How Do You Sequence AI Implementation in Finance Operations Without Wasting Budget?

TL;DR

AI implementation in finance operations follows a strict sequence: Crawl (ground truth and process documentation), Walk (simplify and integrate), Run (deterministic automation), then Fly (AI agents). Skip any step and you waste budget on AI that amplifies broken processes. Each phase has specific checkpoints. You do not advance until you pass them. This is not conservative thinking. It is the only approach with a track record of reaching the P&L.

Why Do Most AI Pilots in Finance Operations Fail to Deliver ROI?

The failure rate for AI initiatives in finance hovers around 70 to 80 percent depending on which analyst report you read. The common explanation blames technology limitations or change management. Both miss the point.

The root cause is operational, not technical. Companies attempt to deploy AI agents on top of processes they do not fully understand, using data they have not validated, in workflows they have never mapped. The AI works exactly as designed. It just amplifies the chaos underneath.

At TFR Solutions, we see this pattern repeatedly in mid-market fashion, retail, and distribution companies. A finance team buys an AI tool, points it at their ERP, and expects automation. Three months later, the tool sits unused because the underlying data was inconsistent, the process had undocumented exceptions, or the integration broke when someone changed a field in NetSuite.

The sequencing rule is simple: Simplify, then Integrate, then Automate, then AI Agent, then Orchestrate. Most failures blamed on AI are actually skipped steps one through three.

The Crawl Walk Run Fly framework exists to prevent this. It forces the right order of operations and builds checkpoints that stop you from advancing until the foundation can support the next layer.

What Does Each Phase of Crawl Walk Run Fly Actually Mean?

The framework has four phases. Each maps to a specific operational maturity level and a specific type of work.

Crawl: What Does Grounding the Truth Look Like?

Crawl is about establishing ground truth. You cannot improve what you do not understand, and you cannot automate what you have not documented.

In this phase, you map every finance workflow end to end. Not the way the SOP says it works. The way it actually works, including the workarounds, the manual steps, the tribal knowledge that exists only in one person's head.

You also validate your data. Are your GL codes consistent? Do your item records have complete attributes? Is your customer master clean enough to support segmentation? Most companies assume their ERP data is accurate. Most are wrong.

The AI Action Plan covers this in the first week, sorting every workflow through the Assess Gate. Each workflow lands in one of five categories: Keep As-Is, Simplify, Integrate, Automate (deterministic), or AI Candidate (probabilistic). Not everything needs AI. Some things need a better spreadsheet.

Crawl Checkpoint: You have documented process maps for all finance workflows. You have a data quality baseline with specific metrics. You have classified every workflow through the Assess Gate. You do not advance until all three are complete.

Walk: How Do You Simplify and Integrate Before Automating?

Walk focuses on two things: eliminating unnecessary complexity and connecting disconnected systems.

Simplification comes first. If a process has twelve steps and could have six, you remove the waste before doing anything else. If a report requires manual data pulls from three systems, you question whether that report structure makes sense at all.

One pattern we have seen across 40+ implementations is that companies try to automate their current process exactly as it exists. This is almost always a mistake. The current process evolved through constraints that may no longer apply. Old software limitations, departed employees, compliance rules that changed. Automating it preserves waste.

After simplification comes integration. If your ERP does not talk to your 3PL, your accounting lives in two systems. If your ecommerce platform does not sync orders correctly, your revenue recognition is manual. These integration gaps create the manual work that AI is supposed to eliminate. But AI cannot eliminate manual work that exists because two systems do not share data.

For companies running NetSuite, our integration practice handles connections to Shopify, Amazon, WMS platforms, and financial tools. For Odoo environments, the logic is the same: connect before you automate.

Walk Checkpoint: Unnecessary process steps have been eliminated with documented justification. Core system integrations are operational and tested. Data flows correctly between all connected systems. You do not advance until all three are complete.

Run: What Qualifies as Deterministic Automation?

Run is automation, but not the AI kind. This is deterministic automation. Workflows with clear rules, predictable inputs, and known outputs. If X happens, do Y. No judgment required.

In finance operations, this includes scheduled reports, journal entry templates, approval routing based on thresholds, vendor payment batching, and reconciliation matching based on exact criteria. These do not require AI. They require workflows configured correctly in your ERP.

NetSuite consulting projects often spend significant time in this phase. Clients believe they need AI for invoice processing when what they actually need is properly configured approval workflows and saved searches. The ERP already has these capabilities. They just were never fully implemented.

The distinction matters because deterministic automation is reliable. It does the same thing every time. AI, by contrast, is probabilistic. It makes judgments. That is powerful for certain problems, but it introduces variability you need to manage. Do not use probabilistic tools for deterministic problems.

