AI in practice · 6 min

What AI can and can't automate in 2026

Published
Updated
Reading time
6 min

Why an honest map matters

The AI vendor landscape is incentivized to oversell. Every tool claims to "automate" something, but the gap between a compelling demo and reliable production use is enormous. Teams that understand this gap make better investment decisions and avoid expensive disappointment.

This is a practical map of what works, what partially works, and what does not, based on building AI features into our own products, not marketing claims.

Works reliably today

These are workflows where AI performs at or near human level consistently, with minimal supervision after initial setup.

Document data extraction. Pulling structured data from invoices, contracts, receipts, forms, and reports. Vision-language models handle messy layouts, inconsistent formats, and even handwritten text. Accuracy is typically 90 to 97% depending on document quality, and the remaining errors are caught easily in a validation step. This is the most mature and proven AI automation category.

Text classification and routing. Categorizing support tickets, emails, feedback, or any text into predefined categories. Intent detection, sentiment analysis, urgency scoring, topic tagging. All work reliably when the categories are well-defined and the prompt is engineered to match your specific taxonomy. The key word is "well-defined". Vague categories produce vague results.

Content summarization. Condensing long documents, meeting transcripts, research papers, or email threads into structured summaries. Models are very good at extracting key points and decisions. Works especially well when you specify the output format: "summarize as 5 bullet points covering decisions, action items, and open questions."

Translation. For most business content (emails, documentation, marketing copy, support responses), AI translation quality now matches or exceeds average human translators. For highly specialized or culturally nuanced content, human review still adds value, but as a refinement step, not a replacement.

Code generation for defined tasks. Writing boilerplate, generating test cases, converting between formats, creating API integrations with known specifications. When the task is well-scoped and the developer reviews the output, AI reliably accelerates development.

Works with guardrails

These workflows produce real value but require ongoing human oversight, confidence thresholds, or fallback paths. Deploying them as fully autonomous systems is a mistake.

Customer-facing responses. AI can draft support responses, FAQ answers, and email replies that are accurate and well-written. But "draft" is the operative word. Tone misjudgments, incorrect edge-case handling, and hallucinated policy details happen often enough that a human should review before sending, especially for high-value customers or sensitive topics. The best implementations use AI to draft and a human to approve, cutting response time without cutting quality.

Lead scoring and qualification. AI can enrich leads with company data and score them against your ideal customer profile. It works well as a first pass, routing obvious non-fits to nurture sequences and flagging high-potential leads for immediate follow-up. But scoring models need continuous calibration against actual conversion data, and the enrichment data is only as good as its sources.

Meeting scheduling and coordination. AI assistants can handle most scheduling back-and-forth, but they struggle with nuanced context: "schedule after the board meeting but give me 30 minutes to decompress" or timezone edge cases involving daylight saving transitions. Works well for straightforward scheduling, needs human fallback for anything complex.

Report generation from structured data. Given clean data, AI can produce narrative reports: financial summaries, performance reviews, status updates. The output is coherent and saves significant drafting time. But the analysis layer is shallow. AI will describe trends accurately but may miss the "why" behind a metric change. Use it for the first draft, not the final analysis.

Knowledge base Q&A (RAG). Retrieval-augmented generation works well when your documents are clean, current, and well-organized. It struggles when documents contradict each other, when the answer requires synthesizing information from many sources, or when the question is ambiguous. The difference between a good RAG system and a bad one is almost entirely about document quality and retrieval tuning, not the model itself.

Not ready yet

These are areas where AI either fails too often, requires too much supervision, or creates risks that outweigh the benefits.

Autonomous decision-making with financial impact. AI should inform decisions, not make them unilaterally. Approving refunds, adjusting pricing, committing budgets, or signing off on purchases. These need human judgment. The model can prepare the analysis and recommendation, but the approval must be human. This is not a technology limitation; it is a trust and accountability limitation.

Complex multi-step reasoning chains. Tasks that require maintaining context across many steps, such as long legal analyses, multi-variable financial modeling, or debugging intricate system failures, still exceed what AI handles reliably. Models can assist (breaking down the problem, suggesting hypotheses) but cannot yet drive these processes end-to-end with the reliability you need.

Creative strategy and judgment calls. AI can generate options, draft variations, and brainstorm. It cannot replace the judgment of someone who understands your market, your customers, and your competitive position. "Write me a marketing strategy" produces generic output. "Draft three positioning options for this specific audience segment given these competitive dynamics" produces useful raw material, but the choice between them requires human judgment.

Anything requiring real-time physical-world awareness. Scheduling that depends on physical presence, logistics that depend on weather or traffic conditions, or any workflow where the input data changes faster than the system can process it. APIs and integrations can bridge some of this, but the reliability is not there for production use without significant engineering.

Processes where errors carry outsized consequences. Medical diagnosis, legal advice, safety-critical engineering calculations. AI can assist the professional (surfacing research, checking for common errors, flagging anomalies) but should not be the primary decision-maker. The cost of a wrong answer is too high relative to the time saved.

How to use this map

When evaluating an automation opportunity, place it on this map honestly. If it falls in the "works reliably" category, you can move fast with confidence. If it is in the "works with guardrails" category, plan for human-in-the-loop and budget for the oversight. If it is in the "not ready" category, either wait or reframe the project as "AI-assisted" rather than "AI-automated."

The biggest waste of money in AI is building for full autonomy in a domain that requires guardrails. The second biggest is adding guardrails to a domain that works reliably. You are paying for oversight you do not need.

Before committing to any automation, run the ROI math. Our free calculator factors in oversight overhead, which is exactly what determines whether a "works with guardrails" workflow is actually worth automating.

For a concrete look at five workflows that fall in the "works reliably" category, with ROI numbers and implementation guidance, start there.

If you want someone to map your workflows against this framework and then build the automation, that is a fixed-price project, from €5,000. We start by checking whether it pays, and we tell you if it does not.

Want this in your business?

Fixed-scope, fixed-price AI projects, from €5,000. We work out first whether it pays, then make it work with real users.