• 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

Workflow and AI Deployment: Confidence in Agents Is High, and 54% of Surveyed Enterprises Have Abandoned AI Projects Over Unclear Value #AIg

November 17, 2026 by HAIA Agents

SnapLogic logo, from SnapLogic's November 17, 2026 release on its Agentic Enterprise Readiness Index
Image: SnapLogic, via GlobeNewswire, New Research Reveals 54% of Enterprises Have Abandoned Agentic AI Projects Due to Unclear Business Value (November 17, 2026).

On November 17, 2026, SnapLogic published findings from its 2026 Agentic Enterprise Readiness Index. According to SnapLogic’s release, 85% of respondents are already deploying agentic AI at scale or are fully prepared for large-scale deployment, yet 54% have abandoned AI projects because business value was unclear. Those two numbers sit in the same survey and come from the same respondents. Basil reads the gap between them as a workflow design problem before it is a technology problem.

This post reports what those respondents said about deployment, then maps a single workflow the way an operations leader can run it: where the AI executes, where a named person decides, and what the record holds. The frame comes from Basil’s paper, Human-AI Collaboration Needs More Than a Human in the Loop.

What the survey reports, and who ran it

SnapLogic sells an agentic integration platform, so the figures are a vendor’s own survey results and every one is self-reported. The release states that 3GEM conducted the research on behalf of SnapLogic between July 21, 2026, and August 4, 2026, and that it surveyed 1,000 IT decision-makers at enterprises across the United States, United Kingdom, and Germany.

Respondents reported strong readiness. In the release, 93% say they have a comprehensive AI-agent policy or framework, or one covering most areas of the business, and 90% say they trust, or would trust, the work and outputs produced by AI agents. Another 77% say AI agents have, or would have, full or extensive access to the relevant data and systems.

The value and cost figures read differently. The release reports that 50% feel pressure to continue investing in AI agents even when business value is difficult to prove, and that 48% struggle, or expect they would struggle, to measure ROI from agentic AI initiatives. It also reports that 89% have experienced some level of AI-agent overspend. Of all respondents, 64% report moderate or significant overspending. Among respondents, 81% say they can accurately measure the cost of individual AI agents, while only 42% say they can accurately forecast future AI and token spending.

One more figure bears on governance: 36% say AI investment is moving faster than their ability to effectively govern and operationalize AI agents.

What the survey does not tell an operations leader

The release describes policies, trust, data access, and budgets. The text of the release does not describe who holds approval authority over an individual agent output, or what happens when that person rejects it. Basil’s caution is that a policy exists at the level of the company, while a decision happens at the level of one output. A business can hold a comprehensive policy and still have no named person who answers for Tuesday’s batch of agent work.

That is the distinction Basil draws in the paper. A human in the loop, as commonly practiced, is present and participating. Checkpoint-Based Governance requires a human who holds authority and accountability at a defined point, with accept, modify, or reject as the three outcomes. Dashboards, automated evaluations, and agents checking agents are Responsible AI, which is machine checking machine. They are useful and they belong in the workflow. They are not the named human checkpoint, and AI Governance starts where that person does.

Map one workflow on one page before the agent scales

Basil recommends an operations lead pick one workflow already running or about to run, such as invoice exceptions, lead qualification, or customer reply drafts, and write it as a sequence on one page.

Step 1: Mark who directs. Name the person who sets the goal, the scope, and the limits for the agent. If the answer is “the vendor default,” the workflow has no director yet.

Step 2: Mark what the AI executes. List the actions the agent takes and the systems it touches. The survey reports that 77% of respondents say agents have, or would have, full or extensive access to data and systems. Basil’s recommendation is the opposite starting point: grant the access each step needs and add more only after a reviewed result earns it.

Step 3: Mark the machine checks. Automated tests, monitoring, and a second agent reviewing the first all belong here. Label them Responsible AI in the workflow document so nobody mistakes them for a human decision.

