MIT's Project NANDA reviewed more than 300 enterprise AI initiatives and found a funnel that should be pinned to every data leader's wall: 60% of organizations evaluated enterprise-grade AI systems, 20% reached pilot, and 5% reached production. The attrition is not random. Projects die at predictable points, for predictable reasons — and almost all of them are visible in the data layer before a single agent is scoped.
The ten questions below are the ones we ask in the first week of a readiness assessment. They are not theoretical. Each one maps to a failure mode that has killed real projects.
Foundations: questions 1–4
1. For each entity this workflow touches, is there one source of truth? If "the customer record" lives in the CRM, the billing system, and a spreadsheet — with different values — the agent will act on whichever one it reads first. Conflicting records are the single most common blocker.
2. Do metric and field definitions agree across systems? "Open invoice," "active user," "qualified lead." If two teams define the term differently, the agent inherits the disagreement and executes it at scale.
3. How fresh is the data at read time? An agent acting on yesterday's pipeline run makes yesterday's decisions. Know the actual latency of every source the workflow depends on — not the documented latency.
4. Does every critical definition have a named owner? When two systems disagree, someone must have the authority to resolve it. If the answer is "we'd schedule a meeting," governance is a gap, not a function.
Operations: questions 5–7
5. Do your pipelines hold under load, or just on good days? Agents read continuously, not at reporting cadence. A pipeline that is "usually fine" produces an agent that is occasionally, confidently wrong.
6. Can the agent get access to what it needs — and nothing else? Scoped credentials, row-level permissions, and audit logging are not compliance garnish. They are what make an agent safe to point at real systems.
7. What happens to exceptions today? Every workflow has a long tail of weird cases. If humans currently resolve them through judgment and hallway knowledge, decide explicitly which exceptions the agent escalates — before it encounters the first one.
Control: questions 8–10
8. Can you reconstruct every action the agent took? Full audit trail: inputs read, reasoning steps, actions taken, approvals given. When finance or a regulator asks "why did it do that," a shrug is not an answer.
9. Are the agent's actions reversible? First deployments should act where mistakes can be undone — drafts, queues, recommendations — and graduate to irreversible actions only after accuracy is proven.
10. Is there a kill switch, and has anyone tested it? Monitoring, alerting, and the ability to halt the agent in minutes. Untested kill switches are decor.
Scoring yourself
Seven or more confident yeses: build. Four to six: a short, scoped foundation phase — weeks, not quarters — will pay for itself in the first deployment. Three or fewer: the agent project you are planning is actually a data project. That is normal. It is also fixable, and the fix compounds: in a documented engagement, one month of foundation work unlocked a production agent in two weeks — and every agent since has shipped faster because the substrate was already there.
Key points
- Readiness is measurable. Ten questions, answered honestly, predict production outcomes better than any vendor demo.
- Conflicting sources of truth and drifting definitions block more agents than model limitations ever will.
- Scope readiness work to one workflow; the foundation then serves every agent that follows.
- Audit trails, scoped access, and tested kill switches are launch requirements, not phase-two nice-to-haves.
The bottom line
Agents do not fail mysteriously. They fail at the exact points where the data underneath them was never trustworthy — and those points are knowable in advance. Ask the ten questions before you scope the build, and you join the 5% on purpose instead of the 95% by surprise.
Sources: MIT Project NANDA — The GenAI Divide: State of AI in Business 2025 (July 2025); The Register — coverage of the NANDA findings (August 18, 2025).