A Fortune 100 fintech spent seven months building AI agents to read invoices, extract the fields, and post the correct accounting transaction. Results stayed mediocre — accuracy that never earned production trust. Then the sequence changed: one month rebuilding the data underneath, two weeks deploying the agent workflow on top. The rebuilt version outperformed the seven-month attempt immediately.
That inversion — a six-week data-first build beating a seven-month agent-first one — is the most instructive thing about the engagement. This is a deep dive into why the first attempt stalled, what the foundation month actually contained, and why the industry's agent failure statistics keep pointing at the same root cause. (The documented account lives at In practice — invoice agent.)
Why accounts payable looks easy and isn't
On paper, invoice processing is the ideal agent target. The task is well-bounded: a document arrives, fields get extracted, a transaction gets posted. The volume is high and the economics are visible — Ardent Partners benchmarks the average all-inclusive cost at $9.84 per invoice, with an 18.4% exception rate and only 35.4% of invoices flowing straight through without human touch.
But the extraction step — the part that looks like the whole problem — is the smallest part of the problem. Posting an invoice correctly means resolving it against vendor master records, purchase orders, GL account mappings, approval hierarchies, and payment terms. Every one of those is reference data living in some system, maintained by some team, drifting at some rate.
The models were capable in the fintech's first attempt. What they were reasoning over was not: data scattered across systems, inconsistent, with no source of truth for the records the workflow depended on. Accuracy plateaued at mediocre, and no amount of prompt engineering moved it — because the errors weren't reasoning errors. The agent was faithfully acting on contradictions.
The month that changed the outcome
The counterintuitive move was to stop working on the agents. The engagement spent one month on data engineering: organizing the data estate, building clean datasets, and establishing sources of truth for the records the invoice workflow depended on. No new models. No orchestration framework. Unglamorous pipeline and reconciliation work.
With that substrate in place, deploying the full agent workflow took two weeks — and it reached near-perfect production accuracy, outperforming the seven-month attempt from day one. Same task, comparable models. The variable that changed was whether the agent was operating on facts or fragments.
The industry data says this pattern is the norm, not the exception. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value. MIT Project NANDA's State of AI in Business 2025 found 95% of organizations getting zero measurable return from generative AI pilots, with only 5% of custom enterprise tools reaching production — failures it attributes to brittle workflows and missing context, not model quality. And in MuleSoft's 2025 Connectivity Benchmark, 95% of the 1,050 enterprise IT leaders surveyed named integration as a hurdle to implementing AI effectively — the average enterprise runs 897 applications with only 29% connected.
Three independent research bodies, one diagnosis: the agent layer keeps getting blamed for failures that originate in the data layer.
What data/AI leaders should take from this
Budget the foundation as part of the agent project
The instinct is to treat data cleanup as a separate, deferrable initiative and the agent as the deliverable. The fintech timeline argues the opposite: the foundation month was the agent project. Six weeks of sequenced work beat seven months of unsequenced work. Leaders scoping an AP agent — or any transaction-posting agent — should put the data-foundation phase in the same budget line and the same timeline, first.
Define sources of truth before defining prompts
Before any agent design, the questions that matter are ownership questions. Which system of record owns the vendor master? What does an approved invoice mean across the ERP and the AP tool? Who arbitrates when two systems disagree? These were hard before agents; agents make the cost of leaving them unanswered immediate, because the agent will act on whichever answer it happens to retrieve.
Measure readiness by the second agent
The strongest signal from the engagement isn't the first deployment — it's what followed. Since the invoice workflow shipped, a new agent has gone to production roughly every two weeks, each standing on the same clean foundation, each faster to build than the last. That compounding is the real return on the foundation month. A one-off agent on raw data is a demo with a maintenance bill; a foundation is a platform. That's the architecture argument behind how Finsight builds agentic platforms.
Looking ahead
Key points
- Treat plateaued agent accuracy as a data diagnosis before a model diagnosis — audit the contradictions the agent is inheriting.
- Sequence explicitly: unify and govern the records the workflow depends on, then build the agent. The inverse order is where the canceled 40% comes from.
- Fund the data foundation inside the agent project, not as a someday initiative beside it.
- Judge success by the marginal cost of the next agent, not the demo quality of the first.
The bottom line
The invoice agent was never the hard part. Seven months of tuning couldn't compensate for a missing source of truth; one month of building it made the agent almost easy. For leaders staring at a stalled agent pilot, the fastest path forward usually runs through the least fashionable work on the roadmap.
Sources: Gartner press release, June 25, 2025; MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025; Ardent Partners, State of ePayables 2025; Salesforce/MuleSoft 2025 Connectivity Benchmark Report. Claims about the anonymized Fortune 100 fintech engagement are limited to what is documented in our case studies — no invented metrics.