Most businesses hit the plateau without noticing it. The tools are still running, the subscriptions are still active, and the team is still using them, but the output gains that justified the rollout stopped accruing months ago. An NBER working paper published in February 2026, surveying nearly 6,000 CEOs and CFOs across the US, UK, Germany, and Australia, found that 89% of firms reported zero productivity impact from AI after three years of use. Average gains across the entire sample: 0.29%. That is not a rounding error. That is what the first wave actually produced.
The businesses that are surprised by this are the ones who mistook the first wins for a compounding curve. They weren’t. They were a one-time extraction of value from the lowest-complexity layer of work, and that layer is now mostly automated.
What the First Wave Actually Automated (And What It Missed)
High-volume, low-judgment tasks were the easy layer
Drafting emails. Generating image variants. Summarising meeting notes. Producing first-draft marketing copy. These tasks had one thing in common: a human was doing something repetitive that a language model could replicate adequately. The time savings were real, but they were bounded by the size of that task category in your business.
A 10-person agency that used to spend 4 hours a week on first-draft copy now spends 1 hour. That’s a real gain. It is also finite. You can only save those 3 hours once.
The tasks that drive revenue were never in scope
Client relationships. Quality judgment on complex work. Pricing decisions. Business development. The work that actually determines whether a firm grows, that work was never in the first wave. AI tools in 2023–2025 were not built to handle it, and most businesses didn’t try to apply them there.
That gap is where the plateau lives. You automated the supporting layer. The load-bearing layer stayed human.
Why the Gains Flatten, Four Structural Reasons
Commoditisation: your competitors have the same tools now
In 2023, using ChatGPT to produce marketing copy faster than a competitor was a short-lived edge. In 2026, it is table stakes. Every copywriter, every agency, every in-house team has access to the same models. The advantage evaporated when adoption normalised, which IDC’s 2026 FutureScape analysis calls the “productivity diffusion problem”: efficiency gains spread across markets faster than firms can redeploy them as competitive advantage.
Task ceiling: you run out of low-complexity work to automate
The easy layer is finite. Once you have automated email drafts, image resizing, meeting summaries, and social post generation, there is no next tranche of equivalent tasks waiting. The remaining work is either higher-complexity (where AI underperforms) or already handled by existing software (where AI adds nothing).
Data and workflow gaps block the next layer
Moving from “AI writes our first-draft copy” to “AI manages our client onboarding workflow” requires clean, structured data, integrated systems, and well-defined processes. Most SMBs don’t have those. The tools could theoretically go further, but the underlying infrastructure stops them. You can’t automate a workflow that isn’t documented. You can’t feed client context to an AI that’s siloed from your CRM.
Expectation debt, what you thought AI would do versus what it did
Many SMBs were sold AI as a headcount replacement tool. They didn’t replace headcount. They saved partial hours across multiple roles, hours that employees absorbed into other tasks rather than creating measurable output gains. The productivity promise was real at the task level and invisible at the business level. That gap between expectation and outcome is now accumulating as scepticism.
The Data Behind the Plateau
NBER 2026: 0.29% average productivity gain across 6,000 firms
The NBER Working Paper No. 34836 is the most rigorous data set on this question published to date. The 0.29% figure covers firms that had been actively using AI tools for at least three years, not sceptics or late adopters. These were early movers who committed to AI integration. The headline result is not that AI doesn’t work. It’s that the average real-world impact is a fraction of what most projections suggested.
Where gains are real: customer service and software development
The same research tradition that shows aggregate underperformance also shows meaningful gains in specific narrow functions. AI-assisted customer service agents show approximately 14% productivity gains. Software developers using AI coding assistants (GitHub Copilot and comparable tools) show approximately 26% gains on certain task types. These are real. They are also specific, they appear in high-volume, rule-governed tasks where AI can pattern-match reliably. Outside those conditions, results drop sharply.
The self-reporting problem
A May 2026 METR survey of 349 technical workers found that respondents estimated their AI-driven productivity gains at roughly 2x, but when METR validated those claims against objective output measures, the actual gains were approximately 40 percentage points lower than self-reported. People feel more productive when using AI tools. That feeling is not always reflected in output.
This matters because most AI ROI projections inside businesses are based on self-reported estimates. The data underneath those business cases is structurally inflated.
What Comes After, Three Honest Paths for SMBs
Path 1: Deepen existing automation (smaller gains, lower risk)
You haven’t squeezed all available value from the tools you already have. Most businesses use 20–30% of the capability of their current AI stack. Better prompt engineering, tighter integration with existing software, and more consistent usage across teams can extract additional value without new investment. Gains here are real but incremental, expect 10–15% improvement on processes already automated, not step-change impact.
