• Skip to primary navigation
  • Skip to main content

Puglisi Consulting for Artificial Intelligence

Independent AI Reviews: Strategy, Governance, and ROI, Before and After You Deploy

  • Home
  • Advisory Reviews
  • The Advisor
  • Marketing Across Verticals
  • Track Record
  • AI & Digital Marketing Blog
    • AIO Prompts
  • Basil Puglisi’s Blog

AI Studies and ROI: ISG Reports Visible Operating Gains While Most Financial Outcomes Fall Short #AIg

November 18, 2026 by HAIA Agents

Cover of ISG's State of Enterprise AI report, The AI Value Gap
Image: ISG, State of Enterprise AI: The AI Value Gap (2026).

A reported AI return is only as good as the decision trail behind it. A study can tell a finance or operations leader that something is happening; it cannot tell that leader who agreed to the baseline, who owns the number, or when anyone will look at it again. I use my paper on enterprise AI ROI as the frame for reading new studies, and the latest one worth reading closely comes from the advisory firm ISG.

What ISG reports, and who is saying it

ISG’s 2026 State of Enterprise AI report says AI is creating visible operational gains for many enterprises, but “it’s not yet consistently producing business and financial value.” That is the headline. The detail underneath is more useful to a business owner than the headline.

ISG says employee productivity gains came in slightly below what was expected and most business and financial outcomes fell far short. Those are ISG’s findings, published by ISG.

Two figures stand out for anyone planning a budget. ISG reports that 69% will keep increasing AI budgets if they continue gaining value at the current rate. ISG’s average increase in employee decision quality is 4.3%, which it calls AI’s strongest workforce impact, level with the creation of new AI talent roles.

What the page does not show

The public page I could read does not give a sample size or fieldwork dates. That is not a criticism of ISG; a landing page is a landing page. It is a reason a business owner should not carry a percentage from it into a board deck as if it were a measured benchmark. ISG is an advisory firm with an AI Advisory practice, so the report is also a calling card.

I also can’t tell from the page how the “value” in “business and financial value” was defined, who at each enterprise signed off on it, or whether the same people who ran the programs scored them. Those are the questions I would ask of any vendor-published or advisor-published result, and they are the ones the full report may or may not answer.

A second primary source tells you what kind of evidence this is

Stanford’s 2026 AI Index gives a useful contrast. Stanford’s 2026 AI Index reports that “Productivity gains from AI are largest in structured, measurable work where outputs are easy to monitor.” It cites studies reporting gains of 14% to 15% in customer support, 26% in software development, and 50% in marketing output.

Its corporate section is a different kind of evidence. The AI Index says its survey-based corporate data, drawn on McKinsey’s annual State of AI surveys, is self-reported and should be viewed as directional rather than comprehensive. In that survey data, respondents more often tied the highest cost savings to software engineering and manufacturing functions (56%) and revenue gains to marketing and sales (67%). The AI Index adds that the number of respondents who believed AI had improved various organizational measures was often similar to the number who did not believe it had any effect. In plain terms, the self-reported returns look a lot like opinion until someone ties them to a ledger.

Read together, the two sources point the same direction: measured gains show up where work is structured and outputs are easy to monitor, and executive-reported gains are harder to pin down. That is my reading of the pair, not a finding either source states.

Where the human sits in ISG’s numbers

ISG puts the human-led share of AI work at 55% today and expects it to shrink below 40% by the end of 2027. It expects 13% of AI work to be autonomous by the end of 2027, against less than 7% today. ISG says nearly a quarter of AI work is reviewed by humans today and almost 14% involves humans only for exception handling.

Here the vocabulary matters. A dashboard, an automated score, or one model checking another model is Responsible AI: machine checking machine. It can be fast and useful. It does not put a named person on the hook. AI Governance is the named human checkpoint, where one person looks at the evidence and accepts, modifies, or rejects what the system proposes, and that person’s name goes on the record. I describe that structure on my Checkpoint-Based Governance page.

ISG warns that requiring people to validate these functions can create bottlenecks when AI produces results faster than employees can review them. I take that as an honest description of the problem. If AI outruns the reviewer, the reviewer’s sign-off stops meaning what it says. Whether a checkpoint is real depends on whether the person has the time and the evidence to refuse.

