$2.67T
Global AI spending in 2026
Gartner forecasts worldwide AI spending of about $2.67 trillion in 2026, up 49.5% from 2025, with infrastructure the largest share.[4]
Free Guide
Everything a leadership team needs to know before betting on AI.
This guide is written for owners, executives, and information technology (IT) leads at small and mid-sized businesses. It explains the technology without the hype, shows how to set AI up safely, and is honest about the costs and the people it takes to keep AI working. Sources are listed at the bottom and linked throughout.
The Basics
AI, machine learning, deep learning, and generative AI.
These terms are nested, not competing. Every generative AI tool is machine learning, and all machine learning is AI, but most AI in your business today is not generative.
| Term | Plain-English meaning | Business examples | Main risk to manage |
|---|---|---|---|
| Artificial intelligence | Software that perceives, reasons, predicts, or decides in ways that normally need a person. | Chatbots, routing rules, recommendation engines | Unclear ownership and accountability |
| Machine learning | Statistical models trained on historical data to make predictions. | Fraud scoring, demand forecasting, lead scoring, email filtering | Biased or stale training data |
| Deep learning | Machine learning using large neural networks, good with images, audio, and language. | Document optical character recognition (OCR), call transcription, defect detection | Hard to explain individual decisions |
| Generative AI | Models that produce new content by predicting the most likely next piece of output. | Drafting email and proposals, summarizing meetings, writing code, creating images | Confident wrong answers (hallucinations), data leakage, intellectual property (IP) questions |
| Large language model (LLM) | A generative model trained on huge amounts of text to predict the next word or token. | The engine inside ChatGPT, Copilot, Claude, and Gemini | Prompt injection and over-trust |
| AI agent | An LLM connected to tools, data, and permissions so it can act, not just answer. | Booking meetings, updating customer relationship management (CRM) records, triaging tickets | Excessive permissions and unintended actions |
Under the Hood
How a generative AI request actually flows.
An LLM does not look facts up. It predicts likely output based on patterns learned in training, plus whatever context you give it at the moment you ask. A well-built business deployment wraps the model in controls on both sides.
Training vs. inference
Training is when a model learns patterns from a very large dataset. It is expensive and done by the model provider. Inference is each time you use the model. You pay for inference through per-seat licenses or per-token usage fees.
Why AI "hallucinates"
Because the model predicts plausible text rather than retrieving verified facts, it can produce confident, fluent, and wrong answers, including invented citations. Grounding answers in your documents, asking for sources, and keeping a human reviewer on anything consequential are the standard mitigations.[2]
Tokens and context windows
Models read and write in tokens, roughly three-quarters of a word each. The context window is how much text the model can consider at once. Bigger windows allow longer documents but cost more per request.
Configure & Deploy
Ways to bring AI into your business, from simplest to most custom.
Most small businesses should start at the top of this table and only move down when a clear use case justifies the extra cost and responsibility.
| Option | What it is | Data control | Effort & cost |
|---|---|---|---|
| Free consumer chat apps | Personal accounts on public AI chat sites | Lowest. Prompts may be retained or used to improve models depending on the vendor's terms and settings. | Free, but the hidden cost is data exposure. Not for company data. |
| Business or enterprise AI plans | Team and enterprise tiers of ChatGPT, Claude, Gemini, and similar | Better. Admin console, single sign-on (SSO), and contractual commitments not to train on your data (verify in the contract). | Per-user monthly license. Fast to roll out. |
| AI built into tools you already use | Microsoft 365 Copilot, Google Workspace with Gemini, CRM and helpdesk AI features | Inherits your existing tenant permissions, which means over-shared files become over-shared answers. | Add-on license per user, plus permission cleanup first. |
| Application programming interface (API) + retrieval (RAG) | Your own app or chatbot calling a model API and searching your approved documents | High, if built well. You choose what data is retrieved and logged. | Development, hosting, and pay-per-use token costs. |
| Fine-tuned models | A base model further trained on your examples to match your tone or task | High, but training data must be cleaned and protected. | Higher. Needs machine learning (ML) skills and an evaluation process. |
| Self-hosted open-weight models | Running models such as Llama, Mistral, or Gemma on your own servers or private cloud | Highest. Data never leaves your environment. | Highest. Graphics processing unit (GPU) hardware or cloud spend, patching, and monitoring are on you. |
| AI agents | Models that can act through connected tools, email, files, or APIs | Depends entirely on the permissions you grant. | Moderate to high. Requires strict least privilege and approval steps. |
A safe-configuration checklist for any AI tool
- Use business or enterprise plans with single sign-on (SSO) and multi-factor authentication (MFA), never personal accounts for company work.
- Confirm in writing that your prompts and files are not used to train the vendor's models, and set data-retention to the shortest period that meets your needs.
