AI solutions built into real workflows.

• Start with a chatbot when the same questions already have written answers.

  If the answers already exist in pages or documents, a chatbot grounded in that content is the most direct first project.

• Start with document intelligence when staff re-key documents by hand.

  When documents are read and retyped into another system, extraction with validation and review removes the copying, not the checking.

• Start with a knowledge base when nothing maintained exists to answer from.

  An assistant is only as good as the content it retrieves, so scattered or outdated answers are organized first.

• Look to workflow automation when the step follows a rule, not a judgment.

  A step that can be written as a condition needs no model: a rule-based workflow, covered under integrations and automation, is cheaper and predictable.

• Look to web apps and SaaS when the AI feature is really an application.

  A feature that needs its own screens, users and permissions is an application build, covered under web apps and SaaS.

A business can use six broad types of AI solution: AI integrations that add a model to existing software, AI-powered applications built around it, AI automation inside workflows, AI agents that take steps with tools, knowledge-based AI that answers from approved content, and AI-powered internal tools for staff. One project can combine several.

• AI integration: a model added to existing software for one task.
• AI-powered application: a product in which AI is part of the experience.
• AI automation: a workflow step where a model interprets text or documents.
• AI agent: a system that takes steps toward a goal using tools.
• Knowledge-based AI: answers drawn from approved documents and records.
• AI-powered internal tools: assistants and interfaces built for staff.

An AI-powered application is software in which AI is part of the experience rather than an add-on: a support portal that drafts replies, a platform that reads uploaded documents, or a product that answers questions about its own data. It is still an application, so it needs accounts, permissions, records and screens like any other.

The application itself is built under web apps and SaaS, and the AI features inside it are designed here: what the model may see, how its output is checked and what happens when it fails. A language feature added to an existing application is an LLM integration.

AI automation is a workflow in which one or more steps use a language model to interpret something rules cannot, such as the meaning of an email or the fields in a document. The workflow around it stays rule-based: it validates the model output, applies business rules, routes uncertain cases to a person and records every run.

The example below follows one inquiry. The model only classifies the message: the triggers, approvals and logging around it are workflow automation, and the CRM update follows the rules of any CRM integration.

1. An inquiry arrives by email or through a website form.
2. A language model classifies it by service, urgency and intent, returning a structured result.
3. The result is validated, and an unclear classification goes to a person.
4. Business rules route the inquiry to an owner.
5. The contact and the inquiry are created or updated in the CRM.
6. The owner is notified, with the original message and the classification.
7. A person reviews the drafted reply or the next step before anything is sent.
8. A follow-up is scheduled and the outcome is recorded.

An AI agent is a system that uses a language model to work toward a goal in several steps, choosing which tools to use, such as searching records, calling an API or drafting a message, and checking results as it goes. It can do only what its tools and permissions allow, and consequential actions should wait for a person.

The difference from a chatbot is action: a chatbot answers, an agent does. Excessive agency usually has one or more of three root causes: excessive functionality, excessive permissions and excessive autonomy. The tools and permissions an agent has are therefore kept to a minimum, high-impact actions require approval from a user, and logging and rate limiting are treated as ways to limit the damage rather than prevent it.

Agents run on the same foundations as any automated workflow, covered under workflow automation. The steps below show one agent task, and the design work is in limiting each of them.

1. A person or a workflow sets a goal, such as preparing a quote.
2. The agent reads its instructions and the business rules that apply.
3. It gathers context from the records it is allowed to see.
4. It chooses tools, such as a CRM search, one step at a time.
5. It takes actions within its permissions, pausing for approval where required.
6. The result is returned with a log of every step.

RAG, retrieval-augmented generation, is a technique in which a system first retrieves the passages of approved content most relevant to a question, then gives them to a language model to answer from. A knowledge base is the maintained source of that content, so the quality of its documents and of the retrieval limits the quality of the answers.

Retrieval-augmented generation combines two kinds of memory: the parametric memory a model gains in pre-training and non-parametric memory, a searchable index of documents outside the model, used together for language generation.

Protections for a retrieval system include fine-grained access controls with permission-aware vector and embedding stores, and validation of the knowledge sources it draws on. A knowledge base used by both staff and customers therefore retrieves only what the person asking is allowed to see.

1. Business knowledge is gathered from documents, FAQs and records.
2. It is split into passages and indexed so it can be searched by meaning.
3. A question retrieves the relevant passages the person may see.
4. Those passages are given to the model as context.
5. The response cites the passages it used, or says the content has no answer.
6. Gaps found in conversations go back to the content owners.

Intelligent document processing reads documents a business already handles, such as forms, invoices and reports, and turns them into structured data. Each document is read, using optical character recognition (OCR) where it is a scan, then classified; its fields are extracted and validated against rules, and anything uncertain goes to a person before the data reaches another system.

Extraction is never perfect, so the design plans for errors. Each field is checked against rules, such as a total matching its line items or a date falling in a valid range, and fields that fail, or that are marked as uncertain, go to a review queue instead of the next system.

Document types are decided for each project from real samples, and long reports can also be summarized for the reviewer.

1. A document arrives by upload, email or a shared folder.
2. Its text is read, with OCR where it is a scan.
3. It is classified by type, such as an invoice.
4. The fields that type requires are extracted.
5. Each field is validated against rules.
6. Uncertain or failed fields go to a person.
7. Approved data is sent to the system that uses it.

AI assists a support team by classifying incoming tickets by topic and urgency, suggesting replies drawn from approved answers, summarizing long conversations, routing each case to the right person and answering routine questions through a chatbot. People stay responsible for what is sent, for complaints and refunds, and for anything the approved answers do not cover.

Wherever possible, a person reviews model outputs before they are used, especially in high-stakes domains, so suggested replies start as drafts that the support team edits and sends.

Deflection, the share of questions resolved without a person, is measured after launch rather than promised. The public chatbot itself is covered under AI chatbot integration.

AI-powered search inside a website or app finds content by meaning rather than by exact keywords, so a visitor who types a question in their own words still finds the right page, product or document. It works by indexing content as numerical representations of meaning, often combined with keyword search, and ranking results against each query.

Semantic search does not understand everything. Results depend on the content indexed, how it is split and labeled, and how relevance is tuned against real queries. Permissions apply at retrieval, so a search across internal documents never returns a file the person could not open.

This is search on your own website or application. Being found and cited by AI-powered search engines is a different discipline, covered under SEO, AEO and GEO.

AI-powered internal tools are interfaces built for staff rather than customers: an assistant that searches internal documents, a tool that summarizes reports or meetings, a reporting helper that answers questions about business data, or a workflow screen that drafts the next step. Each solves a specific internal task and respects the access every employee already has.

Internal tools carry a quieter risk than public ones: people trust them. A model can produce misinformation, meaning false or misleading information that appears credible, and people can fall into overreliance, placing excessive trust in that output without first verifying it, so internal tools show their sources, and decision-support tools leave the decision with a person.

For content work, analytics then show which drafted pages help visitors.

An AI feature is a stack of layers, and the model is only one of them.

• Answers are grounded in the approved content of the business through retrieval, rather than in what a model absorbed in training.
• The data-use settings of each model provider are checked and recorded before anything is connected.
• API keys, prompts and limits stay on the server, never in browser code.
• A set of real questions or documents with known correct results is used to evaluate the system before launch and after every change.
• Every answer and action is logged so it can be reviewed, with personal data kept out of logs that do not need it.
• No accuracy, savings or ranking result is guaranteed: accuracy and cost are measured on each project.