Back to Insights Educational article

The Hidden Cost of Not Using AI, Honestly Counted

The honest cost of not using AI in 2026 is a 10 to 20 percent efficiency tax on a few specific workflows, not extinction. Here is how to tell which workflows, and when the tax has actually come due for yours.

Last updated: May 24, 2026 AI Adoption

~9 min read Reading time 7 Sections Data guide Data-rich article

The honest cost of not using AI in 2026 is not extinction. It is a 10 to 20 percent efficiency tax on a handful of specific workflows, and any consultant telling you it is existential is selling you something.

01 The frame I want to push back on

Inaction is a cost, not a death sentence

I read three of these articles a week. Most of them sound the same. The pitch is some version of go fast or die. The Kodak reference shows up by paragraph four. The Blockbuster reference shows up by paragraph six. The reader is told that any business not deploying generative AI right now will be out of the market in eighteen months. This framing is wrong, and worse than wrong, it is unhelpful for operators trying to make budget decisions. The pattern I keep seeing in conversations with founders and operations leads is that the FUD has the opposite of its intended effect. It produces either paralysis or panic spending, and both of those waste more money than doing nothing for another quarter would.

So let me try a different version. There is a real cost to ignoring this technology, but it is narrower and more concrete than the doom-mongers say. From the few production deployments I have helped with, and from talking to operators running their own, I think the honest picture is this. A small number of workflows get measurably faster or cheaper when you put an LLM in the loop. Most do not. The companies that win in 2026 are not the ones that bought the most credits, they are the ones that figured out which of their specific processes belong in the small bucket, and were disciplined enough to ignore the rest.

02 Where the cost is real

Three workflows where inaction actually shows up

Drafting, lookup and classification. Those are the three categories where the productivity delta is large enough and consistent enough that you can defend a number to a CFO. The GitHub research on Copilot from 2022, the original randomized controlled study, found developers using the tool completed a specific server-writing task about 55 percent faster than the control group. That study has limits. The task was narrow, the population was self selected and later field studies on real PR throughput came in lower, more like 10 to 25 percent on net cycle time depending on language and seniority. But even at the low end, on a team of fifteen engineers, you are looking at the equivalent of one or two additional headcount worth of throughput per year for the price of a per seat license.

Customer support tier one is the second obvious one. Klarna very publicly announced in early 2024 that their OpenAI powered assistant was handling two thirds of customer service chats in its first month, equivalent to seven hundred full time agents. That number went around the internet a hundred times. Less reported is that by mid 2025 Klarna walked the framing back, said they had cut customer service quality in places and announced they were hiring human agents again. The honest read is not that the AI failed. It is that the AI handled volume well and handled nuance badly, and the right answer was a hybrid model rather than the original wholesale replacement. Intercom Fin charges around 99 cents per resolution and quotes resolution rates in the 50 percent range on tier one queries. Those numbers are real. The trap is assuming they generalize past tier one.

The third is internal lookup and document classification. Anything that used to require a knowledge worker to read a long document and answer a structured question about it. Insurance claim triage, sales call summarization with CRM field updates, contract red flagging, RFP intake. The economics there are not as dramatic as drafting code, but they are stable. You can usually take a process that took someone fifteen minutes and get it to two minutes with a retrieval pipeline plus a smaller model. Multiply that across a year. The customer support playbook covers the support specific version of that calculation in detail, and our note on measuring AI ROI covers what to actually count and what to ignore.

03 Where the cost is fake

What inaction does not cost you

Most of the things consultants tell you AI is costing you, AI is not actually costing you. High stakes judgment, novel reasoning, original strategy work, hands on craft, anything where the bottleneck is wisdom rather than information throughput. There is no version of 2026 where the small architecture firm with two principals loses to a fully automated competitor. The bottleneck for that firm was never the typing speed of their drawings, it was the judgment built up over twenty years of project work.

The strongest recent evidence on this is the MIT NANDA State of AI in Business 2025 report, which got a lot of press for the line that 95 percent of corporate generative AI pilots show no measurable P&L impact. The honest read is not that AI does not work. The honest read is that most enterprises are deploying it badly, in places it does not belong, with no clear measurement, with no integration into the actual workflow. The 5 percent that did work tended to follow the same pattern. One specific workflow, clear baseline metrics, integration into the system of record and an owner accountable for the number. Boring stuff. The boring stuff is what works.

“ If the headline number on a vendor pitch sounds like it would solve your entire P&L, the vendor is either lying or talking about a workflow you do not have.

04 The contrarian section

Why doing nothing is sometimes the right call

I do not say this often in public because it is bad for our business, but here it is. There are operators I have spoken with where the right answer to do we need to start using AI was no, not yet, not this year. Three patterns. First, the business has under twenty employees and the partner ecosystem in their vertical has not produced a SaaS that wraps the relevant LLM call cleanly. Building it yourself costs more than the inefficiency it would remove. Wait six months for the off the shelf tool. Second, the business is in a regulated vertical where the disclosure and audit overhead of putting customer data through a third party model exceeds the labor savings. Healthcare back office and parts of regulated financial services keep falling into this bucket. Third, the business is in a moment of organizational instability where the change capacity is already spent. Layering an AI rollout on top of an ERP migration or a leadership transition is asking for the kind of failure that poisons the well for the next attempt.

None of those three patterns are permanent. The right move is to revisit in two quarters, not to declare permanent abstention. But the FUD playbook treats any delay as catastrophic, and that is wrong. Delay is sometimes prudence.

