Why Most AI Contracts Fail Before They Begin
The problem is rarely the technology. It's what wasn't asked before signing.
A CFO at a distribution company in Mexico engaged an AI consultancy to automate financial reporting. The project ran eight months. At close, the team had a functional dashboard that no one knew how to operate, complete dependency on the vendor for any adjustment, and a final cost that tripled the original budget. The projected savings never materialized.
That wasn't bad luck. It was the absence of specific questions at the right moment.
This article compiles the questions an experienced CFO should ask before committing budget to any AI consultancy, regardless of project size.
Question 1: How Is Success Measured, and Over What Timeframe?
This is the most important question — and the one most often avoided.
A serious consultancy must be able to answer with concrete metrics before work begins: manual work hours eliminated, error reduction in specific processes, month-end close time, cost per report. Not generic projections of "improved efficiency."
If the answer includes phrases like "it depends on the context" without offering a clear measurement framework, that's a warning sign. Context always varies; the measurement method should not be an unknown.
What you should require: a success-metrics document agreed upon before the project starts, with current baseline values and targets for week 6, month 3, and month 6.
Question 2: Who Operates the System Once You Leave?
This question separates consultancies that build internal capability from those that build dependency.
The most common scenario in the market is this: the consultancy delivers a functional system, the internal team doesn't know how to modify it, and any change requires re-engaging the vendor. In practice, this converts a closed project into an open-ended maintenance contract with variable costs and no client control.
A delivery model oriented toward client autonomy means the internal team participates in the build from the start — not just in final validation. The difference is significant: a team that built alongside the consultancy can operate, adjust, and expand the system. A team that only received a handoff cannot.
A useful follow-up question: Can you show me a case where the client's team was operating autonomously in under 90 days?
Question 3: What Is the Real Time-to-Value?
Not the project timeline. The time to the first measurable result in production.
There is an important difference between a 12-month project that delivers value in month 11 and one that delivers the first results in week 6. For a mid-size company, that difference is budget committed without return for months, plus the opportunity cost of not having prioritized another initiative.
A reasonable benchmark for financial or operational process automation projects: the first agents in production should be active between week 4 and week 8. Not in an internal pilot. In real production, with real data.
If the proposal does not specify a production milestone within the first 6 to 8 weeks, it's worth asking why.
Question 4: What Happens If Results Don't Arrive?
This question makes many consultancies uncomfortable — and that is precisely why you need to ask it.
A vendor that trusts its own method has a prepared answer. It may be a scope-review mechanism, a partial results guarantee, or a defined escalation process. What should not happen is that the question generates ambiguity or a response that transfers all risk to the client.
In contractual terms, it's worth reviewing: are payment milestones tied to verifiable deliverables? Or is the payment structure time-and-materials with no outcome conditions?
Question 5: How Is the Cost of AI in Production Managed?
This point is frequently overlooked during the contracting phase and surfaces as a surprise in month 3.
AI agents in production generate variable costs: language model calls, storage, processing. Without an observability and cost-control system in place from the start, those costs can scale in unplanned ways.
A vendor with real production experience should be able to show how it monitors and controls the operational cost of the agents it delivers — not as an add-on service, but as part of the standard delivery model.
A Concrete Example: Month-End Close Automation
A professional services firm with 80 employees was spending between 40 and 60 hours per month consolidating data from multiple sources for the financial close. The process involved three people, multiple file versions, and an error rate that required additional manual reviews.
With a data consolidation and validation agent deployed in production by week 7, close time dropped to a range of 12 to 18 hours per month. The internal team was operating the system autonomously by month 3. The project ROI — accounting for the cost of hours freed and error reduction — landed in the 3x to 4x range over the first six months.
That outcome was possible because the right questions were asked before signing: agreed success metrics, a defined autonomy model, a production milestone at week 6.
Conclusion: The Contract Begins in the Conversation Before It
A CFO who arrives at the signature with these questions answered holds a fundamentally different position of control than one who relies on the vendor's commercial proposal.
This isn't about distrusting the consultancy. It's about aligning expectations, metrics, and responsibilities before committing budget. Consultancies that operate with rigor have no difficulty answering these questions in detail.
If you want to apply this filter to your situation and assess whether an automation project makes sense for your company right now, you can request a free diagnostic. No prior call required, no commitment.
→ Request free diagnostic
