The short answer: The obvious scam is real but boring: Gartner reckons only about 130 of the thousands of firms selling agentic AI are the real thing, and calls the rest agent washing. The scam I care about is the agency that delivers exactly what it promised, where the business still does not improve because nobody checked whether the thing being automated was the actual constraint. Automation multiplies whatever process you point it at, so point it at one that leaks and you get a bigger leak with better reporting.
Most SEO and digital marketing agencies were a scam.
Most AI automation agencies are the same scam in a new hoodie.
That’s an opinion, not a finding. And it’s a heavy word, so I’ll be specific about what I mean by it, because I don’t mean the obvious thing.
What is agent washing? The boring version of the scam
There is a straightforward version of the scam and it’s the boring one. In June 2025 Gartner estimated that of the thousands of companies selling agentic AI, only about 130 were real. They have a name for the rest, agent washing, which they define as rebranding existing products like assistants, RPA and chatbots without any substantial agentic capability underneath. That’s ordinary misrepresentation. You don’t need an essay to spot it.
The one I care about is the agency that actually delivers.
The chatbot works. The scraper pulls 500 leads a week. The n8n workflow fires at 9am every morning and it has fired every morning for four months. There’s a Loom walkthrough, a Notion handoff doc, an invoice that got paid. Nobody lied to you. Technically, it’s clean work.
It’s a scam because of what nobody asked before the build started.
Where I’m standing, so nobody has to guess
I build AI software for a living. Taxformify sorts client bank statements against CRA and IRS codes for tax firms, and it works well enough that people pay for it. I am not about to tell you AI doesn’t work.
AI tools are parts. Some are sophisticated, some are dead simple, and a good part in the right vehicle is worth every dollar you spend on it. What nobody has ever managed is fixing a car they hadn’t diagnosed by ordering more parts.
That’s the transaction I’m describing. Not the parts. The ordering.
Six holes, and you can only see one
Picture a ship taking on water. There are six holes in the hull.
Five are below the waterline. An offer nobody wants at that price. A sales process that loses people between the form fill and the phone call. No follow-up after day three. Pricing that folds the second someone pushes back. Churn nobody has ever sat down and measured.
One hole is above the waterline. Visible, embarrassing, and the one the owner points at when you ask what’s wrong.
“We don’t have enough leads.”
So the agency patches that one. Not out of malice. That’s the hole with a budget attached to it, and it’s the hole the client described in the discovery call.
Here’s the part that should bother you. The leaks below the waterline aren’t hypothetical, and they’re not rare.
In 2014, InsideSales published an audit rather than a survey, run the year before. They attempted a web form submission at 14,061 companies; 9,538 of those had a form that worked and received it. Of those 9,538, four thousand four hundred and seventy-two never responded at all. Forty-seven percent. Among the ones that did pick up a phone, 36% made exactly one attempt and stopped. Median time to a first call was three hours and eight minutes. The average was sixty-one hours.
Three years earlier, Harvard Business Review published a similar audit of 2,241 US companies: 23% never responded to an online lead, and among those that did, the average response took 42 hours.
Both are more than a decade old and both trace back to the same commercial interest, which I’ll come back to. But sit with the shape of it. Roughly half the businesses with a working lead form on their site never answered the leads it produced. I can’t tell you what share of those would have named lead generation as their bottleneck if you’d asked them, because nobody has run that study. I can tell you the ones I’ve sat across from usually did.
Meanwhile the thing they aren’t measuring is worth more. Reichheld and Sasser’s 1990 paper in HBR, the original rather than the version everyone paraphrases, found that reducing customer defections by 5% generated 85% more profit in one bank’s branch system, 50% more in an insurance brokerage, and 30% more in an auto-service chain. That’s the hole nobody builds a dashboard for.
Why AI automation projects fail
More leads arrive into a process that loses half of them. More conversations hit an offer that wasn’t converting at that price to begin with. More customers enter a delivery system that was already losing them.
The owner burns cash faster, gets more stressed, and lands on the conclusion that AI doesn’t work for their business.
AI worked fine. It did precisely what it was paid to do.
There’s now data on this at the enterprise scale, which is the only scale anyone bothers to measure. Gartner surveyed 350 executives at billion-dollar companies already piloting or deploying AI agents, intelligent automation or other autonomous technologies, with fieldwork in late 2025 and publication in May 2026. Around 80% reported workforce reductions. The useful part is the cross-tab: the rate of reduction was nearly equal between the organizations reporting strong ROI and the ones reporting only modest gains or negative outcomes. Gartner’s own line was that workforce reductions “may create budget room, but they do not create return.”
