TL;DR
AI projects fail in finance and operations because companies skip foundational steps. They deploy intelligent automation on broken processes, incomplete data, and disconnected systems. The fix is a strict sequence: simplify, integrate, automate deterministically, then add AI. Skip any step, and your project stalls.
Why Does the AI Failure Rate Exist?
McKinsey, Gartner, BCG, and a half-dozen other research firms have published variations of the same finding: Most enterprise AI initiatives fail to deliver expected value. The numbers vary slightly by methodology, but the pattern holds.
The instinct is to blame the technology. The AI was not smart enough. The model hallucinated. The vendor overpromised.
That is almost never the real cause.
At TFR Solutions, we have seen this pattern across 40+ ERP implementations in fashion, retail, distribution, and manufacturing. The root cause is operational, not technical. Companies try to fly before they can walk. They layer probabilistic AI on top of processes that do not work, data that does not exist, and systems that do not talk to each other.
Most failures blamed on AI are actually failures at steps one, two, and three: simplifying, integrating, and automating the predictable stuff first.
What Is the Order of Operations for AI in Finance?
We use a methodology called Walk Before Fly. The sequence is non-negotiable:
Crawl: Ground the truth. Document what actually happens, not what the SOP says happens. Establish baselines.
Walk: Sort the work. Every workflow goes through an Assess Gate and lands in one of five buckets: Keep As-Is, Simplify, Integrate, Automate (deterministic), or AI Candidate (probabilistic).
Run: Build the foundations. Simplify first. Integrate second. Automate the predictable stuff third. Each step has to work before you move on.
Fly: Add AI where it actually makes sense. Scoped roles with human owners and human-in-the-loop checkpoints. Never a single do-everything bot.
This is not a philosophical framework. It is the practical sequence that separates the few that succeed from the many that fail.
What Does Skipping Steps Actually Look Like?
Here is a real pattern we see in mid-market finance teams:
A company decides to implement an AI agent for accounts payable. The pitch sounds great: automatic invoice matching, exception handling, payment scheduling. The CFO signs off.
Six months later, the project is dead. What happened?
Nobody simplified first. The AP process had 47 exception types, most of which existed because of upstream problems in procurement and receiving. The AI was trying to handle exceptions that should not have existed.
Nobody integrated first. Invoice data lived in email attachments, a legacy portal, and two different ERP modules. The AI spent 80% of its cycles on data extraction instead of actual intelligence.
Nobody automated the predictable stuff first. Straightforward three-way matches were still being done manually. The AI was handling simple cases and complex cases with the same probabilistic logic, when the simple cases needed deterministic rules.
The AI was not the problem. The AI was asked to solve problems that simpler interventions should have handled.
How Do You Know If a Problem Is Actually an AI Problem?
This is the Assess Gate question. Not every problem needs AI. Most do not.
At TFR Solutions, we sort every workflow into one of five categories:
Keep As-Is: The process works. The cost of change exceeds the benefit. Leave it alone.
Simplify: The process is overcomplicated. Remove steps, eliminate handoffs, standardize exceptions. No technology required.
Integrate: The process breaks because systems do not talk. Connect them. This is a Celigo or MindCloud problem, not an AI problem.
Automate (deterministic): The process is predictable and rule-based. Use workflows, scripts, or RPA. If-then logic. No judgment required.
AI Candidate (probabilistic): The process requires pattern recognition, natural language understanding, or decisions under uncertainty. This is where AI actually adds value.
Most finance workflows land in categories two, three, and four. The AI candidates are smaller in number and narrower in scope than vendors want you to believe.
Is Your NetSuite Holding You Back?
Most mid-market companies are only using 40% of what NetSuite can do. Let's find the other 60%.
Book a Free Discovery CallWhy Do Finance Teams Struggle More Than Other Departments?
Finance and operations teams have a specific set of challenges that make AI implementation harder:
Data quality is worse than they think. Chart of accounts has been patched for 15 years. Vendor master has duplicates. Item records have inconsistent units of measure. AI cannot make good decisions on bad data.
Processes are more manual than documented. The close checklist says 5 days. The actual close involves 23 spreadsheets, 6 email threads, and 2 people who know where the bodies are buried. AI cannot automate what nobody has mapped.
Exceptions are the rule. Fashion and retail especially. Every customer has special terms. Every season has different SKU structures. Every vendor has different EDI capabilities. AI trained on averages fails on exceptions.
Stakes are higher. A wrong recommendation in marketing is embarrassing. A wrong number in revenue recognition is a material misstatement. Finance teams are appropriately cautious, which means failed pilots create lasting skepticism.
These challenges are solvable. But they require foundational work before AI enters the picture.
What Should a CFO or COO Do Before Approving an AI Project?
Five questions to ask before any AI investment:
1. What is the baseline? If you cannot measure the current state, you cannot measure improvement. No baseline, no project.
2. What have we simplified? If the process has not been reviewed for unnecessary complexity in the last 18 months, simplify first.
3. What is not integrated? If the AI will spend cycles extracting, matching, or reconciling data that should flow automatically, integrate first.
4. What is already automatable? If rule-based automation could handle 70% of the volume, automate that first. Let AI handle the 30% that actually requires judgment.
5. Who owns this agent? Every AI agent needs a human owner who understands its scope, monitors its outputs, and handles its failures. No owner, no deployment.
The AI Action Plan we offer at TFR Solutions covers these questions in the first week, sorting every workflow through the Assess Gate before any recommendations are made.
How Long Does the Foundation Work Take?
This is the question CFOs actually want answered.
For a mid-market company in the $10M to $100M range, with a modern ERP like NetSuite or Odoo already in place:
Simplify phase: 2 to 6 weeks per major process area. Map reality, identify waste, eliminate unnecessary steps.
Integrate phase: 4 to 12 weeks depending on the number of systems. This is where integration work pays off.
Automate phase: 4 to 8 weeks for deterministic automation. Workflows, scripts, alerts.
AI phase: 6 to 16 weeks for properly scoped AI agents with human-in-the-loop checkpoints.
Total timeline for a single major workflow: 4 to 10 months done properly.
That sounds long compared to the vendor pitch of "deploy AI in 30 days." But the 30-day deployment is how your project stalls. The 4 to 10 month timeline is how you build something that actually works.
What Happens When You Get the Order of Operations Right?
The outcomes are measurable:
Reduced cycle time. Close processes that took 12 days take 7. AP processing that took 4 touches takes 1.5.
Lower error rates. Exception rates drop because you fixed the upstream causes, not just the downstream symptoms.
Sustainable adoption. Teams use the tools because the tools actually help. No shelfware.
Compound returns. Each foundation step makes the next step easier. Clean data makes integration easier. Good integrations make automation easier. Solid automation makes AI more accurate.
This is the compounding effect we build toward in our Finance Operations engagements. The goal is not a single AI win. The goal is a finance function that gets better every quarter.