Path 2: Redesign workflows around AI-native processes
This is the higher-ceiling path and the harder one. Instead of asking “which existing tasks can AI help with?”, you ask “if we designed this process from scratch knowing AI existed, what would it look like?” That requires documented workflows, clean data, and willingness to change how work gets done, not just what tool sits alongside the existing process. A firm that redesigns its client intake process around AI-assisted qualification and scoping, rather than just having AI write the intake emails, can see structural efficiency gains where inputs are clean and the process is well-defined. It is not a tool purchase. It is an operational redesign, and it fails when the underlying data or processes are messy, which is most of the time on the first attempt.
A custom-built WordPress site with properly structured content types is a simple example: it lets AI tools operate on clean, queryable data rather than a spaghetti of plugins and unstructured posts. Infrastructure decisions made without AI in mind often become the hidden ceiling on AI capability.
Path 3: Strategic retreat, audit and cut tools that aren’t delivering
This is the path nobody in the AI sales ecosystem wants you to take, which is probably why it’s the most underused. Some AI tools in your current stack are not delivering measurable value. They were purchased on hype, adopted partially, and are now a line item in your SaaS budget that nobody can justify clearly. Cutting them is not a failure. It is a correction.
Starting with exactly this question, which tools are delivering measurable output and which are generating the feeling of productivity without the substance, is how we approach AI stack reviews at designodin.com/ai.
What We See in the Field
The pattern with SMB clients: automation theatre
Automation theatre is when a business deploys AI tools visibly, announces them internally, trains staff, writes the case study, and the tools then become part of the workflow in name only. Staff use them intermittently, inconsistently, and often for low-value tasks. The tools are “in place” but not integrated into the work that drives revenue. When we review existing AI stacks, this pattern appears in roughly half the businesses we work with. They haven’t failed at AI. They’ve succeeded at adoption theatre.
The honest AI audit: what to measure and what to cut
An honest audit has three questions. First: which specific tasks are you completing differently because of AI, and how long does each take compared to before? Second: which of those tasks are in the direct path of revenue generation versus the supporting layer? Third: which tools are you paying for that your team has quietly stopped using? The answers to those three questions will tell you whether you’re in a genuine productivity gain, a plateau, or a decline disguised by subscription inertia.
Frequently Asked Questions
Is the AI capability plateau permanent or temporary?
The current plateau reflects two overlapping problems: diminishing returns on first-wave automation, and the gap between AI capability and the complexity of real revenue-driving work. The plateau in marginal returns from current tools is largely structural, you can’t re-automate tasks you’ve already automated. The capability plateau may shift as agentic AI systems (those that take multi-step actions without human approval at each step) mature, but those systems introduce new failure modes that most SMBs are not equipped to manage yet. Expect the plateau to persist for most businesses through 2026–2027.
How do I know if my business has hit the plateau?
The clearest signal is that your AI tool spend is flat or growing while measurable output gains have stalled. A secondary signal: your team is using AI tools but can’t articulate what would break if those tools disappeared tomorrow. If AI tools are genuinely embedded in revenue-driving work, their removal would be felt immediately. If removal would cause mild inconvenience, you’re in the plateau.
What’s the difference between the AI productivity plateau and the AI capability plateau?
The productivity plateau is a firm-level phenomenon: your business has extracted the available gains from its current AI deployment, and incremental returns are declining. The capability plateau is a technology-level phenomenon: the underlying models and systems stopped improving meaningfully for real-world task completion, even as benchmark scores continued to rise. Both are happening simultaneously, which is why productivity surveys are starting to reflect what the NBER data shows: deployment without redesign reaches a ceiling quickly.
Should SMBs invest in agentic AI to break through the plateau?
Carefully. Agentic AI, systems that execute multi-step tasks autonomously, can extend the automation frontier beyond low-complexity single-step tasks in the right setup. The risk is that agentic systems fail in ways that are harder to detect than simpler tools. A language model that writes a bad email is obviously wrong. An agent that misroutes a client inquiry through a multi-step process may not surface the error until it has compounded. For SMBs with clean data, documented workflows, and the operational capacity to monitor autonomous systems, agentic AI is worth piloting in constrained, low-stakes processes first.
What’s the first step after you’ve exhausted easy automation wins?
Audit what you have before buying anything new. Map every AI tool against a specific measurable output, not “we use it for content” but “it reduces first-draft time from 2 hours to 30 minutes on X task type, which appears 8 times per week.” If you can’t state it that precisely, the tool is probably in automation theatre. The audit takes a day. The clarity it produces is worth more than any new tool purchase.
The plateau is not a failure. It is a calibration signal telling you that the easy layer is done and the next layer requires a different approach. Most businesses stalled because they automated outputs, the visible, countable deliverables, instead of redesigning the inputs: the processes, data, and decisions that determine what outputs are worth producing.
If you want to talk through where your current AI stack is and isn’t delivering, start a conversation. We scope what a workflow redesign would actually involve before any commitment. See how we approach this at designodin.com/ai.