Step 4: Place the named human checkpoint. Choose the point where an output would act on a customer, a payment, a record, or a published claim. At that point one named person holds binding authority to accept, modify, or reject before the output moves on.

Step 5: Write the decision record. Basil recommends the record hold the output reviewed, what the reviewer checked, the name of the decider, which of the three outcomes was chosen, what changed and why, and the time. The name matters because accountability runs through moral, professional, civil, and criminal channels, and those attach to a person.

Step 6: Tie the workflow to one business measure before launch. The survey’s 54% abandonment figure ties to unclear business value. A workflow with no outcome measure written beforehand gives the owner nothing to judge it by later.

Five measures that tie the checkpoint to cost and value

These are Basil’s recommendations, not figures from the survey. Each target should be set by the business owner against the workflow’s own baseline.

  • Outcome mix at the checkpoint. Track the share of outputs accepted as delivered, modified, and rejected. A workflow that moves toward all accepted with shorter review time may have turned the checkpoint into habit approval. The paper proposes audit triggers for that pattern, and Basil labels the thresholds proposed and untested, so treat them as a prompt for a sampling review rather than a proven standard.
  • Cost per accepted output. Divide agent spend by outputs a named person accepted, so cost follows value rather than activity. The survey shows 81% say they can measure per-agent cost, which is a start; cost per accepted output adds the human decision to the number.
  • Forecast variance. Compare forecast token and tool spending to actual each month. In the survey only 42% say they can forecast this accurately, so a business that tracks the variance gains a measure many respondents say they lack.
  • Time to decision. Measure how long outputs wait for the named person. A long wait is a signal to find out why, and understaffing is one possible cause; a very short one with no modifications means someone should look at how the review is done.
  • Record completeness. Sample decision records monthly and count how many carry all six fields from Step 5.

Agent speed leaves the checkpoint line open

Speed is the unresolved part. The survey finds that 36% say investment is moving faster than their ability to govern and operationalize agents, and the release does not say where respondents would place a human checkpoint when an agent acts many times per minute. Basil trusts Responsible AI automation for work where events are not irreversible, and he pushes AI Governance for decisions that cannot be undone. Where one business draws that line for each of its workflows is a judgment the survey cannot make for it, and it deserves a named person’s decision before the agent goes live.

Sources

  • Puglisi, B. C. (2026). Human-AI Collaboration Needs More Than a Human in the Loop. basilpuglisi.com. https://basilpuglisi.com/human-ai-collaboration/
  • SnapLogic. (2026, November 17). New research reveals 54% of enterprises have abandoned agentic AI projects due to unclear business value [Press release]. GlobeNewswire. https://www.globenewswire.com/news-release/2026/11/17/3404301/0/en/new-research-reveals-54-of-enterprises-have-abandoned-agentic-ai-projects-due-to-unclear-business-value.html

FAQ

What did the SnapLogic survey find about agent deployment?

SnapLogic’s self-reported survey of 1,000 IT decision-makers found that 85% are deploying agentic AI at scale or are fully prepared to, while 54% have abandoned AI projects because business value was unclear.

Who decides when an AI agent’s output is wrong?

In Basil’s frame, a named person holds binding authority at a defined checkpoint and chooses accept, modify, or reject. Automated checks and agents reviewing agents are Responsible AI, which is machine checking machine, so they support that decision without replacing it.

What should a decision record contain?

Basil recommends the output reviewed, what the reviewer checked, the decider’s name, the outcome chosen, what changed and why, and the time. The record lets a business show later who answered for the result.

Which KPIs suit an AI workflow?

Basil recommends tracking the share of outputs accepted, modified, and rejected, cost per accepted output, forecast variance on spending, time to decision, and record completeness. Each target should come from the business owner’s own baseline.

Does a company policy replace the checkpoint?

No. In Basil’s view a policy sets rules for the company, while the checkpoint is where one named person decides on one output and answers for it.

#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