing enterprise orchestration, yet only 6% trust agents to autonomously run core end-to-end processes such as source-to-pay and hire-to-retire. Those are the cross-department processes. Companies are buying orchestration and still not trusting the result.
Even the platform built for this assumes you redraw first. Salesforce's own Agentforce Operations requires you to break a process into deterministic steps before an agent can run it, per Parulekar above. Salesforce reports the payoff at 50 to 70% faster cycle times and 80% less manual data entry in back-office processes like auditing and onboarding, though those figures cover back-office work in general and are not broken out for the handoff steps specifically. The tool earns the number after the redesign. The redesign comes first. Layering orchestration onto an unchanged process is the same move as bolting AI onto a workflow you never rebuilt, just one level up.
Redrawing a process that crosses three departments is a leadership call. Get the next piece in your inbox.
Who owns the process when one agent spans several departments?
One named human. The failure mode underneath most broken handoffs is that no single person owns the process across the old department lines, so when an agent drops context at the seam, there is nobody accountable for the whole loop.
MIT Technology Review found that most organizations are making this worse, not better. They are bolting agents onto the existing human operating model instead of redesigning reporting lines and process ownership. New roles are appearing to fill the gap: Agent Supervisor, Eval Owner, Exception Handler, Human-in-the-Loop Reviewer. The common thread is that every deployed agent needs one named human owner. When ownership stops at each department's border, the same way the old swimlane did, the agent inherits an orphaned handoff and so does the org chart.
The fix is one accountable loop. Give the whole cross-department process a single owner accountable for the outcome. That owner is who you go to when the handoff drops, and who has the authority to redraw the step that dropped it.
How do you redraw a workflow so one agent owns it end to end?
Start from the outcome the customer or the business actually wants, then collapse the lanes that only existed because a person sat at each border. The goal is one flow with the context traveling inside it, owned by one loop.
A practical sequence:
| Step | What you do | Why it matters |
|---|---|---|
| 1. Pick the process | Choose one that already crosses departments and stalls at the border: sales-to-cash, source-to-pay, hire-to-retire | These are where the 6% trust number lives, so the upside is largest |
| 2. Map the human handoffs | At each lane crossing, write down what the person actually does: confirm, translate, carry context | This is the unwritten work the agent is missing |
| 3. Collapse the lanes | Rebuild it as one flow with shared state, so context travels with the work instead of in someone's head | Removes the seam rather than automating around it |
| 4. Name one owner | Assign a single accountable owner for the whole loop rather than one owner per department | An orphaned handoff is an unowned handoff |
| 5. Keep humans on exceptions | Route people to the judgment calls and edge cases; leave the routine crossings to the agent | The person adds value on the exception once the reconciliation runs itself |
What that looks like in practice: TheNoah.ai describes a sales-to-cash process spanning CRM, finance, and customer success running as one coordinated agent flow, passing context and triggering actions across the old department lines without a manual handoff at each border. That is the shape you are after. One process, one owner, no seam.
A caution before you rebuild everything at once. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 on cost, weak business value, or thin risk controls. Redraw one high-value cross-department process, prove the loop holds, then move to the next. Pointing an agent at every handoff in the company at once is how the cancellations start.
Frequently asked questions
Is this an agent-quality problem or a workflow-design problem? Almost always the workflow. If an agent runs cleanly inside one function and fails when the work crosses into another, the model is fine and the process is underspecified. A human used to improvise across that boundary, and that step was never written down. As Salesforce's Sanjna Parulekar put it, the brokenness is usually in your requirements document.
What is a swimlane, and why does it break for an agent? A swimlane is a process map where each lane is a department and the arrows crossing lanes are handoffs. At each crossing, a person used to confirm the work, translate it into the next team's system, and carry the context in their head. An agent inherits the visible steps in the lane. It does not inherit the person standing at the crossing, so context drops exactly at the border.
Will an orchestration platform fix cross-department agent handoffs? Not on its own. Orchestration coordinates agents and shares state, which helps, but it does not decide what the process should be. HBR Analytic Services found 74% of companies pursuing orchestration and only 6% trusting agents to run core end-to-end processes. Redraw the process first, then use the platform to run the new one.
Who should own an agent that spans multiple departments? One named human, accountable for the whole loop rather than a single lane. MIT Technology Review reports new roles emerging for exactly this, including Agent Supervisor and Exception Handler, because every deployed agent needs one owner. When ownership stops at each department's border, the handoff between them belongs to nobody, which is where it drops.
What to decide before you buy anything else
Look at where your agent stalls. If it is the border between two departments, more orchestration software will not reach the problem, because the problem is a process drawn as swimlanes with a person filling the gap at every crossing. Pick one cross-department process, write down what the humans do at each handoff, collapse it into one owned loop, and put people back on the exceptions. Then buy the tool that runs the new process.
We write about making operations, sales, and marketing AI-native, with the structural decisions an operator actually has to make. Join the newsletter to get the next piece.
Sources
- Celonis, "The 2026 Process Optimization Report" (roughly 1,600 business and operations leaders). celonis.com
- VentureBeat, "Enterprise agentic AI requires a process layer most companies haven't built" (2026). venturebeat.com
- VentureBeat, "Salesforce launches Agentforce Operations to fix the workflows breaking enterprise AI" (April 2026). venturebeat.com
- Salesforce Newsroom, "Salesforce Launches Agentforce Operations to End Back-Office Bottlenecks" (April 29, 2026). salesforce.com
- Harvard Business Review Analytic Services, sponsored by Workato and AWS, "Enterprise Agentic AI: The Foundations of Trust" (603 leaders, fielded July 2025). workato.com
- MIT Technology Review, "Rethinking organizational design in the age of agentic AI" (May 26, 2026). technologyreview.com
- Moxo, "How agentic AI orchestration reduces handoffs, queues, and rework" (2026). moxo.com
- Dataiku, "Agent orchestration explained: How enterprises manage multi-agent AI workflows" (2026). dataiku.com
- TheNoah.ai, "How businesses can automate cross-department processes with AI agent flows" (2026). thenoah.ai
- Gartner, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 25, 2025). gartner.com
By Christopher Kliebenstein. We build and run AI-native workflows for operators who want results, not demos.