Why Agentic AI Adoption Is Underwhelming

by

From time to time, I’ve cited many success stories of Enterprise Agentic AI:

  1. Bajaj Finance: Agentic AI powers 10% of quarterly loan disbursements, worth Rs. 1,600 crore, by analyzing millions of calls and creating tailored offers that added Rs. 325 crore in volumes.
  2. DBS: Inhouse AI agent built on top of ChatGPT cuts down KYC and onboarding of private wealth management customers from days to hours.
  3. Sanofi: Inhouse AI agent built over Claude Code performs work previously handled by ServiceNow ITSM software.

Notwithstanding these sporadic case studies, Agentic AI has not exactly set the enterprise technology space on fire.

“Enterprise adoption of AI is very slow.” – Mohit Joshi, CEO, TechMahindra.
“AI adoption has not been smooth sailing in larger businesses. AI labs haven’t built a product that boosts productivity for most white-collar workers.” – Ramp AI Index via FORTUNE EYE ON AI newsletter.

Personally, I haven’t had any success with Agentic AI.

As many readers would know, GTM360 has been running a bimonthly newsletter for close to 10 years. It takes 40+ steps to produce each issue of the newsletter. The end-to-end “omnichannel” workflow crisscrosses desktop, blog, email service provider, and other environments.

I’ve been trying to automate the production of these newsletters for years. A few years ago, I attempted RPA (Robot Process Automation). I gave up when I found out that it cost at least $80,000. When Agentic AI burst on the scene around two years ago, I gave it a shot. But I still haven’t made any headway. In How AI Can Propel Personalization To A Higher Orbit, I predicted:

Agentic AI can go beyond personalizing the campaign and run the whole personalized marketing campaign on its own. That will be the topic of a follow-on post. Watch this space!

More than a year later, that follow-on post has not yet been written – because my super-optimistic prediction hasn’t aged well, and this newsletter continues to be produced manually as before.

One could say that I haven’t tried Agentic AI seriously enough. One wouldn’t be wrong.

But the point is I haven’t tried Generative AI / Large Language Model – the bedrock of Agentic AI – any more seriously, but I can still boast of several success stories of it.

Use Case Before ChatGPT With ChatGPT
Spreadsheet → actionable data analysis For more than a decade, I’d been searching for a software or website to which I could simply upload an Excel file and have it analyze the data and discover insights automatically. I tried Excel Pivot Tables, DataRobot and Gartner’s Smart Data Discovery products. They required too much manual data preparation work and still didn’t work as advertised. I uploaded the raw XLS file to ChatGPT and prompted it for actionable insights. I got them instantly.
Blog semantic search I had tried WordPress Search, Algolia, and Miso.ai to find a post on my blog based on its theme. None of them helped – they all worked on the basis of exact keywords that I couldn’t remember. My prompt “Can you find the post where I’ve written about whether to broach the topic of price early or late in the sales cycle?” yielded the correct blog post Should Sales Dodge The Price Question Or Answer It Upfront? in seconds.
Stock-price lookup for Excel portfolio tracker I had tried Excel web scraping on Google Finance / Yahoo Finance websites. The output was fragile and required repeated manual cleanup. ChatGPT gave me the step-by-step instructions to create the Excel-to-Google Finance web link, which worked perfectly.
Coding Given my dubious coding skills, I had struggled to make responsive design tables in my WordPress blog posts. My attempts to write the HTML code by hand and use table plugins both produced meh results. I make tables easily on Microsoft Word, upload the DOC file to ChatGPT, and ask for the HTML code for an equivalent table in WordPress. I copy whatever code ChatGPT gives me and blindly paste it in the “Code” view of WordPress editor. Always works!

(Click on the use case link to find out more details of each case study.)

On the contrary, I can’t claim even one success story with Agentic AI.

It’s not only me.

I’ve line of sight into 100+ prospects and customers of enterprise AI. All of them use genAI extensively (mostly Copilot or ChatGPT) . Around 20 companies kicked off Agentic AI initiatives, but the going is, how shall I put it, not smooth.

So far, I only had personal experience and anecdata about the lukewarm adoption of Enterprise Agentic AI.

Now, there’s research data.

According to a recent report by KPMG, many Agentic AI projects have been put on hold or aborted.

