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

I keep getting asked the same question by operators planning their 2026 budget. Where are we on the curve. The honest answer is that there is no single curve. LLM assistants inside software people already use are mainstream and boring. Autonomous agents are a demo culture with a thin production tail. And the model labs themselves are quietly walking back some of the most ambitious 2024 promises. If you're choosing what to fund this year, the useful exercise is sorting your shortlist into those three buckets before you read another vendor deck.

The data that shaped this piece comes from a few places I trust more than vendor surveys. The MIT NANDA group's 2025 GenAI Divide study put a hard number on a feeling everyone in this work already had. Roughly 95% of enterprise GenAI pilots produced no measurable P&L impact. The Stack Overflow Developer Survey 2024 shows the opposite picture in one specific corner: code assistants. Daily developer use jumped past 60% in a year. And the McKinsey State of AI tracker keeps showing the same uncomfortable gap between adoption and reported financial benefit.

The fight that's already over is the one almost nobody is reporting on. Code assistants, transcription, drafting inside email and docs, retrieval inside CRM and helpdesk software. GitHub Copilot, Cursor and Windsurf together moved the median engineering org to a place where opting out feels strange. Microsoft Copilot inside Office and Salesforce's Einstein layer turned LLM features into checkboxes the buyer barely thinks about. None of this is interesting anymore, and that is the point. When a category gets boring, it has crossed the chasm.

From the few projects we've shipped this year, the pattern that holds is the same. The thing that gets a result fast is almost always an LLM stitched into a workflow the team already runs, not a green-field agent. A draft-generation step inside an existing ticket pipeline. A retrieval layer on top of a knowledge base nobody had time to maintain. A structured-extraction job on PDFs that used to take a junior two hours each. If you're trying to figure out where to put the first dollar, our customer support playbook walks through one of the boring-but-pays variants in depth.

The clearest tell is Klarna. They very publicly launched an OpenAI-powered support bot in January 2024 and spent a year telling the press it replaced 700 agents. By mid-2025 they were quietly rehiring humans and Sebastian Siemiatkowski was on stage admitting service quality had slipped. That isn't a one-off. It is the predictable shape of the cycle. Salesforce Agentforce launched into the same narrative. Cognition's Devin became a meme about the gap between demo video and real PR throughput. Microsoft's own internal Copilot rollout numbers, where leaked, suggest engagement that drops off after the first month.

“

The autonomous agent wave is roughly where chatbots were in 2017. Impressive demos, terrible at the long tail and most enterprise teams will regret going first.

I'm not saying agents won't work. I'm saying the gap between what a clean demo shows and what holds up against a real customer base on day 90 is much wider than vendor decks suggest. The MIT NANDA number was 95% of pilots producing no ROI, and most of those pilots were exactly the agent-shaped ones. If you're being pitched a multi-step autonomous workflow that touches money, refunds or anything legally regulated, that's the bucket I'd put it in. The piece on agents vs chatbots vs copilots sketches the boundary in more detail.

The other thing I find people aren't tracking is the tone shift inside the model labs. Apple Intelligence shipped late, with the more ambitious Siri overhaul pushed past 2026. Anthropic's positioning between Claude Sonnet 4.5 and Opus 4.7 is much more about cost per task and reliability than about new capability tiers. OpenAI keeps re-pricing access rather than announcing a generational leap. None of this is a crisis. It is the normal flattening that happens after a bubble of expectations. Capability is still improving, but the marginal next-year gain looks more like 15% on hard tasks than the 5x jumps that drove 2023 and 2024.

The practical implication for an operator is the opposite of what most people read into it. If capability is plateauing, the advantage stops accruing to whoever has the smartest model and starts accruing to whoever wires the existing models into their business the most carefully. That makes 2026 a build year, not a wait year. Which is a strange thing to say given the cooling vibes, but I think it's correct.

If you have one quarter of attention and a modest budget, here is what the pattern I keep seeing in production says to do. Pick one workflow with a number attached to it. Response time, draft hours per week, manual extractions per month, qualified leads per rep. Pick an LLM-inside-existing-tool approach, not a new agent. Give it a six-week kill date. If the number moves, expand. If it doesn't, stop and pick a different workflow. Skip anything that requires a press release. The teams who do this in 2026 will be in a much better position to evaluate agents in 2027, when the agent stack has had another year of failure data to learn from.

If you want a second pair of eyes on which workflow is the right first pick, that's the conversation we have at the start of every engagement. The services overview describes how we run it.