Generative AI solutions in 2026: what to know before you start
What generative AI solutions are, how they're built on platforms like Amazon Bedrock, where they work, where they stall, and how to start small.

In February 2024, Klarna said its new AI assistant was doing the work of 700 full-time customer service agents. About a year later, its CEO was hiring humans again. That arc tells you most of what you need to know about generative AI solutions: they're software products built on large AI models that write, summarise, answer questions or create images from your own data, and they can be very useful, but only when the job is narrow and someone checks the output.
If you're a business owner or manager weighing one up, my view is clear. Start small, buy before you build, and measure one number you already track. The rest of this guide explains why.
The background: models, prompts and your company's data
Generative AI means software that produces new content (text, images, audio, code) instead of only sorting or predicting from existing data. Most of today's tools trace back to the transformer, a model design Google researchers published in a 2017 paper called "Attention Is All You Need". Train a transformer on a huge pile of text and it learns to predict the next word very well. Scale that up and you get the large language models behind ChatGPT, Claude and Gemini.
A model on its own isn't a solution, though.
The word "solution" usually means three layers stacked together. At the bottom sits a foundation model that someone else trained at enormous cost. In the middle is your data, fed in through a technique called retrieval-augmented generation (RAG), where the system searches your documents first and hands the relevant passages to the model along with the question. On top is the workflow: the chat window, the approval step, the link into your CRM or helpdesk. In my view, that top layer is where most projects succeed or fail.
An everyday comparison helps here, and it's the only one I'll use. A foundation model is like a very well-read new hire on day one; RAG is the company handbook you put on their desk; the workflow is the manager who decides what they're allowed to send to customers without a second look.
The big platforms behind most gen AI solutions
Most businesses don't touch a model directly. They rent one through a cloud platform. Amazon's gen AI offering centres on Amazon Bedrock, which AWS says is used by more than 100,000 organisations and gives access to hundreds of models from different companies. It bundles Knowledge Bases for RAG, Guardrails for filtering harmful content, and AgentCore for building agents that take actions. AWS also states that Bedrock doesn't store or use your data to train models, which is the first thing most legal teams ask about.
Google's equivalent used to be called Vertex AI. It's now branded the Gemini platform, offering Google's Gemini 4 Argon alongside Anthropic's Claude models and open models like Gemma. Microsoft's is Microsoft Foundry, which wraps models, an agent service and Foundry IQ (the successor to Azure AI Search) into one place.
They're more alike than their marketing suggests. I'd usually pick whichever cloud your company already pays for, because data access and security reviews are the slow part, not the model choice.
Generative AI solutions in real companies
Klarna's assistant is still the cleanest public example, partly because the company published hard numbers. In its first month, the OpenAI-powered assistant handled 2.3 million conversations, two-thirds of Klarna's customer service chats. Average resolution time fell from 11 minutes to under 2, repeat inquiries dropped 25%, and Klarna estimated a $40 million profit improvement for 2024. It ran in 23 markets and more than 35 languages.
Then the story got messier. In 2025, CEO Sebastian Siemiatkowski told Bloomberg that the company's AI-heavy approach to support hadn't been the right one and that the "quality of human support" was now the focus. Klarna started recruiting remote human agents again. I don't read that as failure. I read it as the normal second chapter: the bot took the easy, repetitive chats, and the company found out the hard ones still needed people.
Other common uses look less dramatic and probably deliver more steady value:
- Internal search over policies, contracts and wikis, with answers that cite the source document.
- First drafts of sales emails, product descriptions and support replies that a person edits before sending.
- Summaries of calls, tickets and long reports for managers who don't have time to read them.
- Code suggestions and test generation for developers.
- Pulling fields out of invoices, forms and PDFs into structured data.
For a longer list by industry, see our guide to generative AI examples.
Benefits and limitations
The benefits are real but uneven. Deloitte's 2026 State of AI in the Enterprise survey of more than 3,000 business leaders found that companies had raised the share of workers with approved AI tools from under 40% to around 60% in a single year. Yet only 34% said they were using AI to "deeply transform" the business. Most of the gain shows up as individual productivity: faster drafts, quicker answers, fewer minutes spent hunting through folders.
Getting from pilot to production is where things stall. That same Deloitte survey found only 25% of respondents had moved 40% or more of their AI pilots into production. MIT's NANDA initiative was blunter in its July 2025 report, "The GenAI Divide": despite $30–40 billion of enterprise spending, it said 95% of organisations were seeing no measurable return, and just 5% of custom enterprise AI tools reached production. The authors blamed brittle workflows and systems that "do not retain feedback, adapt to context, or improve over time."
Those figures come from surveys with their own definitions of "return", so I wouldn't treat 95% as a precise failure rate. Still, two separate studies pointing the same way is hard to ignore.
