Human middleware is the biggest unpriced risk in agentic enterprise workflows

Human middleware is the biggest unpriced risk in agentic enterprise workflows

The people quietly patching the gaps between systems that don't talk to each other never appear on a budget line, until the day they are gone.

Every weekday morning in one regional office of a Fortune 500 manufacturer, someone opens a queue of new customer orders and keys each one by hand into the system that runs fulfillment. The two systems were never connected, and nobody can say exactly when the workaround stopped being temporary. The people doing this work have done it long enough to catch errors the software would miss, which is the problem: when they are out, the queue backs up, and when they leave, the process leaves with them, because it lives nowhere except in their heads. No project ever budgeted for this work. It is simply headcount, filed under titles that describe almost none of it, and it is the layer every plan for agentic workflow execution quietly runs through.

There is a name for this layer, even if no finance ledger carries it: human middleware. It is the group of people whose real function is moving information between systems, teams, or vendors that were never integrated to do it themselves. Every large enterprise has some. Few have priced what it costs, or what happens the day it fails.

The people who hold the systems together

Human middleware rarely announces itself as a job title. More often it is a logistics coordinator re-entering the same freight booking by hand into several carrier portals, because the systems share no common record and none of them flags a booking that silently fails. Or it is an approvals process that technically lives in a workflow tool but really lives in a handful of shared inboxes, watched all day by someone whose actual job is making sure nothing important sits unread. Sometimes it is a load-bearing spreadsheet an entire process depends on, maintained by the one person who knows its quirks. Most often, it is a name: the one person in a department everyone quietly understands to be the real interface between two systems that were supposed to talk to each other and do not.

A formal operational assessment inside one large enterprise recently put a number on the pattern for the first time: a register of dozens of manual workarounds, many run daily, each substituting a person for an integration that was never built. The assessment flagged the pattern itself, not any single workaround, as the largest unpriced operational risk on the books. That framing is rare: audits routinely price technology, contract, and reputational risk. Almost none price a process that only functions because one employee shows up.

Why the cost never survives a budget cycle

The reason human middleware persists through every cost review is structural, not careless. It never appears as a project, so there is nothing to cancel. It is absorbed into headcount, spread across a dozen ordinary-looking job descriptions. Cutting it looks like cutting a person, not a process, so nobody proposes it. And the real costs, key-person dependency, slower cycle times, error rates, live in operations, not finance, where nobody tallies them against a single line.

Ask a CIO or a chief supply chain officer to name the risk and they usually can, quickly. Ask what it costs and the answer gets vague. That gap between recognition and measurement is why the problem survives budget cycle after budget cycle: everyone senior enough to fix it already has a hundred priced risks competing for the same funding, and this one has never been priced.

Why agentic workflow execution stalls for the same reason

The pattern shows up again in why enterprise AI stalls. Eighty-eight percent of AI pilots never reach production, according to IDC research. Ask researchers why, and the answer is rarely model capability: the blockers are data quality, integration complexity, change management, and unclear ownership, according to the same body of research, the same conditions that produce human middleware in the first place. A workforce re-keying data by hand and a stalled pilot share the same root cause: nobody funded the handoffs.

Fixing agentic workflow execution starts with governance and integration, not agents

There is an honest complication most vendors skip. Pointing an AI agent directly at a human-middleware process, the coordinator's daily re-entry, the inbox of stuck approvals, before fixing the handoff does not remove the risk. It automates it. An agent that inherits an undocumented workaround runs it at machine speed, without the judgment calls the human was quietly making: the same errors, faster, harder to trace back to a cause.

Real AI workflow automation for the enterprise has to start earlier than the agent: map the handoff, decide who owns the exception cases, fix the connection, then let agents take the repetitive load off a named owner who still signs off on judgment calls. That sequence is the heart of the named-expert AI delivery model. A different delivery structure treats that sequence as the first deliverable, not an afterthought bolted onto an agent rollout. That is what an AI-native transformation engine is built to do.

Future Works runs that sequence as an operating partner, in 12-week cycles: named experts, not a rotating bench, orchestrate a fleet of AI agents, with a third of the fee outcome-staked against a value target the client's own finance team verifies. That combination, AI agents plus named experts, is built to replace the rotating bench, not the judgment calls. It is a slower first move than pointing an agent straight at the mess, and that is deliberate: some workarounds survive the fix precisely because the judgment call needs to stay human, integration or not.

The person re-keying orders every morning is the evidence: somewhere beneath that routine sits an integration nobody funded and a risk nobody priced, and no enthusiasm for agents changes the arithmetic until the handoff itself gets fixed. That is the sequencing point research on agentic operations keeps confirming: governance and integration first, agent fleets second, or the workaround simply moves faster.

If an enterprise still runs on people quietly connecting systems that were never built to connect, Future Works treats that gap as the first deliverable. Start with the AI operating system.

Matt Leta, Founder and CEO, Future Works.