AI Is Eating Software. Here's What SMBs Should Actually Build in 2026.

Every April, I find myself writing some version of the same calibration exercise. Not predictions — I have no idea what the next breakthrough model will be or when it’ll drop. What I can do is take stock of what’s actually working in production, what’s still mostly a demo, and where the ROI is clear enough that any business should be acting on it now.

This is that post for 2026.

I’ve been building AI systems for SMBs for several years. The vantage point is useful not because I’ve seen what’s possible in research labs, but because I’ve seen what’s actually deployable, maintainable, and valuable for businesses that don’t have a dedicated ML team, a data science budget, or a tolerance for systems that require constant babysitting.

What follows is not a hype piece. It’s a field report.

What’s Real and Working Right Now

Document intelligence at production quality

Extracting structured data from unstructured documents — contracts, invoices, receipts, forms, emails — is genuinely solved at a level that makes it worth operationalizing. Not “sometimes works in demos” solved. “Runs reliably in production at 90%+ accuracy with appropriate human review for the edge cases” solved.

If your business processes documents — and almost every business does — and you have humans manually keying information from those documents into a system, that’s one of the highest-ROI AI automation opportunities available right now. The technology is mature. The cost is low. The ROI case is straightforward.

RAG-based knowledge retrieval

Building a retrieval-augmented generation system over your internal documents — SOPs, wikis, product documentation, historical project notes — works well today and costs very little to operate once built. The era of “our knowledge is scattered across Google Drive and three Notion workspaces and nobody can find anything” is over if you want it to be.

The caveat: your data quality determines your retrieval quality. RAG doesn’t fix bad documentation; it surfaces it more efficiently. If your internal knowledge is outdated, inconsistent, or incomplete, a chatbot over that knowledge will be outdated, inconsistent, and incomplete. The value of RAG scales directly with the quality of what you’re retrieving from.

AI-assisted QA and testing

I called this the highest-ROI AI use case in March, and I’ll say it again here: if you ship software and your testing story involves manual QA cycles, automated tests that break constantly, or “we’ll catch it in production,” AI-assisted testing is one of the best investments you can make. LLM-generated test cases, adaptive Playwright selectors, log analysis for CI triage — these are production-viable right now and dramatically reduce the cost of maintaining test coverage.

Content operations pipelines

The Discord + n8n + LLM approval loop I wrote about in March is one example, but the broader pattern applies to any team that produces content regularly. Drafts, summaries, social reformats, SEO metadata, internal reports — these are high-volume, low-risk candidates for AI assistance. The key word is “assistance”: the content still passes through human judgment before it ships, which is what keeps the quality bar where it belongs.

Business team reviewing digital dashboard with AI insights — practical tools for daily operations

What’s Overhyped (For Now)

Fully autonomous agents doing complex multi-step business work

Autonomous agents that can take a complex business goal — “go find us three qualified sales leads, draft personalized outreach, schedule calls, and follow up” — and execute it end-to-end without human involvement are impressive in demos. In production, at scale, for anything of consequence, they’re still brittle.

The current state: multi-agent systems work well when the scope is narrow, the tasks are well-defined, the tools are reliable, and there’s a human approval checkpoint for anything high-stakes. They work poorly when you hand them an ambiguous goal and expect them to figure it out. The fully autonomous autonomous AI agent that makes good decisions across genuinely complex business contexts is coming — it’s just not quite here yet in a form that’s safe to hand consequential work to without oversight.

Build agentic workflows. Don’t remove the humans from them yet.

AI replacing your entire customer service operation

I have watched this project fail enough times to have opinions about it. The dream: AI handles all support tickets, escalates only the truly complex ones, saves the cost of a support team. The reality: the edge cases are more numerous than expected, the AI’s handling of emotional conversations is notably bad, and customers who don’t feel heard leave.

Where AI wins in customer support: first-response drafting for human review, automatic triage and routing, FAQ deflection for genuinely simple queries, post-resolution summaries. Where AI loses: handling frustrated customers with complex, emotionally charged situations where the person needs to feel like someone actually gives a damn about their problem.

Use AI to make your support team faster and better. Don’t use it to eliminate the team.

Custom LLM fine-tuning as a first step

Fine-tuning a model on your specific data sounds like the natural first step for any AI project. It’s rarely the right first step. Prompt engineering on a base model — the cheap, fast, no-training-required approach — solves most business problems without the cost, complexity, and ongoing maintenance overhead of a fine-tuned model. Fine-tuning is the right tool for specific classes of problems: needing a much smaller, faster model, producing output in a highly specific format, or achieving performance on a very narrow domain that RAG can’t match.

If someone is pitching you fine-tuning as the foundation of a first AI implementation, ask why prompt engineering with retrieval wouldn’t accomplish the same goal at 10% of the cost and timeline.

Where the Actual Opportunity Is for SMBs

The biggest AI opportunity for most SMBs in 2026 isn’t the most sophisticated implementation. It’s the most targeted one.

The companies getting the clearest ROI from AI right now are not the ones building the most ambitious systems. They’re the ones who identified one expensive, time-consuming, high-volume process — document extraction, test generation, content drafting, data classification — and built something narrow and reliable that handles that process well. Then they built confidence in the technology and expanded from there.

The no BS consulting playbook I wrote in January still applies: start with the pain, build the vertical slice, ship it, iterate. The AI technology has never been more accessible or more affordable. The constraint is almost never technology. It’s scope discipline and the willingness to define success narrowly enough to actually achieve it.

Single thread of light in darkness — the power of a narrow, focused implementation

What to Stop Waiting For

Two years ago, the reasonable position was “let’s wait for this to mature before investing.” That window has closed for most use cases. The technology is mature enough. The tooling is mature enough. The cost is low enough. The companies that are building now are building institutional knowledge, internal muscle, and competitive positioning. The companies still waiting are falling behind.

This doesn’t mean build something complex. It means stop using “the technology isn’t ready” as the reason not to start. Because for the use cases I listed above — document intelligence, RAG, AI-assisted testing, content operations — the technology has been ready for eighteen months, and the cost of not acting is accumulating.

The question isn’t “is AI ready?” It’s “what’s the one thing I could build in the next 90 days that would pay for itself by the end of the year?”

That’s the question worth sitting with.

I’ve been publishing field notes on what’s actually working in AI for SMBs since January — if you want to go deeper on any of the topics I covered above, the full series is here:

And if you want to talk through what the right first AI project looks like for your business — I’m available.