Automation vs AI: what your small business actually needs.

Start with the work taking too much time. Write down what happens, who handles it, and where it gets stuck. That usually makes the next move clearer: a simple automation, help from AI, or a better handoff to the next person.

What the failure figures actually say

You've probably seen some big numbers about AI projects failing. They're worth a closer look. A project that's been abandoned isn't the same as one that hasn't yet earned back its cost, and a forecast isn't a measured result. The chart below keeps those differences visible, with the sources beside it.

AI projects: four different measures

Reported shares, estimates and a forecast

Organizations with no measurable GenAI return

95%

MIT NANDA · 2025 · preliminary report

AI projects estimated to fail

>80%

Estimate cited by RAND · 2024

Companies abandoning most AI projects

42%

S&P Global · 2025 survey

GenAI projects expected to stop after a pilot

≥30%

Gartner · 2024 forecast for the end of 2025

For your own project, the reasons matter more than the headline percentages. RAND found that teams can misunderstand the problem they are trying to solve. Gartner named data quality and unclear business value as risks. Before choosing a tool, agree on its job and how you will tell whether it helps.

BCG's 10-20-70 framework puts most of the emphasis on people and process: 70 percent, alongside 20 percent for data and technology and 10 percent for algorithms. It is not a budget formula. It does make one point clearly: leave time for the people using the system and the work that needs to change.

Three roles in a working system

Automation, AI, and human review do different jobs. Give each a clear boundary and the system is easier to trust.

01 / Rules

Automation

For work with a clear rule.

Someone fills out a form. A job record is created and a confirmation goes out. That's automation: rules you set, run the same way each time. You can follow what happened and see where something went wrong. Google's engineering guide makes a similar case for starting with simple rules before machine learning.

Best forIntake, routing, notifications, reminders, and moving data between tools. Any step where you can write the rule in one sentence.

Watch forToo many exceptions. If every request needs a different rule, the automation gets harder to maintain.

02 / Judgment

AI

For work where the input varies.

AI can help summarize a long thread, draft from approved material, or sort requests that arrive in different formats. Give it a specific job and a way for someone to catch mistakes. McKinsey's review of teams using AI agents also notes that many problems are still better handled with ordinary automation.

Best forSummarizing before a handoff, first drafts from your own material, classifying inbound requests, pulling background before a call.

Watch forConfident mistakes. AI can invent a number, misread a request, or promise something your business can't deliver.

03 / Control

The human check

For decisions that need approval.

When a mistake would affect a quote, booking, payment, or customer, build in a chance to check it. The system can prepare the next step; a person reviews it and approves or changes it before anything goes out.

Best forQuotes, bookings, payments, and anything sent to a customer with your name on it.

Watch forRushed approvals. Give the person enough context and time to check, plus a clear way to make a correction.

Start by writing the rule. If you can describe it clearly, automate it. If each case needs judgment, AI may help. If the step is still hard to describe, map it before adding a tool.

Map the workflow before you pick the tool

Start with the process. Map one workflow from beginning to end, then look for the steps that repeat, stall, or carry real risk.

• See the real process, including the workarounds people use.
• Mark steps that happen the same way each time. Those are automation candidates.
• Find where work waits or gets typed twice. A routing rule can often fix that.
• Mark steps where an error costs money or lets a customer down. Decide what needs checking before it goes ahead.

That split shows up in our own work. Strata Match uses deterministic intake, matching rules, and notifications because a marketplace needs consistent handling. In Evalia, the scoring engine calculates each number deterministically; AI can explain the result but does not calculate it. Recon uses AI for drafting and research, with a person approving actions before they happen.

Voice agents: where the split shows up

Voice AI makes the boundary between useful and risky very clear. Realtime APIs can handle a short inbound conversation, capture a structured message, ask a few qualifying questions, or check a calendar. That suits bounded work, such as after-hours intake. ReplyFirst is built around that use case for home-service trades: capture the call and route an urgent lead to the right person.

Names, addresses, and phone numbers are easy to mishear. A long request can lose its shape, and an agent may agree to something it should have escalated. Build for those limits. Let the agent collect and route the request, send consequential actions for confirmation, and give callers a clear route to a person.

There is a legal side too, and it depends on the call, the data collected, and where you operate. Start with the Office of the Privacy Commissioner's guidance on call recording and the CRTC's rules for automated solicitation. If you are recording calls, making automated decisions, or placing outbound calls, get advice for your use case. Our practice is simple: tell callers they are speaking with an automated system and explain what will happen to their information.

Start with one workflow

Pick the workflow causing the most extra work or delays and write down its steps. Try one repeatable task, check whether the change helps, then decide what deserves to happen next.

The AI Readiness Check is a practical starting point. It walks through six areas of your business and points to the work worth doing first.

Common questions

Are automation and AI the same thing?

No. Automation follows defined rules, so the same input produces the same result. AI produces likely answers from patterns in its input. A workflow can use both: rules for a form submission or reminder, and AI to sort an unusual request or prepare a first draft.

When should a small business use automation or AI?

Use automation for repeatable steps with clear rules, such as routing an enquiry or sending a reminder. Use AI where the input varies too much for a practical rule, such as summarizing a long request or drafting from approved material. Keep a person in the loop for quotes, bookings, payments, and other decisions where an error matters.

What is the 30% rule in AI?

It refers to Gartner's 2024 forecast that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025. Gartner cited data quality, rising costs, and unclear business value. It was a forecast, not a count of projects that had already been abandoned.

How much does business AI cost?

Subscription prices vary widely, but the licence is only part of the budget. BCG's 10-20-70 framework puts most of the emphasis on people and process: mapping the work, preparing data, and changing how a team works. Plan for that work before buying a tool.

Explore how we build workflow automation around clear rules and human checks.