Sooner or later, the board asks the CIO two questions about AI: what are our agents actually doing, and is it worth what we're spending?
Few can answer either with confidence. But the work it takes to answer them builds something worth far more than an ROI number: the foundation for letting agents run whole functions.
The thesis: the data you collect to prove AI ROI today is the context data you will need to govern agents tomorrow. Measuring outcomes forces you to trace work end to end, and that same trace of human and agent work shows which work can safely run on its own.
Why AI ROI is so hard to prove
Adoption isn't the problem. McKinsey's 2025 survey found 88% of organizations use AI in at least one function, yet only 39% report any EBIT impact, and for most of those it's under 5%. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing cost, unclear value and weak risk controls.
The reason is structural. AI tools report on themselves: seats, prompts, suggestions accepted. That's activity, not outcomes. And agent spend, from licenses to compute, sits in its own budget line, disconnected from the work it paid for.
Real work doesn't live in one tool. A single feature moves from a customer request to a Jira ticket, a GitHub branch, a CI build, a change approval and a cost line in finance. An AI assistant sees one of those hops.
So enterprises building an ROI case end up asking three questions. What tooling do we have? How are we using it? And how does it connect to the other systems, including systems of record, that make up how we actually work?
That third question is where ROI lives. Answering it means seeing work end to end, across organizations, functions, processes and data silos.
From ROI to governance
Engineering got agents first, so that's where the ROI question showed up first. It won't stay there.
Agents now work wherever work happens: coding agents like Claude Code, OpenAI Codex and Cursor, assistants like Microsoft Copilot, Glean for company knowledge, and agents on data platforms like Snowflake. Most work across systems rather than inside one. Finance, HR, ITSM and revenue teams will ask engineering's questions: what did the agents do, did it help, and what did it cost?
Then come the harder questions: was the agent allowed to do that, and which work is ready to hand over to agents entirely?
The governance gap is wide today. A Cloud Security Alliance survey found 68% of organizations can't reliably tell an agent's activity apart from a person's. A second CSA survey found only 28% can trace an agent's actions back to a human sponsor across all their environments.
Each platform governs its own agents, inside its own walls. But one request can pass through a Copilot agent, a coding agent and a data agent on Snowflake before it's done. No single platform sees the whole path.
This is where the two problems meet. The end-to-end trace you build to prove ROI (who or what did each step, under whose authority, with what result) is exactly the trace you need to govern people and agents working side by side.
ROI and governance aren't two projects. They're two readings of one dataset.
A full-lifecycle ontology, not a point-in-time one
What ties it together is a full-lifecycle ontology: a supervised model of the objects, actors, handoffs and decisions that make up how work gets done, across the whole life of the work rather than a snapshot of it. It's built from two sources: event logs from the systems work already runs through, like Jira and GitHub, and agent traces, the step-by-step record of what every agent did.
Flora, our Full-Lifecycle Ontology Reasoning Agent, puts it to work: people ask Flora questions, and Flora briefs agents with the context they need.
Two things make the ontology different from a process map.
It's living. Every time an agent joins a process, a human step is automated, or an agent's access widens, the model changes. A static map goes stale in weeks.
It's supervised. Agents learn from the context they're given, including the record of other agents' work. Good patterns spread fast. So do bad ones. OWASP now lists poisoned agent memory among the top risks for agents.
Supervision means humans set policy, review changes and can trace any agent decision back to the context that shaped it. That's how you govern a hybrid workforce: not just who an agent is, but what it may do, where and for whom.
Agents learning from agents' work creates a recursive loop. Supervised, it's the path to autonomous maturity: agents act, the ontology records the results, humans correct the context, and agents learn from the better record.
Each cycle earns trust, moving a process from humans doing the work, to agents assisting, to agents acting with approval, to agents running it alone. A process moves up a stage only when the record shows it's ready: outcomes holding steady, little rework, and few agent actions that humans have to override. Autonomy isn't switched on. It's earned, one cycle at a time.
What to do now
Start with ROI, because that's where the budget and urgency already are. But build it so the same data serves governance next, and autonomy after that. Three moves get you there.
- Measure the workflow, not the tool. Define ROI on cycle time, rework, quality and cost per outcome across the full process, with agent traces joined to your systems of record.
- Give every agent an identity and a human sponsor from day one. Govern what it may do in each process, at each step.
- Treat the context you feed agents as an asset: versioned, reviewed and reversible.
At Bloomfilter, we start in the software development lifecycle, because that's where agents went to work first. We build the full-lifecycle ontology across the tools engineering runs on and show whether agents are actually driving business outcomes. Because it brings human and agent activity into one view, the same data that proves ROI also governs agents alongside people.
From there, once the record shows how work actually flows, every function can redesign it around humans and agents together: engineering today, then finance, HR, ITSM and revenue.
The enterprises that win won't just have the most agents. They'll have the clearest map of how their agents and people get work done, and the confidence to hand whole functions to agents.
If you're building the ROI case for AI in engineering, build the trace first. We'd like to compare notes.