At ISE 2026, AI strategist Sol Rashidi warned that companies are launching AI pilots faster than they can prepare them for production. She called the trap “POC purgatory”: proof-of-concept projects that demonstrate potential but never become reliable, governed tools. Her message was not to stop experimenting, but to fix the data, security, infrastructure and workforce foundations that make useful deployment possible.
What Rashidi said at ISE 2026
Rashidi’s official Wednesday keynote was titled “The AI Reality Check: What It Takes to Scale and the Future of Leadership.” It took place on February 4, 2026, from 3:00 to 3:45 p.m. in room CC4.1. ISE’s “Push Beyond” event theme and professional AV and systems-integration audience might suggest a focus on equipment, but the keynote addressed enterprise AI, cybersecurity, leadership and operational automation. Those issues matter to integrators and technology leaders building smart spaces, connected facilities and supply-chain systems as well as to executives in other sectors. ISE’s keynote listing gives the session details; its speaker announcement describes the event theme and Rashidi’s Human Amplification Index.
“POC purgatory” was not the formal keynote title. It was the phrase highlighted in EE Times’ post-event report, which described Rashidi’s warning that many AI efforts stall because organizations lack dependable data, governance and operating foundations. ISE’s programme recap and speaker biography provide the event’s own account of the session and her background.
What “POC purgatory” means
A proof of concept (POC) is a limited experiment intended to show whether a proposed technology or use case is feasible. A successful demonstration is not the same as a production-ready system. A pilot may work with a curated dataset, a small group of users and close technical supervision, yet fail when it must handle real records, routine exceptions, security controls, support and integration with existing operations.
Recommended Free Tools
#1 Best Overall
POC purgatory is the pattern of repeatedly launching such experiments without a credible path to either production or a deliberate stop. A canceled pilot can still be a good outcome if it prevents an unsafe or uneconomic deployment; it becomes wasteful when no one owns the decision, the learning is not captured, and the same obstacles recur in the next trial.
EE Times reported that Rashidi put the share of AI initiatives paused, stopped or canceled at the proof-of-concept stage at 74%–88%. The report also said that, among more than 200 initiatives associated with her experience, about 63 reached production and 39 remained active. These are figures attributed to Rashidi’s keynote, not independently established industry-wide rates. The report does not define the full denominator or provide enough methodology to determine how “initiative,” “production” or “active” was counted, so the percentages should not be treated as a universal forecast. EE Times’ account of the keynote contains the reported figures.
Why AI pilots fail to scale
Weak data and system foundations
An AI model cannot reliably compensate for records that are incomplete, stale or defined differently across departments. EE Times reported that Rashidi linked scaling problems to weak master data management (MDM) and enterprise resource planning (ERP) foundations. A demonstration may rely on a clean extract assembled for the pilot; a production system may need to reconcile those records continuously with purchasing, inventory, customer or factory systems.
- Are the relevant records accurate, current and consistently defined?
- Who owns data quality and resolves conflicting records?
- Can the system access live information, or does the pilot depend on a one-off manual extract?
- Will the same integrations and permissions exist outside the test environment?
Governance and accountability left until later
A pilot often exposes governance questions that its sponsor has not answered: who approves the use, who owns the model and its outcome, what data it may use, which actions need approval, what must be logged, and who can pause the system. A company that postpones those decisions may discover that a technically successful pilot cannot pass review or be supported as an operational service.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallRank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
Security permissions change at deployment
A low-risk test may use synthetic or limited data. A production deployment might touch employee, customer, financial, factory or logistics records. Read-only access is not equivalent to permission to change a shipment, update an ERP record or issue a command to operational technology (OT). An AI tool that answers a question also presents a different risk from an agent that can execute a workflow.
No business owner or operating case
A technical sponsor can produce a persuasive demo without securing the budget and responsibility needed to run the system. Before a pilot begins, leaders need to identify its users, the workflow it changes, a baseline for comparison, the operational target, integration funding and the team responsible for support. Without those, “success” may mean only that the model performed well during a presentation.
Workforce consequences are not planned
Automation can remove tasks without automatically building the higher-level skills workers need to take on. Rashidi warned that eliminating junior analytical work could weaken the development of judgment and problem-solving experience. That is a workforce-design risk as well as a job-displacement concern: if less-experienced employees no longer learn by doing foundational work, organizations may later have fewer people able to challenge AI outputs or supervise complex decisions.
Rashidi’s “4 D’s” test for useful automation
EE Times reported that Rashidi urged companies to prioritize industrial tasks that are dull, dirty, dangerous or involve massive data processing. The first three are the familiar alliterative categories; the fourth is a description of data-intensive work, not another D-word. Reported examples included janitorial cleaning and sending robots into hazardous-material environments. The EE Times report describes the examples and the framework.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The practical test is whether automation can reduce repetitive or hazardous work, or process information at a scale people cannot reasonably handle. That is Rashidi’s proposed orientation, not a universal rule that every other use of AI is inappropriate. Automation can augment people and eliminate particular tasks at the same time. Nor does keeping a person nominally “in the loop” guarantee good oversight: a reviewer needs adequate time, expertise and authority to understand and challenge what the system proposes.
Why agentic AI raises the stakes
Generative AI often produces text, code or recommendations for a person to review and act on. An agentic system may retrieve information across tools, choose which tool to use, trigger a workflow, update records or route resources. That shift from suggesting to acting makes identity and permission design central to the safety case.
