In my years of building and deploying enterprise AI systems, I have observed a recurring pattern - leadership teams evaluating new AI initiatives often fall into what I call the soft ROI trap.
And that’s what I dive deep into this edition of the newsletter - how unit economics and workflow integration go hand in hand. And why it’s important for every business leader to understand this, if you are going to make AI work at your org.
- Rahul Thota, CEO, Akaike (about me)
Learn About My AI Workshop for Business Leaders
I work with Pharma, CPG and other industry leaders to identify which stalled pilots are technically feasible, commercially useful and worth shipping - without disrupting existing workflows.
A Tale of Two Pilots
One of our clients is a major global consumer durables and home appliances brand. It operates thousands of franchised retail showrooms and experience centers, worldwide.
They are a well-funded enterprise with a dedicated innovation budget and a leadership team that consciously wanted to experiment with cutting-edge vision AI to explore what was possible. They wanted to test whether in-store CCTV camera feeds could digitize manual compliance checks - verifying display bays, brand signage, promotional setups, and product placement across 4,000 outlets, while also tracking customer foot traffic and floor movement across display zones.
From an engineering standpoint, the pilot was a success.
We built models that successfully processed the feeds, generated territory-level compliance reports, and achieved roughly 90% accuracy in counting unique walk-ins while filtering out repeat visitors.
The experiment delivered on its technical goals. However, when leadership looked at transitioning from an exploratory pilot to an enterprise-wide rollout, it failed on both critical pillars:
Broken Unit Economics: Continuous 24/7 video streaming and GPU inference across thousands of camera feeds is one of the most compute-heavy workloads you can run. Infrastructure costs scaled linearly with every store and camera feed added.
High Workflow Friction (Inspection, Not Acceleration): Knowing that an experience center’s promotional display was misaligned or that a customer lingered near a smart refrigerator for three minutes before walking out was merely passive observation. It didn’t give store reps a direct, real-time action to close a high-ticket sale.
The vision AI pilot served its purpose as an exploratory test, but scaling it permanently did not make economic or operational sense.
During our ongoing work with their operations team, a different opportunity emerged.
The client’s 7,000 retail outlets received approximately 200,000 inbound phone calls every week. These calls were handled by local store staff or routed through a central customer support desk. Store reps picked up some calls, missed others, and rarely had the time to manually sift through call logs to track down serious buyers.
We designed an event-driven voice analytics pipeline using NLP (natural language processing) to transcribe and analyze these call recordings across multiple regional languages.
This second pipeline succeeded in scaling across the business because it nailed both pillars:
Feasible Compute Economics: Unlike continuous 24/7 video streams, voice processing runs only when a call actually takes place. Compute costs remained low, predictable, and directly tied to activity.
Zero Workflow Friction (Direct Revenue Link): Instead of auditing store displays, the system scanned call conversations to isolate specific intent signals - like a customer asking for local store pricing, stock availability, or delivery timelines. Out of 200,000 weekly calls, the model isolated the top 2,000 high-intent leads and pushed them directly into the CRM call-back queues store reps were already using every day.
By focusing follow-ups on those 2,000 high-intent leads without forcing reps to learn an alien dashboard, the client achieved an immediate 2% to 3% uplift in retail sales - translating to roughly 2,000 additional high-value sales closed.
90% of AI Pilots Fail to Scale either due to Bad Unit Economics or Piled-On Workflows or Both
This comparison reflects a broader trend across enterprise deployments:
Piled-On Tools vs. Core Workflows: Research from MIT shows that most enterprise AI budgets go toward front-office teams like sales and marketing. But these pilots often report the lowest ROI because the tools get piled on as extra daily work - like asking reps to log into a new inspection dashboard, rather than taking a tedious task off their plate.
Specialized Tools Beat DIY Platforms: The same MIT study found that focused, domain-specific tools built for one targeted job succeed in production at roughly twice the rate of massive, generic “do-it-all” internal builds.
The 10% Scaling Trap: McKinsey’s enterprise AI benchmarking reveals that while nearly two-thirds of organizations experiment with AI pilots, fewer than 10% ever scale a project to the point of showing real profit or cost savings. Most remain “science experiments” that look impressive in board meeting demos but break when exposed to real-world compute costs or operational friction.
When evaluating AI initiatives, it is easy to get drawn to technically complex media types like multi-modal vision AI . But long-term operational scaling depends entirely on affordable compute efficiency and frictionless workflow integration.
Three Questions to Ask Before Scaling an AI Initiative
To avoid falling into The Soft ROI Trap, leadership teams should stress-test every proposed AI project against three operational criteria:
Is the compute model continuous or event-driven?
Model your inference costs at scale before committing to a permanent rollout. Continuous processing (like 24/7 video feeds or constant scraping) causes GPU costs to scale rapidly. Event-driven inference (processing data only when a call, document, or transaction occurs) keeps operational costs aligned with actual business value.
Is this tool performing inspection or acceleration?
Tools that merely audit compliance, monitor adherence, or generate passive visibility dashboards are difficult to justify on a balance sheet long-term. Prioritize pipelines that directly accelerate a core operational process - like isolating high-intent sales leads, cutting document drafting time, or reducing inventory stock-outs.
Does the output fit into existing daily interfaces?
If an AI system requires frontline employees to open a separate dashboard or adopt an alien workflow, usage will drop off as soon as executive attention moves elsewhere. The output should flow directly into the tools, CRMs, or messaging queues your team already relies on.
- Rahul Thota, CEO, Akaike (about me)
Learn About My AI Workshop for Business Leaders
I work with Pharma, CPG and other industry leaders to identify which stalled pilots are technically feasible, commercially useful and worth shipping - without disrupting existing workflows.