Run Checkpoint: All deterministic workflows are automated within the ERP or connected tools. Automation has been running for at least 30 days with measured error rates below defined thresholds. Manual intervention is documented and minimized. You do not advance until all three are complete.

Fly: When Is AI Actually the Right Solution?

Fly is where AI agents enter. These are workflows that require judgment, pattern recognition, natural language understanding, or decisions based on unstructured data.

In finance operations, genuine AI candidates include anomaly detection in transaction data, cash flow forecasting with multiple variables, vendor communication parsing, document extraction from inconsistent formats, and decision support for complex accruals.

Notice what is not on that list: data entry. Basic reconciliation. Approval routing. These are deterministic. They belong in Run.

At TFR Solutions, every AI agent we deploy has a human owner and human-in-the-loop checkpoints. AI augments your team. It does not replace them. The finance professional reviews the AI's output, validates the judgment, and maintains accountability. This is not a limitation of current AI. It is the correct design for financial controls.

Fly Checkpoint: AI is deployed only on workflows that passed through Crawl, Walk, and Run. Every agent has a documented human owner. Performance is measured against the baseline established in Crawl. You do not advance to orchestration until error rates and adoption metrics meet defined criteria.

How Long Does Each Phase Take for a Typical Finance Team?

Timelines vary based on company size, ERP maturity, and process complexity. For a mid-market company in fashion or distribution running NetSuite or Odoo, rough benchmarks look like this:

• Crawl: 2 to 4 weeks for initial assessment. The AI Action Plan delivers a baseline, classified backlog, and sequenced roadmap in two weeks.
• Walk: 4 to 12 weeks depending on integration complexity. Simple integrations take days. Complex multi-system connections take months.
• Run: 4 to 8 weeks for initial automation. Ongoing refinement continues.
• Fly: Varies widely. Simple AI use cases deploy in weeks. Complex orchestration takes quarters.

The critical point is that you cannot shortcut Crawl and Walk to get to Fly faster. Attempting this is how companies end up in implementation recovery situations.

What Are the Warning Signs That You Skipped a Step?

You know you skipped steps when:

• Your AI tool requires constant exception handling
• Automations break when someone changes a record in the ERP
• Finance staff do not trust the outputs and manually verify everything
• You cannot measure improvement because you never established a baseline
• The AI vendor blames your data quality

These are not AI problems. They are foundation problems. The fix is not a different AI tool. The fix is going back to the step you skipped.

How Do You Get Started With Proper Sequencing?

Start with an honest assessment of where you are. Most companies believe they are further along than they actually are. The ERP advisory process includes this reality check.

If you want a structured starting point, the AI Action Plan delivers exactly what Crawl requires: a baseline, a classified backlog sorted through the Assess Gate, and a sequenced roadmap. The engagement takes two weeks and starts at $5,000. You keep the deliverables regardless of what you decide to do next.

The sequence is the whole game. Get it right and AI becomes a multiplier. Skip steps and it becomes an expensive distraction. The choice is yours, but the physics do not change.

Frequently Asked Questions

Can you skip directly to AI if your processes are already documented?

Documentation is necessary but not sufficient. You still need validated data quality, tested integrations, and deterministic automation running reliably. Most companies with documented processes still have not completed Walk and Run. The checkpoints exist because assumptions about readiness are usually wrong.

How do you know if a workflow is an AI candidate or should be deterministic automation?

The Assess Gate sorts this. If the workflow has clear rules and predictable inputs, it is deterministic. If it requires judgment, pattern recognition, or handling unstructured data, it may be an AI candidate. The key word is may. Many workflows that seem to need AI actually need better process design or integration.

What happens if you deploy AI and then realize you skipped a step?

You pause the AI deployment and go back to the skipped step. This is frustrating but faster than continuing to patch a broken foundation. We see this regularly in recovery engagements. The right answer is always to fix the foundation, not to add more technology on top of it.

How do you measure success at each phase checkpoint?

Each phase has specific metrics. Crawl measures process coverage and data quality baselines. Walk measures integration uptime and error rates. Run measures automation reliability and manual intervention frequency. Fly measures AI output accuracy against the Crawl baseline. No success claim without a baseline and a measured result.

Does this framework work for companies not using NetSuite or Odoo?

The sequence is ERP-agnostic. The principles apply whether you run SAP, Microsoft Dynamics, or a collection of spreadsheets. The specific implementation details change, but Crawl Walk Run Fly logic does not. Process before technology. Humans before agents. Outcomes before outputs.

How does human-in-the-loop work practically for finance operations?

Every AI output has a designated human reviewer before it affects the ledger or triggers a payment. For high-volume, low-risk tasks, review can be sampling-based. For high-risk tasks like vendor payments or revenue recognition, review is 100 percent until error rates justify reducing oversight. The human owner is accountable for the AI's work, period.