Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 — killed by escalating costs, unclear business value, and inadequate risk controls. Read the fine print and the prognosis changes: the projects at risk are the ones that started with a demo. The ones that survive start with the data.
For leaders who know agents matter but don't know where to begin, that is the actual answer. The beginning is not a vendor evaluation or a model bake-off. It is a 90-day sequence that puts one real workflow into production — and builds the muscle to repeat it.
Why projects die between demo and production
The failure pattern is consistent enough to name. A team picks an impressive use case, wires a model to a few APIs, and produces a demo that looks alive. Then production asks harder questions: which system owns this field, what happens when the pipeline is late, who is accountable when the agent acts on a stale number.
MIT's Project NANDA put numbers to the pattern in its GenAI Divide report: 60% of organizations evaluated enterprise-grade AI systems, 20% reached pilot, and just 5% reached production. The stated causes were not model quality. They were brittle workflows, missing context, and misalignment with how work actually happens — data problems wearing an agent costume.
The 90-day plan exists to surface those problems in week two, when they are cheap, instead of month seven, when they are fatal.
Days 1–30: Audit what's underneath
Pick one candidate workflow — narrow, repetitive, measurable — and audit everything it touches. Not the whole data estate. One workflow's worth.
- Which systems of record hold the entities this workflow acts on?
- Do definitions agree across those systems, or does "active customer" mean three things?
- How fresh is the data at the moment the agent would read it?
- Who owns each definition when two teams disagree?
Most firms discover the same thing: the workflow they wanted to automate runs on data nobody fully trusts. That is not a reason to stop. It is the scope of the real project.
Days 31–60: Build the foundation for one workflow
Now the unfashionable work. Unify the records the workflow depends on into a source of truth. Shape the datasets the agent will actually read. Put definitions, ownership, and quality checks in writing so they hold under load, not just in a meeting.
This is the phase teams are tempted to skip, because it produces no demo. Skip it and the next phase produces nothing else. In a documented Finsight engagement, a Fortune 100 fintech spent seven months tuning an agent directly against a raw data estate with mediocre results; after one month of foundation work, the full workflow deployed in two weeks at near-perfect production accuracy. The sequence is the strategy. (Full account: In practice — invoice agent.)
Days 61–90: Ship one narrow agent
With a trusted substrate underneath, the agent build becomes the smallest part of the project. Orchestration, tool calls, retrieval, evaluation — all of it is easier when the data stops lying.
Ship narrow on purpose. One workflow, clear success metric, human review on the exceptions, and a kill switch that works. The goal of the first agent is not transformation. It is proof — to the board, to the team, and to the next use case — that the foundation carries weight.
What this plan is not
It is not a year-long data program before anyone touches a model. The audit and foundation phases are scoped to one workflow precisely so the first agent ships inside the quarter. It is also not a vendor-first path: platforms accelerate construction, but as the platform vendors themselves concede, the surrounding data work is the hard part, and no license completes it for you.
Key points
- Sequence beats selection: audit, foundation, one agent — in that order.
- Scope the data work to one workflow so the first agent ships in 90 days, not year two.
- Treat the readiness audit as a reusable asset, not a project tax.
- Judge the first agent by whether the second one ships faster.
The bottom line
"Where do we start?" has a boring answer, and boring is the point. Start with the data under one workflow, build the foundation that workflow needs, and ship one narrow agent on top of it. The firms on the right side of Gartner's 40% will not be the ones with the best demos. They will be the ones whose first agent made the second one easy.
Sources: Gartner — Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025); MIT Project NANDA — The GenAI Divide: State of AI in Business 2025 (July 2025).