AI SaaS in 2026: What to Act On and What to Ignore

Act on AI SaaS that fits one documented workflow and proves itself on your own data. Ignore products that promise to run a whole function on their own. The evidence says most organisations use AI, few see large returns, and the difference lies in redesigning the work, not in picking a smarter tool.

This post gives you a filter for the AI SaaS pitches landing in your inbox, plus a rule for when to build instead of buy.

McKinsey's State of AI 2025 survey of 1,993 respondents in 105 countries found 88% of organisations use AI in at least one function. Only about 6% reported significant enterprise-wide impact, and those high performers were nearly three times as likely to have fundamentally redesigned workflows (McKinsey, November 2025). It also found 23% scaling an agentic AI system somewhere and 39% experimenting.

Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising cost, unclear value or weak risk controls, and it estimates only around 130 of thousands of vendors are real agentic offerings (Gartner, June 2025).

Act on narrow, checkable claims. A tool that classifies incoming tickets, drafts replies for approval or flags unusual invoices can be tested against last month's data in a week. A good vendor will let you do that before you pay.

Look for a human approval step, an audit log and a clear price. These are signs the vendor understands that your name is on the output.

Ignore the word agent when it is a label on an old chatbot. Gartner calls this agent washing. Ignore savings percentages with no baseline. Ignore any promise that the system will replace a team by next quarter.

Also be wary of products that need broad access to your systems to function. OWASP warns that excessive functionality, permissions or autonomy lets unexpected model output cause real damage (OWASP LLM06).

Buy when the need is common and the product fits your process as it stands. Build when the process is what sets you apart, when your data cannot leave your control or when you need the AI woven into an existing system rather than sitting beside it.

A middle path often works best: buy the commodity layer, such as a model API, and build the workflow around it. We explain the wider case in why custom software still wins when the workflow is complex.

Pick one workflow, record how long it takes today, run the tool in parallel for a month and compare. Keep a person approving outputs. Write down what would make you stop. If it saves time with no rise in errors, widen it. If not, you have lost little.

If you want a second opinion on a vendor shortlist or a pilot design, talk to us.

Write six lines before any demo. The workflow the tool will touch. Today's time and error rate for that workflow. The data the tool needs and where it will be stored. The person who approves outputs. The total yearly cost including your staff time to supervise. The condition under which you would cancel.

The supervision cost is the one most buyers leave out. A tool that saves ten hours a month but needs eight hours of checking has saved two. Count both before you compare prices. If a vendor cannot help you fill in the page during the sales process, that tells you how they will behave after the contract is signed.