Decisions a named person makes before claiming a return

Here is how I would turn the study into action, with every item below being my recommendation and not a sourced finding.

1. Take the baseline before launch. Measure the current cost, hours, error rate, and turnaround for the workflow for at least a few weeks. If the before-launch measure doesn’t exist, the return can only be estimated, not measured. 2. Name one owner for the number. One person, with a title, decides what counts as a saving. Hours saved are not a return until a budget line, a headcount plan, or a revenue figure changes. ISG’s own list of what enterprises need includes “clarify the economics of AI,” and I’d start there. 3. Set a review date and write down the stop rule first. My suggestion is 90 days, with the continue, modify, or stop thresholds written before the pilot starts so nobody adjusts them afterward. 4. Separate machine checks from human checks. Let dashboards and automated scoring do the first pass. Then have a named person sample the outputs and record accept, modify, or reject, with a reason. 5. Record the decision, not just the metric. A one-line note with the date, the owner, the evidence, and the call is enough to rebuild the reasoning a year later.

KPIs a business owner can track

These are suggestions drawn from my own practice. The targets are mine, and you should replace them with numbers your baseline supports.

  • Cost or time per unit of work, compared with the pre-launch baseline.
  • Rework and exception rate, because a faster process that produces more fixes has not saved anything.
  • Reviewer load: items reviewed per person per day, and the share of reviewed items that were modified or rejected. A reject rate that sits at zero for months deserves a second look; it may mean the checkpoint has become a rubber stamp.
  • Money moved: the dollar change on a budget line, attributed to one owner, not hours estimated.
  • Decisions logged on time: the share of scheduled continue, modify, or stop reviews that actually happened.

The part nobody has settled

ISG expects the human-led share of AI work to fall while the autonomous share grows. At some point on that path, a named person can no longer look at the work before it counts. Where that line sits, and what an owner signs for when review shifts to exceptions only, is a boundary I don’t think any study has settled, and a business owner will have to draw it for each workflow.

Sources

  • – ISG. (2026). State of Enterprise AI: The AI Value Gap [Report landing page]. https://isg-one.com/advisory/artificial-intelligence-advisory/state-of-enterprise-ai-report-2026
  • – Stanford Institute for Human-Centered Artificial Intelligence. (2026). AI Index Report 2026, Chapter 4: Economy [PDF]. https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf
  • – Puglisi, B. C. (n.d.). Enterprise AI ROI: Seven landmark reports, five decisions. basilpuglisi.com. https://basilpuglisi.com/enterprise-ai-roi-seven-landmark-reports-five-decisions/
  • – Puglisi, B. C. (n.d.). Checkpoint-Based Governance. basilpuglisi.com. https://basilpuglisi.com/cbg/

FAQ

What did the ISG report find?

ISG’s public page describes visible operational gains at many enterprises but says most business and financial outcomes have fallen far short of expectations. Treat it as an advisory firm’s published summary, not as your own benchmark.

Can I use these percentages as a benchmark for my business?

Use them as context, not as a target. The page gives no sample size or fieldwork dates, and your own baseline is the only comparison that matches your costs, customers, and workflows.

What is the difference between Responsible AI and AI Governance?

Responsible AI is machine checking machine, such as dashboards, automated scoring, or one model reviewing another. AI Governance adds a named human checkpoint who accepts, modifies, or rejects the output and answers for the decision.

What should I record before I claim an AI return?

Record the baseline, the one owner of the number, the review date, and the continue, modify, or stop thresholds. Add a short note of each human decision with its date and reason.

How long should a pilot run before I decide?

That depends on the workflow, and no source here sets a number. My suggestion is a 90-day review with the thresholds fixed in advance; adjust it to how quickly your work produces enough volume to measure.

#AIg

Share this:

  • Share on Facebook (Opens in new window) Facebook
  • Share on X (Opens in new window) X

Like this:

Like Loading…

Filed Under: AI Consulting, Digital & Internet Marketing, Mobile & Technology

Reader Interactions

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

© 2005 to 2026 Puglisi Consulting | basil@puglisiconsulting.com | basilpuglisi.com

%d