- Clean up file-sharing permissions before turning on copilots that search your documents.
- Turn on data loss prevention (DLP) and sensitivity labels so regulated data such as patient health information (PHI), card data, controlled unclassified information (CUI), or client secrets, cannot be pasted or retrieved.
- Grant connectors and agents the minimum permissions they need, and require human approval for sending, paying, deleting, or changing records.
- Enable audit logging, keep logs for your retention period, and review them on a schedule.
- Publish a one-page AI acceptable-use policy and train employees on it before rollout.
- Map controls to a recognized framework such as the National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF)[1] or International Organization for Standardization and International Electrotechnical Commission (ISO/IEC) 42001[11] so customers and insurers can review them.
Prompt Engineering
Better instructions, better answers.
Prompt engineering is the skill of giving an AI model clear context, a specific task, and a defined output. It is the cheapest way to improve AI results, and it requires no code.
The six parts of a strong prompt
- Role: who the AI should act as ("You are a healthcare privacy compliance analyst").
- Context: the background it cannot know (audience, company size, constraints).
- Task: one clear verb ("summarize," "compare," "draft," "critique").
- Format: the shape of the answer (table, five bullets, 150 words).
- Examples: a sample of good output, if you have one.
- Guardrails: what to avoid, and permission to say "I don't know."
Techniques that work
- Ask it to think step by step, or to list assumptions before answering.
- Provide the source documents and tell it to answer only from them.
- Ask for citations and spot-check every one of them.
- Break big jobs into a chain of smaller prompts.
- Ask the model to critique its own draft against your criteria.
- Save prompts that work as shared templates for your team.
Weak promptWrite something about our security for a customer.
Strong promptYou are a security program manager at a 60-person accounting firm.
Context: a new client asked how we protect their tax documents. Use only the attached policy summary.
Task: draft a reply email in plain English.
Format: under 200 words, three short paragraphs, no jargon.
Guardrails: do not promise certifications we do not hold; if the policy does not cover something, say we will follow up.
Security note: never paste passwords, API keys, client records, or regulated data into AI tools that are not approved for that data. Be aware that text hidden in documents, emails, or web pages can carry instructions that hijack an AI assistant. This is called prompt injection and it is the number one risk on the Open Worldwide Application Security Project (OWASP) Top 10 for LLM Applications.[12]
Privacy & Regulations
The AI rules that matter to U.S. small and mid-sized businesses.
There is no single U.S. AI law yet. Existing privacy, consumer-protection, and sector rules already apply to AI, and new AI-specific laws are arriving at the state and international level.
| Rule or framework | What it means for you | Status (as of Oct. 2026) |
|---|---|---|
| Existing U.S. law | Health Insurance Portability and Accountability Act (HIPAA), the Gramm-Leach-Bliley Act (GLBA) and Federal Trade Commission (FTC) Safeguards Rule, state privacy laws, and FTC rules on unfair or deceptive practices all apply when AI handles personal data or makes claims to customers. | In force today |
| Federal AI policy | Executive Order 14365 (December 2025) set up a federal push to preempt state AI laws, and the White House released a National Policy Framework for AI in March 2026. | State AI laws remain in effect unless Congress acts or courts rule otherwise[10] |
| Colorado Senate Bill (SB) 26-189 | Replaced the original Colorado AI Act. Requires notice before using automated decision technology for consequential decisions, plain-language explanations after adverse outcomes, and a path to human review. | Effective January 1, 2027[9] |
| Texas Responsible Artificial Intelligence Governance Act (TRAIGA), House Bill (HB) 149 | Prohibits certain harmful AI uses and sets disclosure duties, mainly for government agencies and health care. | In force since January 1, 2026[13] |
| EU AI Act | Applies if you sell to or serve people in the EU. The 2026 Digital Omnibus (Regulation (EU) 2026/1744) moved high-risk Annex III obligations to December 2, 2027 and Annex I to August 2, 2028. | Prohibitions and general-purpose AI rules already apply; transparency labeling for some systems from December 2, 2026[8] |
| NIST AI RMF and generative AI (GenAI) Profile | Voluntary U.S. framework (Govern, Map, Measure, Manage) that regulators, customers, and insurers increasingly ask about. | Published; widely referenced[1][2] |
| ISO/IEC 42001 | Certifiable international standard for an AI management system. | Published 2023; certification available[11] |
This is general information, not legal advice. AI law is changing quickly; confirm current requirements with your counsel. See Regulatory Watch for cybersecurity rule updates.
Privacy practices that keep AI out of trouble
- Know which AI tools touch personal data, and list them in your records of processing and vendor inventory.
- Minimize: send the AI only the data the task needs, and strip identifiers where possible.