05 Vendors and what they actually do

Calibrating against real product launches

FIG. 01 - VENDOR PROMISES VS REALITY What three high profile launches actually delivered

Launch – The headline – What it actually looks like
Klarna OpenAI assistant, 2024 – Replaces 700 agents, saves 40M USD – Walked back in 2025; hybrid model now; volume coverage was real, quality was not
Salesforce Agentforce, 2024 – Autonomous agents handle sales and service – Mostly used as a better workflow builder; full autonomy still rare in production
Intercom Fin, 2023 to 2026 – Resolves customer queries end to end – Strong on tier one FAQ; escalates the moment context gets complex; per resolution pricing keeps it honest

None of these are bad products. The pattern is that vendor marketing always overshoots actual deployed value by 30 to 50 percent. If you build your buying decision around the launch press release, you will be disappointed. If you build it around the case studies that are six to twelve months post launch, you will get a much more accurate number. McKinsey publishes a running version of this calibration in their State of AI survey, which is worth reading not for the headline adoption numbers but for the section on where dollar value is actually being captured.

06 Reading the signals

When the efficiency tax has come due

Honest reasons to move now

A specific workflow eats more than 20 hours of staff time per week
You have a baseline metric on that workflow and a way to remeasure it
Your competitors have already shipped a customer facing AI feature you cannot match
Senior engineers in interviews ask whether AI coding assistants are allowed

Bad reasons that look honest

A board member read an article and now wants an AI strategy by Friday
A vendor offered a discount that expires this quarter
A competitor announced something on LinkedIn (announcements are not deployments)
Fear of being seen as the laggard at the next industry conference

What I tell operators in this position

If you cannot point to a workflow that costs you more than the implementation will, the right answer is to wait. Buy a per seat coding tool for the engineering team if you have one, that decision pays back fast. Beyond that, watch your customer support volume and your sales cycle length for the next two quarters. If either degrades by more than 10 percent against your baseline and your competitors have shipped something obviously AI flavored, that is the actual signal to start scoping. Until then, the efficiency tax is small enough that calm beats panic.

07 How we would actually approach this

Boring discipline beats heroic adoption

When operators bring us in, the first conversation is almost never about model selection or vendor choice. It is about which one workflow to pick. We push hard against the urge to do three things at once. One workflow, one owner, one baseline number, ninety days to a deployed system the team uses without prompting. If that works, the second one is easier to fund and easier to scope. If it does not work, you have a single bounded failure to learn from rather than a portfolio of partial pilots that nobody can interpret. The ninety day implementation roadmap walks through what that first cycle actually contains. The implementation cost note covers what to expect on the bill so the budget conversation is real.

If you want to talk through which workflow that should be in your specific case, that is most of what we do. No deck, no AI maturity assessment, just a working session on your actual operations.

Frequently asked questions

Is there a number for the cost of waiting another year?+ For most mid market operators the realistic range is a 10 to 20 percent efficiency tax on two or three specific workflows, not on the whole P&L. Whether that is meaningful depends on how much of your cost base sits in those workflows. The ROI measurement note covers how to size it on your own numbers rather than borrowing a vendor average. Is the AI advantage permanent or will it level out?+ Some of it will level out as off the shelf tools catch up. That is good news if you have not started yet. The gap between an early adopter and a late adopter on tier one customer support, for example, will narrow as the underlying SaaS converges. The gap on proprietary internal workflows that someone has spent eighteen months tuning will not narrow as fast. Can a small business actually compete with AI enabled rivals?+ Yes, in most verticals. The off the shelf tooling for small businesses in 2026 is far better than what enterprises had access to in 2023. The risk for a small business is not that AI exists, it is buying three overlapping tools without ever wiring any of them into the actual workflow. Is this all just a hype cycle that will deflate?+ Parts of it will deflate. Autonomous agents in particular are getting marketed well past what is shipping. The underlying productivity gains on drafting and lookup are not a hype cycle, they are a real efficiency shift on a narrow set of workflows. Bet against the marketing if you want, do not bet against the underlying capability on those specific tasks. What is the worst case if we just keep doing things the old way?+ Not extinction. Margin compression of a few points per year on the affected workflows, slower hiring of senior engineers who now expect AI tooling and a harder catch up project in 2027 or 2028 when you decide to start. Manageable, not existential. How do we tell a real signal from FUD when reading these articles?+ If the article talks about Kodak or Blockbuster in the first three paragraphs, lower your confidence in the rest of it. If it gives you a percentage without naming the source study, lower it further. If it names a specific company, a specific workflow and a specific measurement window, that is the part worth keeping.

ANM SOLUTIONS / CONTACT US

Need help applying this to your business?

We turn AI insights into measurable business outcomes. Tell us about your workflow and we'll show you the highest-impact place to start.

AI Adoption

The 2026 AI Adoption Curve: Where Things Actually Sit at Mid-Year

An operator's read on the 2026 AI adoption curve. What's gone mainstream, what's cooling off and where to spend a 2026 budget without getting burned by agent hype.

Read article

Implementation

The 90-Day AI Implementation Roadmap: From Scope to Production

A practitioner's view of the first 90 days of an AI implementation. Boring stack, eval-first, kill switch by day 90. Where most plans go wrong and what to do instead.

ROI

Measuring AI ROI: How to Get a Number You Can Defend (2026)

An opinionated guide to measuring AI ROI for CFO-leaning operators. Three return layers, a defensible formula, honest evidence standards and when to wind a project down.

Published: 2026-05-04 · Author: A&M Flow