Transform Internal Comms Chaos Into Clarity
Simplify internal comms with Haystack. Publish updates, maintain approval workflows, and track engagement—all from a single platform designed to reduce chaos and keep employees aligned.
This week, we’re handing the mic to our friend Ed Powers. He has great perspective on why AI is failing to deliver and how we’re applying it to the wrong parts of the system.
Keep reading for Ed’s thoughts on solving AI’s productivity problems…
Time is money in business. People are the biggest expense in our service and software-driven business-to-business economy, and given the fixed number of workdays in a year, saving time on execution directly saves money. To justify substantial AI investments, buyers must be able to point to significant time savings.
But as several recent studies have shown, productivity gains from AI haven’t yet materialized. A 2025 MIT Media Lab report said 95% of organizations deploying AI had seen no return on investment. And despite some large layoffs publicly attributed to AI, an October 2025 report from Yale’s Budget Lab showed little overall impact on job displacements. Still, AI vendors remain bullish. Dario Amodei, co-founder of Anthropic, said that AI models could soon surpass human intelligence, paving the way for machines to do even more work.
The Productivity Blocker
Perhaps one reason why AI hasn’t delivered its promised productivity gains is that it’s rarely aimed at eliminating bottlenecks.
What are bottlenecks? They are steps in the process that limit throughput because they take a substantial amount of time. In 1984, Eli Goldratt’s Theory of Constraints showed that the maximum speed of any production line is determined by its slowest step. He demonstrated that boosting speeds for anything but the bottleneck doesn’t help. Inventory just keeps piling up in some places while other work centers sit idle.
The same is true in non-manufacturing environments. Take airport security lines, for example. What’s always the bottleneck? The X-ray machine. Despite new tech that checks your ID and finds hidden objects on your person, lines remain long. Unless the carry-on X-ray goes any faster, people will still queue in large numbers.
Not surprisingly, it’s no different when applying AI in business processes. Some steps, such as summarizing text, have shown significant time savings. But unless AI eliminates bottlenecks, the overall process won’t go any faster.
Software development is an excellent example. The 2025 Annual Stack Overflow Developer Survey said 84% of coders were using AI. But much of the time saved by generating new code is offset by the time required to correct the slop the AI produces, such as glaring security vulnerabilities. Muhammad Awais writes, “As the cost of generating code approaches zero, the economic and cognitive bottleneck of software development has inverted. The primary challenge is no longer the speed at which we can write code, but the rate at which we can understand, review, and safely integrate the massive quantities of code being produced.” Simply replacing one bottleneck with another delivers no net productivity gain.
How to Fix It
So how can managers like you place their AI bets in the right places? By successively identifying process bottlenecks and using AI as just one tool for continuously improving the system. Here are the seven key steps:
Map the workflow. You can’t improve what you can’t see, and most knowledge work is invisible. Document the process so you know what you’re dealing with.
Measure the time it takes. You must gather the duration (end time minus start time) and task time (the time actively spent doing the task), both of which reveal hidden wait times.
Identify the longest step. That’s your bottleneck.
Apply lean practices:
Simplify– Reduce complexity in the offering. A better design can eliminate the bottleneck.
Triage– Sort inputs by type and reassign them to other work centers that can handle them faster. This can bypass the bottleneck.
Gate– Limit the input rate to match the maximum speed of the system, such as scheduling when the work begins. This reduces queues that form in front of the bottleneck.
Boost capacity– Duplicate the slowest step and divide the volume between them. Doubling down may appear to be more expensive, but removing the bottleneck often reduces total time and cost.
Make serial parallel– If possible, start work in the slowest area earlier and add its output later. Eliminating a dependency often eliminates the bottleneck.
Reduce rework– Poor quality in a previous step adds steps for inspection and correction. Improving quality can remove the bottleneck.
Lower the switching time– Too much multitasking can slow things down because more time is needed to catch up after focusing on something else. Limiting distractions or dedicating resources can reduce the bottleneck.
Remove non-value-added– Tasks like generating management reports that go unread don’t benefit anyone. Eliminating unnecessary work can remove the bottleneck.
Selectively automate– Having machines do certain high-volume, low-complexity tasks instead of relying on humans can increase completion rates and eliminate bottlenecks.
Develop, test, and refine. Try the fix and see what happens on a small scale, ensuring that you don’t solve one problem and create another.
Implement. Roll it out at scale.
Repeat, starting at step 3. Quality and productivity improvement never ends.
More Work to Do
Thus far, AI has proven beneficial for alleviating some bottlenecks, but there’s more work to be done.
Customer Support bots help triage incoming requests, answering “how do I do it?” and “where do I find it?”-type questions so humans don’t have to. Scheduling AI’s gate inputs by arranging service appointments, smoothing out demand, and reducing queues.
And, of course, agentic AI can selectively automate certain tedious steps, provided it doesn’t create rework. As described, this is often the case for AI writing software and other tasks.
AI has great potential to increase speed and productivity, but it must be applied in the right ways. Using AI thoughtfully to reduce bottlenecks, rather than to simply shift work from one place to another, is the best path forward.
— Ed Powers
Challapally, A., Pease, C., Raskar, R., and Chari, P. (2025) The gen AI divide: state of AI in business 2025. MIT NANDA.
Gimbel, M., Kinder, M., Kendall, J., and Lee, M. (2025) Evaluating the impact of AI on the labor market: Current state of affairs. The Budget Lab, Yale University.
Jin, B. and Stern, J. (2025). Anthropic CEO says AI could surpass human intelligence by 2027. Wall Street Journal online edition.
Cox, J., & Goldratt, E. M. (1986). The goal: A process of ongoing improvement. North River Press.
Sankhe, P., Patil, N., Ghorpade, M., Prasad, P., Linkesh, M. (2025). Empirical Analysis of AI-Assisted Code Generation Tools Impact on Code Quality, Security and Developer Productivity. International Journal For Multidisciplinary Research. DOI: 10.36948/ijfmr.2025.v07i06.61350
Awais, M.(2025). 5 Inconvenient Truths About Coding with AI Assistants, Medium.
We’re grateful you choose to read each week. When you’re ready for more, there are a couple ways we can help:
» Cover Your SaaS is a financial literacy course for go-to-market leaders. Grab your copy here.
» Promote your product and services to over 5,500+ senior SaaS Customer Success pros by sponsoring our weekly newsletter and podcast.
Was this email forwarded to you? Sign up at ChiefCustomerOfficer.io.