Big 4 consulting firms and pundits alike have advanced many reasons to explain the underwhelming adoption of Agentic AI ranging from hallucination, poor data quality, and security concerns through to non-deterministic nature of models, governance gaps, and questionable ROI.

I don’t disagree with any of those reasons but I think the RCA in the public discourse misses the elephant in the room: Critical GTM Disconnect. More specifically, misguided use cases.

Well-built conventional models are already cheap, transparent, fast and easy to explain. Agentic AI models bring inference cost, latency, hosting/vendor dependency, and limited interpretability. Even if the accuracy gain over a well-tuned statistical model is real, the operational cost may not be worth it. ~ Evident Insights.

Most Agentic AI use cases I’ve come across sound amateurish to me. By pitching customer service, fraud detection, loan underwriting, payments, etc., Agentic AI vendors are posturing that they invented automation.

This is obviously disingenuous. Through the decades, enterprises have used RPA and other automation technologies to automate many business processes including all of the above mentioned low hanging fruits of Agentic AI.

Against that backdrop, the real market for Agentic AI is replacement of existing automated workflows.

However, Agentic AI vendors are currently targeting automation of all workflows as though they’re manual.

As you can see, the real market is brownfield whereas the targeting is of a greenfield market. This is a critical GTM disconnect. I attribute the sluggish adoption of Agentic AI to it.

Let me take payments, the last of the low hanging fruits, to illustrate how this disconnect is playing out in the real world.

A Top 5 bank in UK offers TARGET-2 payments to its customers in UK and EU. These are high value RTGS payments (running into tens or even hundreds of billions of EUR each). For over 15 years, the bank has been using a high value fund transfer system that’s built with traditional automation technologies. On a typical day, the HVFT system processes 80,000 payments. Out of them, 76,000 go through the Straight Through Processing route with zero human touch, whereas 4,000 are swept to the repair queue to be handled manually. In the payments industry parlance, this payments shop has 95% STP Rate (see footnote 1).

By pitching payments as a use case, an Agentic AI vendor is targeting all 80,000 payments (greenfield market). Obviously, most banks would reject this GTM since their existing system already processes 76,000 of them autonomously. This explains the lacklustre adoption of Agentic AI.

If the vendor were to instead target just the 4,000 payments in the repair queue (brownfield market), the bank would respond to its Agentic AI solution more favorably.

Notice that there’s a 20X difference between these two TAMs – 80,000 versus 4,000.

tl;dr: Payments is not a compelling GTM, only payments requiring manual repair is.

As for why Agentic AI vendors use a GTM with such a severe disconnect, I can only speculate.

One possibility is that they’re full of people below a certain age. Like Millennials / Yoomers and Gen Zers / Zoomers tend to believe that all great things in the world were invented by their generation, workers in Agentic AI companies might genuinely think that they invented automation. While they’re obviously wrong, truth doesn’t matter. Until it does. (We saw the same behavior in fintech’s posturing of disrupting traditional banks.)

The other possibility is that they acknowledge the vast gap between the actual TAM and the TAM projected by them but deliberately target the 20X market (of 100) instead of the 1X market (of 5) in order to drive nosebleed valuations for their startups. When they do that, 95 out of 100 use cases will obviously not deliver value, thus vindicating the findings of the aforementioned KPMG study.

But that’s okay. It’s still very early days of the Technology Adoption Life Cycle of Agentic AI. At this point, markets care only about adoption, which is high because the vendor is targeting the projected TAM of 100. By the time markets start caring for value, it’s quite possible that the fast-evolving technology will find many new tenable use cases, thereby vigintupling (see footnote 2) the actual market from today’s 5 to 100 by then.

As they say, Edison invented electricity in the 1870s but it took another 70 years for the full range of electrical appliances to appear in American homes in the 1940s.

Called “Fake It Till You Make It“, this is a charter principle of Silicon Valley. The valley has used it for almost all transformative technologies of the past. I won’t be surprised if it’s using it for Agentic AI now.

Agentic AI is moving very rapidly. Anything can happen anytime. Don’t blame me if the technology achieves some new breakthrough next month and I write a totally contradictory blog post next quarter!

FOOTNOTES:

  1. Numbers changed to protect customer confidentiality but ratios are correct.
  2. Vigintupling is the term for multiplying something by 20 (20x). While “20x” or “twentyfolding” are more commonly used in everyday speech, “vigintuple” is the precise mathematical term, derived from the Latin word viginti, meaning twenty.