Cutting cost through automation showed no relationship with the business getting better. The intervention landed. The outcome didn’t follow. That’s self-reported ROI from 350 executives, so don’t treat it as physics, but it’s the largest published look at whether “the automation works” and “the company improves” are the same event, and the answer appears to be no.
McKinsey’s most recent State of AI, from November 2025, points the same way from a different angle. Nearly nine in ten organizations report regular AI use in at least one business function. Fewer than four in ten will claim it has touched EBIT at all, and most of those put the effect under 5%.
Automation multiplies whatever process you point it at. Point it at one that leaks and you get a bigger leak, running around the clock, with much better reporting. I’ve written before about why most AI projects are solving the wrong problem, and this is the commercial version of the same failure.
The failure is diagnosis, and there’s a trial that proves it
This is the part I find genuinely convincing, and it isn’t from a vendor.
Nicholas Bloom, Benjamin Eifert, Aprajit Mahajan, David McKenzie and John Roberts ran a randomized controlled trial in the Indian woven-cotton textile industry, published in the Quarterly Journal of Economics in 2013. They gave free management consulting to a set of plants and withheld it from a control group. Productivity in the treated plants rose 17% in the first year, through better quality, less waste and lower inventory. Nothing exotic. Standard practices, adopted properly.
The obvious question is why the firms hadn’t done any of it already. The answer is the reason I work the way I do.
Some of it was genuine ignorance. For the practices that were uncommon in the industry, the paper says firms were simply not aware they existed. But slightly over 45% of the initial non-adoption came from firms that had heard of the practice and thought it wouldn’t apply profitably to them.
That isn’t ignorance. It’s a wrong theory about their own business, held confidently, by the person best placed to know better.
Two caveats I’d rather state than have pointed out. It’s 17 firms and 28 plants in one industry in one country, with fieldwork between 2008 and 2011, and I’m not going to pretend it maps cleanly onto a plumbing company in Alberta. And the authors are careful that these were the initial barriers: once the information problem was solved, owner time and ordinary procrastination took over as the constraint. But it’s a randomized trial aimed straight at the thing every automation project assumes away, which is that the client correctly identified the problem before the discovery call started.
An aside on the numbers everyone quotes at you
While I was checking sources for this, two things fell over, and they’re worth knowing if you sit through automation pitches.
You’ve seen the claim that 80% of AI projects fail, twice the rate of ordinary IT projects, usually credited to RAND. RAND does have a 2024 report on AI project failure and it’s a good one, since the interviews with 65 practitioners are real research. But the 80% isn’t RAND’s finding. It appears twice, once in the summary and once in the introduction, both times hedged as “by some estimates,” and the endnote points at a July 2022 Fortune article. A magazine’s estimate became an authoritative statistic by getting footnoted in a think-tank report.
Then there’s MIT’s “95% of GenAI pilots return nothing,” which went around the world in August 2025. Its headline sentence is that 95% of organizations are getting zero return, which is a different claim to the one everybody repeats. In fairness to the people repeating it, the report is loose about this itself: elsewhere it calls the same number a “95% failure rate for enterprise AI solutions,” and the executive summary talks about pilots. Three different denominators, one number. As for the document: it’s a preliminary, non-peer-reviewed working paper out of the Media Lab, circulated as version 0.1, drawing partly on 153 survey responses collected at four industry conferences. No derivation for the 95% appears anywhere in it, and it isn’t listed on the Media Lab group’s own publications page.
I’m not saying either number is wrong. I’m saying almost nobody repeating them knows whether they’re wrong, including the agencies putting them on slide four to sell you the fix.
Apply the same suspicion to me. Both audits I quoted above trace back to one commercial interest: the 2011 HBR piece was co-authored by a co-founder and CEO of a lead-response software company, and the 2014 report is that same company’s. It doesn’t make them false, and behavioural audits with disclosed samples beat surveys, which is why I used them. But you should know who paid before you let a number move you.
What this looked like in my own work
A client once brought me an investment analysis for a sunflower oil production facility. The brief was a financial model: capacity, capex, margins, three scenarios, the usual.
I built the model. The model was not the answer. The thing that would have decided whether that facility lived or died was an export duty structure on the seed side of the border it was buying from, a policy detail sitting several steps upstream of anything in the spreadsheet I’d been hired to produce. Get that wrong and every scenario in the model is fiction, presented in a nice table.
I could have delivered the model, invoiced, and been technically correct. Plenty of people would have.
What’s usually actually broken
Almost none of it is an AI problem.
In a professional services firm, the world I know best, the money doesn’t leak where people point. It leaks out of non-billable time nobody logs. Overtime that became permanent and stopped registering as a signal. Meetings that exist because two people don’t trust the same document. Work done twice because the first version wasn’t findable. A handoff that fails quietly every Thursday and gets absorbed by whoever notices first.
