Where AI development sits in 2026

We work with small businesses who keep getting pitched on AI. Most of the pitches are overstated. Some are not. This is a short accounting of what’s actually shipping in production right now, written for founders and operators who want to know what to take seriously and what to ignore.

Agentic AI is real, but narrower than the demos suggest

If you went to a tech conference in 2025, “agents” was the word. The demos look impressive: a language model picks tools, clicks through a browser, fills in a form, drafts an email, books the meeting.

In production, the useful version of this is much less ambitious. It looks like:

• A support agent that routes tickets, drafts a first reply, and waits for a human to send it.
• A data-entry agent that takes a PDF invoice, pulls out vendor, date, line items, and totals, and writes them to a spreadsheet.
• A research agent that reads your internal docs and answers questions with citations.

All of these are real and shipping. The fully-autonomous, multi-step agent that runs your business while you’re at the beach is not, and the companies trying to sell you one are optimistic.

GitHub’s own numbers back this up: its Octoverse 2025 report puts the year at roughly a billion commits, up about a quarter year over year, with pull-request volume up a similar amount and a Copilot “coding agent” credited with over a million merged PRs in five months. Read that the right way, though: it’s developers using AI as a faster typewriter, not companies firing engineers. More code per engineer, though the same report shows commit comments dropping, which tells you something about how carefully all that code is being read.

Small models got good enough for most tasks

Two years ago, you needed GPT-4 class models for anything non-trivial. In 2026, a small open-weight model (Llama 3.x, Qwen3, Gemma 3, a small Mistral, roughly 8–14B parameters) runs on a single consumer GPU and handles:

• Customer support replies.
• Document summarization.
• Classifying, tagging, and routing.
• Structured extraction from messy text.

For a specific task in a specific domain, a smaller fine-tuned model often beats a larger general one on both cost and accuracy. The frontier labs still matter for hard reasoning problems, but the median use case has moved down.

Practically, that means you can run a lot more AI locally than you could last year. Cheaper per request, no data leaving the building, and no dependency on whoever’s pricing page you’ve been staring at.

Multimodal is where the practical wins are

Text-only models were a bottleneck. A lot of the real work businesses want automated (invoices, receipts, ID documents, chart screenshots, warehouse photos) has always been in images or PDFs with mixed layouts.

In 2026, vision models are good enough to read those inputs reliably. The applications are unglamorous: extracting line items from a receipt, checking a warehouse photo for damage, reading a handwritten form. But they’re the ones that actually save operators time.

Physical AI is still slow

There’s a lot of excitement about robots and “physical AI.” NVIDIA has made the phrase a centerpiece of every keynote, and the humanoid demos from Figure and Tesla are the poster children. In warehouse and manufacturing environments, where the task is constrained and repetitive, there’s real deployment happening. Anywhere else, it remains slower and more brittle than the videos suggest. If a vendor tells you their robot will handle a general task in an unstructured environment, ask for a reference customer.

What this means for a small business

If you own a tiffin service or a painting company, you don’t need a strategy for “AI.” You probably need one or two specific things automated:

• A cheaper first-line support reply that a human reviews before sending.
• Receipt or invoice processing that doesn’t require someone retyping totals into a spreadsheet.
• Search over your own content (FAQs, policies, SOPs) that doesn’t require paying an employee to remember everything.

Those are all tractable with current tools, run cheaply, and don’t require a platform rewrite. The rest (the autonomous agents, the digital workforce, the enterprise transformation) is mostly a longer timeline than the pitch suggests.

Update, August 24, 2026. Seven months on, the call at the top of this post has held: the narrow, human-in-the-loop version of agents is what ships, and the autonomous version is still mostly a demo. What has genuinely changed is underneath, at the plumbing layer, and it is worth one paragraph.

The two competing standards for how agents talk to things have stopped competing. Anthropic’s Model Context Protocol, which standardises how a model reaches a tool or a database, was contributed to the Linux Foundation’s Agentic AI Foundation when that body was formed in December 2025. On August 20, 2026, Google’s A2A protocol, which covers the other axis, how separate agents negotiate tasks and exchange identity across organisational boundaries, moved under the same foundation. The AAIF now lists AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI at its top membership tier.

For a small business the practical read is about lock-in, not capability. A year ago, committing to an AI vendor meant committing to that vendor’s way of connecting to your systems. That is now much less true, and the reasonable posture when a vendor pitches you an agent is to ask whether it speaks these standards or a proprietary integration you cannot take with you.

One thing has not improved and is not on a roadmap: the security limit. Anything you let an agent do autonomously is bounded by prompt injection, which is not a bug anyone is going to patch. That constraint, and the simple test for deciding what an agent should be allowed to touch, is in our piece on agents and prompt injection. It is the thing to read before signing anything.

If you want help figuring out which one of those actually applies to your business, send us a note.