
The Pilot That Never Becomes Production
Adoption rates in finance surveys are high and deployment rates are not. The difference is a set of decisions made at the start of a pilot, usually without anyone realising they were decisions.
- Survey work through 2026 has reported high experimentation with AI in finance alongside much lower rates of production deployment.
- The gap is not usually caused by the technology failing to work in the pilot.
- The common causes are scope chosen for demonstration rather than value, no production owner, and controls left until the end.
- A pilot designed to prove a capability rarely converts, while one designed to run a real process usually does.
- ChatFin works within existing ERP approval routes, which removes the control redesign that stalls many deployments.
Surveys of finance functions consistently report high levels of AI use, and consistently lower levels of anything running in production against real transactions.
The natural reading is that the technology disappoints under real conditions. That is occasionally true and is not the usual explanation.
More often the pilot worked, everyone was pleased, and it did not become production because of decisions taken before it started.
The pilot that cannot convert
A pilot designed to demonstrate that something is possible is a different artefact from a pilot designed to become production, and the difference is established at the scoping stage.
The demonstration pilot uses a clean data subset, runs alongside the existing process rather than as part of it, is owned by someone in a transformation or innovation role, and defers the control and security questions until after the capability is proven.
Each of those choices is reasonable for proving a capability. Together they guarantee that converting to production is a second project, with a different owner, a different scope and a fresh approval process, at a point when the initial enthusiasm has been spent.
The alternative is a narrower pilot that runs a real process end to end for real transactions, with the controls in place from the start. It is less impressive to watch and it converts.
This is a general observation from adoption patterns rather than a claim about any specific organisation.
The four causes
Across organisations the same reasons recur, and all four are decided early rather than discovered late.
None of them are technology problems:
Two pilot designs
The choice is made at scoping, usually without anyone framing it as a choice.
The second is smaller, less impressive and considerably more likely to be running in a year.
Prove the capability
Broad scope, curated data, parallel to the existing process, controls deferred. Produces a compelling demonstration and a conversion project that has to be sold separately.
Run one real process
Narrow scope, real data including the awkward cases, inside the existing control environment, owned by whoever will own it in production. Less impressive and it is already production if it works.

Scoping for conversion
Three questions asked before a pilot starts determine most of whether it converts.
They are not difficult questions and they are frequently not asked.
| Question | Why it matters | What a poor answer means |
|---|---|---|
| Who owns this in production | The production owner should choose the scope and be involved throughout | The pilot solves a problem its eventual owner did not prioritise |
| What controls apply, from day one | Approval routing, evidence and segregation configured in the pilot | Controls become a redesign after the capability is proven, which stalls conversion |
| Does the data include the difficult cases | Exception rates on curated data do not predict production | The business case is built on a rate that will not survive |
All three are answerable in a scoping conversation and all three are commonly skipped in favour of starting quickly.
The exception rate is the business case
Most finance AI business cases rest on the proportion of items that flow without human intervention, and that proportion is the number most distorted by a curated pilot.
A pilot on clean data can show a very high automatic rate. The same process on the full population, including the awkward suppliers, the unusual transactions and the entity with poor master data, produces a materially lower one.
That does not make the case invalid. It makes it different, and discovering the difference after committing to a business case is what turns a successful pilot into a difficult conversation.
Running the pilot on real data including the difficult cases produces a smaller headline number and a business case that holds.
The pilot worked. It did not become production because of decisions taken before it started.
ChatFinDesigning a pilot that converts
Four steps, all at the scoping stage rather than during execution.
Let the production owner choose the scope
Whoever will own it in production selects the process, based on where their effort actually goes rather than on what demonstrates well.
Configure the controls from day one
Approval routing, evidence and segregation as they will be in production. It is a configuration during the pilot and a redesign afterwards.
Use real data including the difficult cases
The awkward suppliers, the unusual entities, the poor master data. A lower headline rate on real data is worth more than a high one on curated data.
Define conversion criteria before starting
What result means this becomes production, agreed in advance, so the decision is not renegotiated from scratch afterwards.

Why the ChatFin model removes one obstacle
The controls obstacle is the one that most commonly stalls conversion, and it is largely structural rather than a matter of effort.
ChatFin operates inside the approval routes already configured in your ERP. Work is prepared as drafts that follow the existing approval path, using each platform's supported interfaces, with the evidence record produced as a consequence of the design.
That means the control question is answered by configuration rather than by redesign. A pilot running this way is already operating under production controls, so conversion is a scope decision rather than a controls project.
The same holds across NetSuite, SAP Business One, JD Edwards, Acumatica, Dynamics 365 and Sage Intacct, which matters where a pilot in one entity is intended to extend to others running different systems.
AI Cost Estimator for the CFO Office
Four inputs, a first year cost range, a timeline, and three delivery models compared. No login required.
Frequently Asked Questions
Why do finance AI pilots so often not convert?
Usually not because the technology failed. The common causes are that the production owner did not choose the scope, controls were deferred until after the capability was proven, the process was selected to demonstrate well, and the pilot ran on curated data that does not predict production exception rates.
What is the single most useful scoping question?
Who owns this in production. If the answer is someone who was not involved in choosing the pilot, conversion requires selling the result to a person who did not prioritise the problem.
Why does deferring controls stall conversion?
Because controls configured during a pilot are a setup task, and controls added afterwards are a redesign requiring approval from people who were not part of the pilot. The second is a project, and projects need sponsors and budgets.
Should a pilot use real data?
Yes, including the difficult cases. Exception rates on curated data are materially better than production and the business case usually rests on that rate. A lower number derived from real data is more useful than a higher one that will not hold.
General information only. It is not legal, accounting or tax advice, and rules change. Check your own position with your advisers before acting on it.
Design the pilot to be production if it works.
Conversion is determined at scoping: who owns it in production, whether controls are configured from day one, whether the data includes the difficult cases and what result means it converts.
ChatFin runs inside your existing ERP approval routes, so a pilot is already operating under production controls.
Book a DemoSee the platform
Run a pilot that is already production if it works.
Bring one process and the person who would own it. We will scope it against real data, inside your existing approval routes, with conversion criteria agreed upfront.