- Tell people when they are interacting with AI or when AI helps make a decision about them.
- Keep a human in the loop for hiring, lending, insurance, health, and other consequential decisions.
- Update your privacy notice and customer contracts to describe AI use and subprocessors honestly.
The Big Picture
How large organizations are using AI, and what it is really delivering.
44%
Scaling AI enterprise-wide
In McKinsey's 2026 survey, 44% of organizations said they are scaling AI across the enterprise, up from 38% a year earlier. Nearly nine in ten use AI regularly.[3]
37%
Seeing enterprise-level profit impact
Only 37% report enterprise-level earnings before interest and taxes (EBIT) impact from AI, essentially flat year over year. Adoption is easier than results.[3]
362
Reported AI incidents in 2025
The Stanford AI Index recorded a record 362 reported AI incidents in 2025, a reminder that governance has not kept pace with deployment.[5]
Where large companies are putting AI to work
- Customer service: AI-drafted replies, chat triage, and call summaries
- Software engineering: coding assistants and agents for tests and documentation
- Marketing and sales: content drafts, personalization, and lead research
- IT and security operations: alert triage, ticket routing, and threat-intel summaries
- Knowledge management: searchable answers over policies, contracts, and manuals
- Finance and operations: invoice processing, forecasting, and anomaly detection
What separates winners from stalled pilots
- They redesign the workflow, not just add a chatbot to it.
- They pick a few measurable use cases and track results.
- They clean up data and permissions first.
- They assign an accountable owner and governance from day one.
- They train people. MIT NANDA researchers reported in 2025 that most enterprise generative AI pilots they studied showed no measurable profit return, largely because tools were not integrated into real workflows.[6]
Cost
What AI really costs to implement, update, and use.
License fees are the visible part. Integration, data cleanup, security, training, and ongoing monitoring usually cost more over the life of an AI program.
| Cost category | What drives it | One-time or ongoing |
|---|---|---|
| Licenses and seats | Number of users and tier. For example, Microsoft lists Microsoft 365 Copilot Business at $21 per user per month, with an $18 promotional first-year price through December 31, 2026.[7] | Ongoing |
| Usage and API fees | Tokens processed, model size, and how many requests your apps or agents make | Ongoing, variable |
| Data preparation and integration | Cleaning documents, fixing permissions, connecting CRM, enterprise resource planning (ERP), or file systems | Mostly one-time, plus upkeep |
| Security and governance | Policy, AI inventory, vendor reviews, DLP, logging, red-team testing | Setup, then ongoing |
| Training and change management | Role-based training, prompt libraries, champions program | Setup, then refreshers |
| Monitoring and model updates | Accuracy checks, drift, new vendor model versions, re-testing prompts and agents | Ongoing |
| Hardware (self-hosted only) | GPUs, servers or private cloud, power, and patching | Capital plus ongoing |
Illustration: 25 employees on a $21 per user per month copilot license is $6,300 a year in licenses alone, before integration, training, or governance. Pilot with the people most likely to benefit, measure, then expand.
Trust
Seven trust factors to check before you rely on AI.
NIST describes trustworthy AI with seven characteristics. Use them as questions for any AI tool or vendor.[1]
| Characteristic | Question to ask |
|---|---|
| Valid and reliable | How accurate is it on our tasks, and how was that measured? |
| Safe | What happens when it is wrong, and who catches it? |
| Secure and resilient | Has it been tested against prompt injection, data leakage, and abuse? |
| Accountable and transparent | Who owns it, and can we see what data and model it uses? |
| Explainable and interpretable | Can we explain an output to a customer, auditor, or regulator? |
| Privacy-enhanced | What data is retained, where, for how long, and is it used for training? |
| Fair, with harmful bias managed | Has it been checked for unequal outcomes across groups of people? |
Vendor trust signals
- A current System and Organization Controls (SOC) 2 Type II report, and ISO/IEC 42001 or ISO/IEC 27001 certification where available
- A clear data processing agreement with a no-training commitment and a subprocessor list
- Published model or system cards that describe limitations and testing
- Admin controls for SSO, retention, logging, and data residency
- A responsible-disclosure program and a history of fixing reported issues
Benefits & Cautions
The upside of AI and ML, and where to be careful.
Benefits
- Time back on drafting, summarizing, and searching for information
- Faster customer response and around-the-clock first-line support
- Better forecasting and anomaly detection from machine learning
- Accessibility gains: transcription, translation, and plain-language rewrites
- Faster security triage and threat-intelligence analysis
- Small teams can compete on output with much larger ones
Cautions
- Hallucinations: fluent, confident, and wrong answers, including fake citations
- Data leakage through prompts, over-shared files, or unapproved tools (shadow AI)
- Prompt injection and over-permissioned agents taking unintended actions
- Bias in training data that leads to unfair outcomes for people
- Intellectual property and copyright uncertainty in generated content
- Deepfakes and AI-powered fraud aimed at your staff (see How to Spot AI Fakes)
- Vendor lock-in, rising usage bills, and skills erosion from over-reliance
People & Resources
Who you need to deploy, update, and maintain AI.
