Sit in on almost any AI budget discussion right now and you'll hear the same shortlist: a sales copilot, a content generator, a chatbot on the website. Nobody proposes the invoice matching queue. Nobody proposes the claims handling flow that runs eleven thousand times a month. Those never make the slide.
MIT's NANDA report, The GenAI Divide: State of AI in Business 2025, put numbers on this. More than half of generative AI budgets go to sales and marketing tools. Meanwhile the same report found the biggest ROI somewhere else entirely: back-office automation. Eliminating business process outsourcing, cutting external agency costs, streamlining operations. And the headline finding that made the rounds: roughly 95% of enterprise generative AI pilots deliver little to no measurable impact on the P&L. Only about 5% see rapid P&L impact.
So the money goes to one side of the business and the returns come from the other. That's not a technology problem. It's an allocation problem, and it has a structural cause worth understanding before you spend another euro.
You fund what leadership can see
Budget follows visibility. Sales and marketing produce demos that work in a boardroom. A generated campaign, a chatbot transcript, a pipeline dashboard: leadership can see these, so leadership funds them. Every one of those functions also has a leader in the budget room arguing for their slice.
Now think about the end-to-end flow behind a customer request. An order, a loan application, a claim. It crosses procurement, operations, finance, customer service, maybe compliance. Each of those functions has an owner. The flow between them has nobody. It doesn't present in the budget room because it isn't anyone's line item. It has no champion, so it gets no AI, no matter how much value is leaking out of it.
This is the same structural problem I wrote about in why your budget structure is your real operating model. You can declare any strategy you want; the spreadsheet that allocates money by function is what actually decides how the company runs. AI budgets are just the newest column in that spreadsheet, inheriting the same blind spot.
Where the money actually disappears
Let me make this concrete with a case from my own work.
A lending process crossed five teams. Each team had optimised its own part, and honestly, each part looked good. Response times inside every function were within target. If you asked any of the five team leads how their piece was performing, they could show you a green dashboard.
The customer still waited weeks.
The waiting time wasn't inside any function. It sat between them: in handoffs, in queues, in requests parked while another team got around to them. Nobody owned that time, so nobody measured it, so nobody's budget was accountable for it. Five locally excellent teams, one slow process, and the cost invisible because it lived in the gaps.
Now imagine that organisation gets an AI budget. Where does it go? Each of the five teams proposes something for their own part, because that's what they own and that's what they can see. The gaps, where the actual weeks are lost, get nothing. The money follows the org chart, and the org chart doesn't have a box for "between functions."
Why the back office is where AI compounds
Here is where it gets interesting, because the MIT finding isn't just an accounting quirk. There's a reason back-office automation is where the ROI shows up.
Look at what generative AI is actually good at today: work that repeats, follows rules, and runs at volume. That description doesn't fit a strategic sales conversation. It fits invoice processing, claims triage, document verification, order handling, compliance checks. Repetitive, high-volume, rule-based work with a defined process.
Three things happen when you automate that kind of work:
The savings compound. Shave two minutes off a task that runs ten thousand times a month and you've bought back over three hundred hours, every month, forever. Improve a sales pitch and you've improved one conversation.
A cost line turns into throughput. Back-office capacity is usually treated as pure cost: headcount, outsourcing contracts, agency fees. This is exactly what MIT found in the winning 5%: eliminating business process outsourcing and cutting external agency spend. When the same team can process double the volume, that cost line has quietly become a capacity line. That's a different kind of return than "the emails sound better now."
The results are measurable. Defined processes have counts, cycle times, error rates. You can see before and after. Much of the 95% that shows no P&L impact fails here, I suspect: not because the AI didn't do anything, but because "better content" and "more productive meetings" never had a baseline to move.
The better question to ask
The allocation we just looked at is what you get when the guiding question is "where could AI be impressive?" That question points at visible, customer-facing work. It produces pilots that demo well and move nothing on the P&L, which, going by MIT's numbers, is most of them.
There's a duller question that works better: where does the same task run thousands of times a month, with clear rules and a defined process?
Ask that and the answers look boring. Invoice matching. Address changes. Document intake. First-line claims assessment. Order exceptions. None of it will excite a board meeting. All of it has the three properties that make automation pay: volume, rules, and a measurable process.
There's one catch, and it's the one that sends most organisations back a step: you can't automate what you haven't defined. If the process exists as tribal knowledge spread across five teams, AI has nothing to attach to. Which brings the work back to operating model territory, where I'd argue resource allocation was always a design decision, not an accounting exercise.
A decision rule you can use
If you're deciding where next year's AI money goes, here's the rule I'd apply:
Fund where the work repeats, not where it shows.
And the sequence that makes it work:
- Pick one end-to-end flow, from customer request to fulfilled outcome. Not a function, a flow. The lending process, the claims process, the order-to-cash process.
- Map it across every team it touches, including the handoffs. In my lending case, this map alone was worth more than any tool, because it made the invisible waiting time visible for the first time.
- Find the constraint. Where does volume pile up? Where does the same manual step repeat endlessly? Where does work wait?
- Put AI there. At the constraint, inside the defined flow, with the cycle time and volume measured before you start.
Then compare that to the sales copilot proposal using the same test: how many times a month does this task repeat, and what number on the P&L moves if it gets faster? One of these investments will survive the question. It's rarely the visible one.
None of this means sales and marketing should get nothing. It means they shouldn't get the majority by default, just because they're the functions leadership can see. If more than half your AI budget sits there, as it does in most companies MIT looked at, the burden of proof should be on that allocation, not on the boring back-office case. This is also where AI stops being a tool purchase and becomes what it should have been from the start: a redesign of how the business runs.
The 95% of pilots with no P&L impact didn't fail on model quality. They failed on placement. The money went where AI would be seen, not where the work repeats.
So before you approve the next AI line item, ask the boring question: what runs ten thousand times a month in this company, and who owns the flow it lives in?
I'm Jenni. Most strategies stay vague, and most AI starts from the tech. I start from where the business actually makes money: mapping where it grows or leaks, designing how it should run, scoring what's worth building, and building the working tools that make it real. Founder of Digital Rebel, a Transformation Studio.