The limitations are well known by now:
- Models still make things up (hallucinate), especially when asked about facts that aren't in the documents you've connected.
- Costs scale with usage, so a tool that's cheap in a 20-person pilot can get expensive at 2,000 users.
- Your data has to be clean and permissioned, or the assistant will happily surface the salary spreadsheet to the whole company.
- New attack types such as prompt injection let a malicious email or web page give instructions to your agent.
That last risk is why I'd put staff training on the list before launch; our roundup of AI security training options covers the practical side. On the vendor end, AWS claims Bedrock Guardrails can block up to 88% of harmful content. That's a useful filter, but "up to 88%" also tells you some gets through.
Practical takeaways before you spend money
There are really four ways to get a generative AI solution, and they differ a lot in cost, speed and risk. The table below is my own rough map, not a vendor's.
| Approach | Typical example | Time to first use | Who you need | Main risk |
|---|---|---|---|---|
| Use an off-the-shelf assistant | ChatGPT, Claude or Gemini business plans | Days | An admin and a usage policy | Staff paste sensitive data into it |
| Turn on AI inside tools you own | AI features in your CRM, helpdesk or docs app | Weeks | The tool's admin | Paying for seats nobody uses |
| Configure a platform with your data | Bedrock Knowledge Bases, Gemini platform, Microsoft Foundry | 1–3 months | One or two developers | Messy data gives confident wrong answers |
| Build or fine-tune your own | Custom model, agents wired into core systems | 6 months or more | An ML team plus a budget for compute | The project never reaches production |
Most small and mid-sized companies, I think, should stop at row two or three. Here's what that means in practice.
Pick one workflow with a number attached. "Reduce average ticket handling time" works; "use AI more" doesn't. Klarna's results were persuasive precisely because they tracked resolution time and repeat contacts before and after.
Buy first. A turnkey tool tells you within weeks whether the use case is worth anything, and if it isn't, you've lost a subscription fee instead of a development budget. Our comparison of turnkey AI systems versus building your own goes deeper on this choice.
Keep a human on anything customer-facing or legally binding. Plan for that review step to cost real hours, and include them in your savings maths.
Write the rules down. NIST's Generative AI Profile (NIST AI 600-1) is free and gives a usable checklist of risks to consider, from confabulated answers to data privacy. You don't need to adopt all of it. Skim it, pick the items that apply, and you'll have a better policy than most firms your size.
The next 12 months for gen AI solutions
Two shifts look likely. The first is agents. Every big platform is now selling systems that don't just answer but act: AWS made a config-based AgentCore agent builder generally available at its New York summit and launched a managed Knowledge Base for building enterprise RAG pipelines. Amazon also stopped offering Amazon Q Business to new customers, steering them to a newer agent product called Quick. If you want the background, our explainer on agentic AI covers how agents differ from chatbots. My guess is that agents will repeat the Klarna pattern: impressive on routine tasks, then pulled back where errors are costly.
The second shift is regulation, at least if you sell into Europe. Under the EU's AI Act, rules for general-purpose AI models took effect in August 2025 and transparency rules followed in August 2026. Providers of generative AI must make AI-generated content identifiable, and deepfakes must be clearly labelled. Stricter obligations for high-risk uses such as CV screening and credit scoring start on 2 December 2027. Our overview of AI regulation covers other regions.
What I'm less sure about is price. Model costs have been falling, but agents use far more tokens per task than a single chat reply. Whether the total bill goes up or down for a typical company seems genuinely open.
FAQ
What does generative AI mean?
Generative AI is software that creates new content, such as text, images, audio, video or code, in response to a prompt. It learns patterns from very large training datasets and produces output that resembles them. ChatGPT, Claude, Gemini and image tools like Midjourney are all generative AI. Older AI mostly classified or predicted; generative AI makes something new. Our generative AI definition page goes further.
How do you build generative AI?
Most companies don't build a model from scratch; that costs millions in computing power. Instead, they pick an existing model through a platform like Amazon Bedrock, Gemini platform or Microsoft Foundry, connect their own documents with retrieval-augmented generation, add guardrails, and wrap it in a basic app or workflow. Fine-tuning a model on your own examples is an optional later step.
What is Amazon's gen AI offering?
Amazon's main gen AI product for businesses is Amazon Bedrock, a managed service on AWS that gives access to hundreds of models from multiple companies. It includes Knowledge Bases for connecting your data, Guardrails for filtering output, and AgentCore for building agents. Amazon Quick is its newer workplace assistant, replacing Amazon Q Business for new customers. Our AWS AI review has more detail.
Are generative AI solutions worth it for a small business?
Usually, yes, if you start with an off-the-shelf assistant and one clear task. Custom builds are harder to justify. I'd spend a month measuring a cheap tool before paying anyone to build something bespoke.