AI is not set-and-forget. Models change, vendors ship new versions, data drifts, and regulations evolve. Large enterprises staff every role below. Small and mid-sized businesses usually combine roles and bring in fractional help.
| Role | Responsibilities | Typical SMB approach |
|---|---|---|
| Executive sponsor | Sets goals, approves budget and risk appetite, owns outcomes | Owner, chief executive officer (CEO), or chief operating officer (COO) |
| AI governance lead | AI policy, inventory, risk register, vendor approvals, regulatory tracking | Fractional virtual chief information security officer (vCISO) or fractional AI advisor |
| Security engineer | Identity, DLP, logging, prompt-injection defenses, agent permissions, red-team testing | Managed service provider (MSP) plus a fractional security lead |
| Data owner or data engineer | Data quality, permissions, pipelines, retention | Existing IT lead or MSP, with project help |
| AI or ML engineer | Builds RAG apps, integrations, fine-tuning, evaluations | Project-based contractor, only if building custom AI |
| Privacy and legal | Contracts, privacy notices, regulatory interpretation | Outside counsel as needed |
| Domain experts and reviewers | Define "good," review outputs, catch errors | Your own staff |
| AI champions and trainers | Prompt libraries, peer coaching, adoption tracking | One or two enthusiastic employees |
A maintenance cadence that works
Monthly
- Review usage, cost, and audit logs
- Triage user-reported errors
- Check for new shadow AI tools
Quarterly
- Re-test key prompts and agents after vendor model updates
- Update the AI inventory and risk register
- Review permissions and connectors
Annually
- Refresh the AI policy and training
- Re-assess vendors and contracts
- Red-team high-risk use cases and run an AI incident tabletop
How Nexus Helps
Services to implement, update, and maintain AI safely.
Nexus Security Advisors can fill the governance and security roles above on a fractional basis, so you get enterprise-grade AI oversight without a full-time hire.
01
AI readiness & use-case sprint
Inventory current AI use, rank use cases by value and risk, and build a 90-day plan.
02
AI governance program
Policy, inventory, risk register, vendor rules, and controls mapped to NIST AI RMF and ISO/IEC 42001.
03
Secure copilot & agent rollout
Permission cleanup, DLP, SSO, logging, and least-privilege design before you switch AI on.
04
AI red teaming & testing
Prompt-injection, data-leakage, and agent-abuse testing, with findings mapped to OWASP and MITRE Adversarial Threat Landscape for Artificial-Intelligence Systems (ATLAS).
05
Ongoing AI oversight
Monthly and quarterly maintenance cadence, vendor and model-update reviews, and regulatory tracking as a fractional AI advisor.
06
Training & incident readiness
Leadership workshops, staff prompt and safety training, deepfake-fraud tabletops, and AI incident response plans.
Sources
References
Sources cited on this page (last reviewed October 31, 2026)
- NIST, AI Risk Management Framework (AI RMF 1.0). nist.gov
- NIST, AI 600-1: Generative AI Profile, July 2024. doi.org/10.6028/NIST.AI.600-1
- McKinsey & Company, The state of AI in 2026: On the road to ROI, August 2026. mckinsey.com
- Gartner, Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026, September 16, 2026. gartner.com
- Stanford HAI, The 2026 AI Index Report. hai.stanford.edu
- The Register, reporting on MIT NANDA's The GenAI Divide: State of AI in Business 2025, August 18, 2025. theregister.com
- Microsoft, Microsoft 365 Copilot plans and pricing, accessed October 31, 2026. microsoft.com
- Orrick, EU AI Act Update: Digital Omnibus Finalizes 8 Compliance Changes, July 2026. orrick.com
- Seyfarth Shaw, Colorado Enacts Artificial Intelligence Replacement Law, 2026. seyfarth.com
- Ropes & Gray, The White House Legislative Recommendations: National Policy Framework for AI and Federal Preemption of State AI Laws, March 2026. ropesgray.com
- ISO, ISO/IEC 42001:2023 Artificial intelligence management system. iso.org
- OWASP GenAI Security Project, Top 10 for LLM Applications 2025. genai.owasp.org
- Digital Policy Alert, Texas Responsible Artificial Intelligence Governance Act (HB 149) enters into force. digitalpolicyalert.org
Related free guides: Responsible & Ethical AI · How to Spot AI Fakes and AI Bots · AI Governance Checklist