How to Evaluate an AI Investment Before You Approve the Budget
Somebody on your leadership team brought up AI again this month. Maybe it was your CEO asking what the plan is. Maybe it was a board member mentioning that a competitor already has something live. Either way, you’re now expected to produce an answer, and a budget line, before the year closes. The problem isn’t that AI is uninteresting. It’s that you’re being asked to make a capital decision on a technology you can’t fully evaluate, on a timeline set by anxiety instead of analysis. That’s not a strategy. That’s a reaction, and reactions rarely survive contact with an actual invoice.
An AI investment case is the internal argument for spending money on an AI tool, pilot, or platform. It states the specific business problem the investment is meant to solve, what it actually costs to run once licensing, integration, and staff time are counted, and how you’ll know within a defined timeframe whether it worked. Without that case, an AI purchase is a subscription with better marketing behind it.
Why “Doing Something With AI” Isn’t a Strategy
Urgency is not the same as clarity, and most AI budget requests we see are built on the first without the second. Somebody read an article, sat through a vendor demo, or watched a competitor’s press release, and the instinct became: we need to be doing this too. That instinct isn’t wrong to have. It’s wrong to act on without a business problem attached to it first.
The results back this up. According to MIT NANDA’s 2025 State of AI in Business report, 95 percent of generative AI pilots at companies fail to deliver a measurable return. That’s not a story about the technology being immature. Researchers found the failures were concentrated in projects with no clear owner, no defined workflow to change, and no way to measure whether anything actually improved. Those are decision-process failures, not technology failures. A better model, purchased under the same conditions, produces the same result.
If your AI conversation started with “what should we buy” instead of “what problem costs us the most right now,” you’re already building the case backward.
Most AI Pilots Fail Because Nobody Defined What Winning Looks Like
The gap isn’t adoption. It’s proof. According to McKinsey’s 2025 State of AI survey, 78 percent of organizations now use AI in at least one business function, yet more than 80 percent report no tangible impact on enterprise-level profit from that use. Only 17 percent say AI contributes 5 percent or more to their bottom line. Adoption has outrun measurement almost everywhere, and mid-market companies feel that gap harder than enterprises do, because they don’t have a data science team to quietly absorb a failed pilot.
The fix isn’t a bigger pilot or a longer trial period. It’s defining, before you spend a dollar, what specific metric changes and by how much. Fewer support tickets escalated. Faster quote turnaround. Lower error rate on a specific manual process. If nobody can name the metric before the pilot starts, nobody will be able to point to it afterward either, and the project will quietly become “the AI thing we tried.”
Ask These Questions Before You Approve an AI Budget Line
A short set of questions, asked honestly, will tell you more about an AI investment than any vendor pitch deck.
What specific task or decision does this change, and who currently owns that task? If the answer is vague, the business case is vague.
What does this cost once you count integration, data cleanup, and the staff time needed to actually use it, not just the licence fee on the invoice?
What happens to the work this replaces? Someone’s job changes. Naming that upfront prevents the quiet resistance that kills adoption six weeks in.
How will we know in 90 days whether this worked, and who is checking? A pilot with no review date is a subscription, not an experiment.
What’s the cost of being wrong? Some AI applications fail cheaply. Others touch customer data, financial reporting, or compliance obligations, and a wrong call there is expensive in ways that don’t show up on the first invoice.
Treat AI Vendors Like Vendors, Not Like Magic
The same governance discipline that applies to any other vendor relationship applies here, and we see mid-market companies suspend it the moment “AI” is in the product name. Contracts get signed faster. Data-handling terms get skimmed instead of read. Nobody asks who owns the output the tool generates, or what happens to your data if you cancel.
That’s backward. AI tools that touch customer information, financial data, or core operating workflows deserve more governance scrutiny than a standard SaaS purchase, not less, because the failure modes are less visible until they aren’t. A model that quietly degrades in accuracy doesn’t send an outage alert. Build AI vendor evaluation into the same discipline you’d apply to any strategic vendor relationship: defined success criteria, a named internal owner, and a review point that isn’t optional.
Approving an AI budget line isn’t the finish line. It’s the start of a governance relationship you need to actually manage.
The pressure to have an AI plan before year-end is real, but a rushed answer isn’t a plan, it’s a liability with a good demo. Before you approve the next AI budget request, make someone write down the specific problem it solves, the full cost of running it, and the metric that will tell you in 90 days whether it worked. If your team can’t produce that case internally, that’s a sign worth taking seriously, not a reason to skip the exercise.
FAQ
How do I know if an AI investment is actually worth it for my business?
Start with the problem, not the tool. If you can name the specific task, cost, or bottleneck it addresses, the metric that will move, and who owns tracking it, you have the makings of a real case. If the justification is competitive pressure alone, treat that as a reason to investigate, not a reason to buy.
What is an AI readiness assessment?
An AI readiness assessment is a structured review of whether your organization has the data quality, process clarity, and internal ownership needed to get value from an AI investment before you spend on one. It typically surfaces gaps, like undefined workflows or messy data, that would sink a pilot regardless of which tool you chose.
Why do so many AI pilots fail?
Most fail for the same reasons any technology initiative fails: no clear owner, no defined success metric, and no plan for how the work changes once the tool is in place. MIT NANDA's 2025 research found this pattern in the large majority of failed generative AI pilots, independent of which model or vendor was used.
Should I just wait until AI technology matures before investing?
Waiting isn't free. The capability gap between AI-fluent and AI-hesitant organizations tends to widen, not narrow, over time. The better move usually isn't waiting, it's narrowing your first investment to a problem small enough to measure honestly within 90 days.
Is it better to build an AI tool internally or buy one?
For most mid-market companies, buying and integrating an existing tool is more defensible than building, since it avoids carrying ongoing model maintenance and data infrastructure costs that rarely get budgeted upfront. The build-versus-buy decision should follow from the business case, not precede it.
When is the right time to engage Deliver Digital?
Ideally before selection begins. But we also help mid-project—when leaders realize what they bought isn’t what they needed. Either way, our goal is clarity, not complexity.