Salesforce’s most recent State of Sales survey, self-reported across 4,050 sales professionals so weigh it accordingly, puts more than half of the average rep’s week on work that isn’t selling. Their own chart splits it 40% selling, 60% not. Whatever the equivalent number is in your business, you probably don’t know it. That’s the finding, not the 60%.
These are communication problems and process problems, and I keep saying so because they’re the good kind: they’re fixable, usually quickly, and usually without buying anything.
Then here’s what makes automation worth doing. Once you know which leak is actually costing you, the tool you buy is a different tool, narrower and cheaper and pointed at something you can measure before and after. That’s the whole difference between an automation that pays for itself in a quarter and one that becomes a Notion doc nobody opens. When I wrote about what a day of tax workflow automation actually changes, the useful part wasn’t the software, it was knowing which step in the week was eating the hours.
The order matters more than the technology. Diagnose, then build. Most people do it the other way around, because the build is the half with a proposal attached to it.
Should you hire an AI automation agency? Three questions first
These aren’t a framework. They’re just what I ask.
If this worked perfectly tomorrow, what breaks next? If you can’t answer that, the build is premature. Every business has a next constraint, and if you don’t know yours you’re about to spend money moving pressure onto it.
Where does the money actually leak? Not where it feels like it leaks, but where the numbers say it does. If you don’t have the numbers, that’s the project, and it’s cheaper than the automation.
Is this a technology problem, or a decision you’ve been avoiding? Usually the second, in my experience, and software is an expensive and sophisticated way to postpone a conversation. Firing someone. Raising prices. Killing a product line. Nobody buys an n8n workflow to dodge a hard decision on purpose. It’s just frequently what the invoice is paying for.
An agency that won’t sit through those three questions with you before it quotes isn’t an agency. It’s a vendor with a subscription.
I don’t start by building. I start by looking for the hole nobody’s pointing at, because a fair amount of the time the thing you were about to buy wouldn’t have moved anything, and a fair amount of the time, what would have moved something costs less than the retainer.
That’s the part I’d want you to leave with. None of this is an argument against automation. I build it. It’s an argument for knowing what you’re automating and why, which is cheaper than the build and almost nobody sells it.
If you’re about to sign an automation retainer, run the three questions first. Or book an hour and I’ll run them with you.
---
Sources
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, June 25, 2025. Source of the “agent washing” definition and the estimate that only ~130 of thousands of agentic AI vendors are real.
- InsideSales.com, 2014 Lead Response Report. Submissions attempted at 14,061 companies; 9,538 had a working form and received one; 4,472 (47%) never responded. Median first phone response 3h08m, average 61h01m.
- James B. Oldroyd, Kristina McElheran and David Elkington, “The Short Life of Online Sales Leads”, Harvard Business Review, March 2011. 2,241 companies audited; 23% never responded; 42-hour average response.
- Frederick F. Reichheld and W. Earl Sasser Jr., “Zero Defections: Quality Comes to Services”, Harvard Business Review, September to October 1990.
- Gartner, “Autonomous Business and AI Layoffs May Create Budget Room, But Do Not Deliver Returns”, May 5, 2026. Survey of 350 executives at $1B+ organizations, fielded Q3 2025.
- McKinsey & Company, The State of AI, November 5, 2025. n=1,993 across 105 nations.
- Nicholas Bloom, Benjamin Eifert, Aprajit Mahajan, David McKenzie and John Roberts, “Does Management Matter? Evidence from India”, The Quarterly Journal of Economics, Vol. 128, No. 1, 2013, pp. 1 to 51. RCT across 28 plants in 17 firms; fieldwork 2008 to 2011.
- James Ryseff, Brandon de Bruhl and Sydne J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND Corporation, 2024. The “more than 80 percent” figure appears as “by some estimates” and is endnoted to Jeremy Kahn, Fortune, July 26, 2022.
- Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT Media Lab / Project NANDA, 2025. Circulated as a preliminary, non-peer-reviewed working paper (version 0.1 in the distributed filename); no stable public URL, and it is not listed on the Media Lab group’s own publications page.
- Salesforce, State of Sales, fielded August to September 2025. 4,050 sales professionals across 22 countries; self-reported.
Further reading
- Building AI Solutions Beyond the Hype: the engineering version of the same argument, about picking the problem before the tool.
- AI Is a Feature, Not a Company: what investors have worked out about AI as a differentiator, which applies just as well to buyers.
- Vertical AI Agents: where agents genuinely earn their keep, once the process underneath them works.
- Tax Workflow Automation: A Day in the Life: a worked example of automating the step that was actually costing the hours.