Most “AI opportunity assessments” start with the technology stack and ask what AI could do with it. That’s backwards. The best AI use cases are in the operational bottlenecks you’ve been working around for years. Here’s how to find them.
The wrong place to look
Run a typical AI opportunity assessment and you’ll see this pattern:
- IT pulls the application inventory
- Someone maps the systems
- A workshop generates ideas about what AI could do in each system
- A long list emerges, scored on some matrix
- Nothing happens
The list looks comprehensive. It’s also useless. It generated ideas about AI capabilities, not about business bottlenecks. The result is a portfolio of plausible-sounding AI projects that won’t move the business. It’s why MIT’s NANDA initiative found that only about 5% of generative AI pilots achieve rapid revenue acceleration while the vast majority stall, delivering little to no measurable P&L impact. Their finding on the cause: it’s not the quality of the models. It’s the “learning gap” between a capability and a workflow that actually uses it.
The reason: AI doesn’t optimize systems; it relieves bottlenecks. That’s the Theory of Constraints in one line, the principle Eliyahu Goldratt laid out in his 1984 business novel The Goal: optimizing anything other than the constraint won’t improve throughput. And bottlenecks don’t show up in application inventories. They show up in the operational drag your team has been working around for years.
Where bottlenecks actually live
Three reliable places to find them.
1. Where work queues up
If there’s a queue, there’s a bottleneck. The queue tells you the constrained resource.
- Inbound RFQs queued waiting for the senior estimator
- Customer service tickets queued waiting for the right rep
- Invoices queued waiting for someone to reconcile
- Contracts queued waiting for legal review
Each queue is an AI candidate. The work in the queue can be drafted, triaged, classified, or routed by AI; the human becomes the reviewer instead of the originator.
2. Where senior people get pulled in repeatedly for the same kind of question
If your most senior people are answering the same question many times a week, that question is a bottleneck.
- The senior estimator who’s pulled in for any quote over $50K
- The COO who’s pulled in for any customer escalation
- The CFO who’s pulled in for any pricing question over a threshold
- The owner who’s pulled in for vendor disputes
The pattern they apply is mostly repeatable. AI can codify the pattern, surface the right context, and draft the answer; the senior person reviews instead of starting from scratch.
3. Where you’re losing visible deals or customers
Loss has a name. Where in your business are you losing, and why?
- Bids lost to faster competitors → quoting bottleneck
- Customers lost to slow response → service bottleneck
- Renewals lost to “we didn’t realize” → account health bottleneck
- Referrals lost to slow onboarding → intake bottleneck
The reason for the loss usually points at the bottleneck. AI can absorb the operational drag that’s causing the loss.
What’s not a bottleneck (and why it doesn’t matter)
Plenty of plausible AI use cases aren’t actually bottlenecks. The work isn’t queuing, senior people aren’t being pulled in, deals aren’t being lost. AI could do the work, but doing it doesn’t change the business.
Examples we see proposed and don’t pursue:
- “AI summaries of internal meetings.” Useful for individuals; doesn’t move a business KPI.
- “AI chatbot on the website.” Almost never the right starting point; rarely a bottleneck.
- “AI sentiment analysis of customer emails.” Cool dashboard; doesn’t change behavior.
- “AI knowledge base for the team.” Often good work; usually a second or third priority, not a starting point.
None of these are bad ideas. They’re just not bottlenecks. If you do them first, you’ve made a non-bottleneck more efficient while the actual bottleneck still controls the throughput. This is the trap the MIT NANDA work flagged directly: more than half of generative AI budgets go to sales and marketing tools, yet the biggest measurable ROI showed up in back-office and operational automation. Most companies are spending where the work is visible, not where the constraint is.
How to find the bottleneck: the 90-minute exercise
Take your leadership team off-site for 90 minutes. Run this.
Minutes 0–15: Where does work queue up? Each person names two or three places they see work queue up in their function. Capture without judging.
Minutes 15–30: Who gets pulled in repeatedly? Each senior leader names the recurring “I need 30 seconds” pattern they get pulled into. Same rule: capture.
Minutes 30–45: Where are we visibly losing? Each person names a specific recent loss (deal, customer, talent) and the proximate reason. Capture.
Minutes 45–75: Cluster the answers. Three lists are now on the wall. Look for overlaps. The clusters where multiple people pointed at the same workflow are the candidates.
Minutes 75–90: Score the top three candidates. For each, score on:
- Volume — how often does this happen?
- Impact — what’s the dollar value of fixing it?
- AI fit — structured work, data exists, human-in-the-loop fits, failure is recoverable
- Speed-to-value — can we ship something in 90 days?
The top-scoring candidate is your first Value Sprint. The next two are quarters 2 and 3.
What this gives you
A short list (one to three items) with operational justification, not a long list with theoretical justification. The difference matters because:
- The short list can actually be funded and executed
- The operational justification means the team understands why
- The bottleneck framing means the AI work will move a business KPI, not just a productivity metric
This is the framework we run with every AI Office client in month one. It usually produces a 90-day priority the team agrees with, including the members who started the engagement skeptical.
When the bottleneck isn’t an AI problem
Sometimes the bottleneck search turns up something AI can’t fix. That’s a useful finding.
If your top bottleneck is “we don’t have enough people,” AI can’t hire. The fix is hiring, with maybe AI as a follow-on to help the new people scale faster.
If your top bottleneck is “our senior estimator’s judgment is brilliant but slow,” AI can probably codify the judgment to make her faster, but the underlying scarcity is real and you need a succession plan, not just AI.
If your top bottleneck is “we don’t know what’s broken because we can’t see across our systems,” that’s a data and visibility problem. AI can help, but the first work might be data integration, a data-readiness Sprint, before any AI build.
The framework surfaces these honestly. That’s part of why it works.
Walk through it on your business
We run this exercise on every AI Office discovery call, usually a compressed version in 20 minutes. If you want to walk through it on your own bottlenecks, that’s the conversation to have.