Rashidi warned, as reported by EE Times, that companies may give agents broad access to IT and operational-technology environments even while putting human employees through lengthy access reviews. The contrast is important: software does not become accountable simply because it is fast or efficient. Before an agent can act, its operator should be able to answer:
- What identity does the agent use, and are its permissions limited to the task?
- Can it make changes, or only read and recommend?
- Which high-impact actions require human approval?
- Are tool calls and material actions logged in a way that supports investigation?
- Could malicious instructions in the data it reads redirect it?
- What happens when it uses stale, incomplete or corrupted information?
- Who can intervene or shut it down, and how quickly?
A useful deployment boundary is to separate recommendation from execution until reliability is demonstrated. An assistant that proposes a revised shipment schedule is not equivalent to an agent with permission to change that schedule in an ERP or warehouse system. The controls should match the consequence of the action, not just the sophistication of the model.
Rank #4
Automated governance: useful forecast, not a substitute for accountability
Rashidi reportedly predicted that organizations will need automated systems to monitor and govern AI agents because people may not keep pace with the speed and volume of agentic activity. Such tools could support continuous monitoring, policy checks, faster anomaly detection and machine-readable audit trails. Her prediction is not proof that one AI system supervising another is an established best practice.
Automated governance can also inherit blind spots, obscure how a policy decision was made, block legitimate work through false positives or become a valuable target itself. If an organization uses automated monitoring, it still needs independent controls, a route for human review and a clear assignment of responsibility when the monitor misses a problem.
The AI “flywheel” and the burden of verification
EE Times described a Q&A discussion about AI helping to design the chips, software and systems that run AI. This creates a feedback loop: AI increasingly contributes to infrastructure on which AI itself depends. The concern is not evidence of an inevitable collapse; it is a growing verification and dependency challenge. AI-generated code or design choices can be difficult to audit, especially when speed-to-market pressure reduces review. Interconnected suppliers and platforms can also concentrate risk, while teams may lose visibility into the assumptions behind a system.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Energy and infrastructure are part of the business case
EE Times reported that Rashidi used striking comparisons: a prompt consuming as much energy as recycling 47 plastic bottles, and full AI adoption by every Fortune 1,000 company potentially requiring power comparable to the entire U.S. electrical grid. The keynote report does not provide enough methodological detail to independently assess those figures, so they should not be treated as measured constants or precise forecasts.
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
The underlying operational point remains relevant: AI workloads require electricity and data-center capacity, and industrial uses may also depend on networks, cooling, storage and redundancy. Energy use varies with model, hardware, workload, prompt length and utilization. A credible deployment case therefore needs to account for energy cost and availability alongside integration, latency, licensing and support costs, rather than assuming that software capability alone determines whether a system can scale.
Workforce capability and the risk of “intellectual atrophy”
Routine analysis may look like low-value work when measured one task at a time, but it can teach new employees how a business operates and how to spot exceptions. If AI performs all entry-level analysis, workers may advance without building the experience needed to assess unusual cases. That can leave inexperienced managers supervising systems they do not understand well enough to challenge.
Rashidi argued that AI lacks prudence and that people retain an advantage in reading context and unspoken nuance. Those are her judgments, not settled measurements of what every AI system can or cannot do. The practical question for employers is whether workers continue to learn the process beneath the automation, and whether they can identify errors instead of merely approving outputs.
What the Human Amplification Index is meant to measure
ISE’s speaker announcement says Rashidi developed the Human Amplification Index™ (HAI), a framework intended to assess whether AI strengthens an organization and its workforce. It is her proposed framework, not an established industry standard. Its value is in widening the ROI discussion beyond speed, volume or headcount reduction.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problems- Does the system improve decision quality, not merely the number of decisions made?
- Can employees explain, question and override its recommendations?
- Does it preserve or expand institutional knowledge?
- Does it reduce dangerous, exhausting or repetitive work?
- Can the organization recover when the model or its data fails?
- Are employees gaining capability, or only being asked to approve outputs?
These questions make “human amplification” practical: leaders can assess not only whether a tool completes work faster, but whether it leaves people and the organization more capable of handling the next difficult case.
A practical scale-or-stop checklist for AI pilots
The following steps translate the keynote’s themes into a decision process; they are an implementation guide, not a verbatim Rashidi framework.
Quick Recap
- Choose a consequential business problem. Define the operational issue and intended user before selecting a model.
- Name the production owner. Assign responsibility for the outcome and ongoing support before the pilot starts.
- Audit data and integration. Identify sources, quality owners, access paths and gaps between test data and live systems.
- Set an operational success threshold. Choose measurable targets that cover reliability and business value, not just demo performance.
- Fix a decision date. At a pre-agreed point, scale, redesign or stop; capture what the pilot established and what it did not.
- Keep recommendations separate from execution initially. Grant action rights only after reliability and safeguards have been tested for the consequence involved.
- Use least-privilege agent access. Limit identities and permissions to the specific workflow; do not treat broad access as a convenience.
- Log actions and establish escalation. Record tool use, identify who reviews incidents, and define a practical shutdown route.
- Test edge cases and hostile inputs. Check unusual conditions, misleading data and failure recovery, not only typical examples.
- Measure workforce and infrastructure effects. Track capability, safety, energy, latency, resilience and support costs alongside productivity.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
