The large majority of mid-market AI initiatives that fail do not fail because finance, sales, or operational planning has the wrong model. They fail because finance, sales, and operations each have different incentives, and therefore different “truths,” and unlimited processing power and model accuracy will never reconcile those differences into a single, shippable reality.
The tool isn’t the problem
If you look at any mid-market company implementing a new AI copilot or agent, you’ll find the same sort of preconditions underneath: a CRM that defines “active customer” differently than the ERP, and a couple of spreadsheets in between that aren’t officially sanctioned by either. Add AI and you don’t get clarity; you get a system that confidently massages three different data realities into one.
That’s the vendor’s problem! You won’t hear about it in marketing materials, but most AI tools have no guardrails that recognize when data sources contradict each other, and quietly resolve the conflict in whichever direction their training or prompting happens to favor. The result is still a confident-seeming answer, but it may take months to realize that, at some point downstream, the “truth” the model is feeding came from a completely different universe of data.
That’s why the industry benchmarks for generative AI are so thoroughly dismal: according to the MIT Sloan Management Review and Boston Consulting Group’s “State of AI” study from 2021, seven out of ten organizations see little to no value from their AI investments, with an even larger portion citing difficulties stemming from “organizational and data” causes rather than technical ones. That’s why it’s such a common sight in mid-market operations for AI initiatives to fizzle out; the projects that did get off the ground never touched the handoffs that needed reconciling, and therefore had nothing to resolve. And because silos don’t just hurt data, but often ownership, there are few people at the table who can claim responsibility for a unified view of the customer, the order, or any other process.
RPA learned this lesson ten years ago when robotic process automation tools hit the same sort of wall. Straightforward bots were great at executing processes with happy path data, but the second a decision had to be made or the data in one system didn’t match the data in another, they had to turn to a human to stop the flow of information, escalate the resolution of the exception to a manager, and make a judgment call about what to do next. RPA tools simply weren’t built to make judgments; they left those to humans, and created a whole support ecosystem around keeping the robots fed with the correct data to keep them functioning.
Where AI tools excel is in taking over that last responsibility, and that’s where teams evaluating ai for midmarket companies have to start looking at their operations differently. The same sort of rethinking has to occur at the earliest stages of implementing any AI initiative, and it’s why purpose-built AI tools have a crucial advantage in being built with that reality in mind: they don’t assume the data is clean, unified, or properly governed, because in most cases it isn’t.
Start with the exception, not the tool
You’ll notice that in this framework, you’re not buying software or building an enterprise data governance program first; those are both expensive distractions that fail to deliver quickly enough to justify the investment. Instead, you’re identifying a specific cross-functional process to tackle, end-to-end, and finding out exactly how and where it fails at the moment.
Quote-to-cash is the canonical example of such a process, where salespeople are regularly incentivized to give discounts that finance can’t account for, and operations can’t fulfill without conflicting with the master data in the system of record. Each of these exceptions represents a judgment call made by an individual at some point downstream, with none of them wanting to take responsibility for the judgment itself.
That’s the target for your AI initiative to focus on; the place where you can actually demonstrate clear business benefits by taking over one or more of those judgment calls and replacing them with either a rule-based or statistical model. That’s where you build the capability to process exceptions faster or more accurately than any individual could, removing a significant chunk of tedious busywork from each person involved in resolving the conflict. That’s not the 80% of the quote-to-cash process that’s already automated without thinking or the portion that runs smoothly through a few integrated tools; it’s the 20% of processing that occurs at the intersection of dissimilar systems which have to be manually entered into downstream tools, or the human judgment that has to be applied to a data set that isn’t actually fit to be used yet, or the one-off decision to grant an exception that no one wants to be responsible for making.
A sequence that actually works
Pick one place to start, not an end state. Focus on unifying data at the point of a single handoff, using whatever middleware you have access to if you’re unable or unwilling to fully unify disparate systems at first. Build the actual AI layer on top of that to recognize, route, or resolve the specific exceptions that have to be addressed at that point.
Do it on a small enough scale that the entire project can be completed within a single quarter to demonstrate tangible results and move on to the next one just as quickly.
Measure outcomes, not adoption
Forget about headcount and utilization or any other metric that serves as a proxy for value realization; if you’re looking to optimize operations, your indicators of success should be measured in reduced staff time spent on a particular handoff, reduced time needed to route or resolve exceptions, or reduced error rates on a specific set of tasks coming off that assembly line.
Having those numbers tell a compelling story about the benefits of AI for operations can significantly raise your chances of securing additional budget and support for future experiments. Even if they don’t, you’ve invested no more than a few months in an educational pilot with minimal risk to your bottom line.
Mid-market companies need fewer tools to solve the problems caused by inconsistent data governance, and more examples of fully realized end-to-end process improvements before anyone considers scaling anything.
