👋 Welcome back to AI for SME Success, your weekly dose of practical AI insights for small businesses.
Today we cover:
- AI caught $45,000 in invoice errors over three years
- What every small business should know about AI-washing
- Vibe-coding in Canva + Google Search and NotebookLM updates
- How people use Claude at work
- 5 tips for greener AI use
Let’s dive in! 👇
💰 AI Found $45,000 in Invoices Over Three Years
Invoice review is where AI shines.
HashiCorp co-founder Mitchell Hashimoto ran a ChatGPT automation over three years of his construction billing, and it found about $45,000 in errors he had already paid.
His lifetime ChatGPT spend was roughly $1,800, so this one job paid for the tool about 25 times over. He shared the full breakdown publicly.
Hashimoto provided AI with read-only access to previous invoices, emails, and meeting notes, and prompted the tool to create a report and email it to him. AI found wrong amounts, duplicate invoices, and bills addressed to wrong people. Hashimoto manually verified the results.
Now each new bill, even low-quality scanned PDFs, gets read, checked against history, and flagged for discrepancies.
Takeaway: invoice review is an excellent low-risk, profitable use case for AI, where the tool uses a recurring, tested workflow and has access to rich non-confidential data.
🧼 Risks of AI Washing
AI washing is overstating, or faking AI capabilities inside products or services. Regulators treat it as deceptive advertising.
The FTC just settled with Cox Media Group and two marketing partners for $930,000 over an “AI-powered” service called Active Listening. They pitched small businesses on a proprietary algorithm that supposedly listened through smart-device microphones and served ads based on overheard conversations. It did none of that. The FTC found no voice data at all, just email lists bought from data brokers and resold at a markup.
AI washing can hurt small businesses in two ways:
- As a buyer, you risk overpaying for AI features that are overstated, underdeveloped, or unnecessary.
- As a seller, making an inflated “AI-powered” claim in your own marketing can put the liability on you.
Takeaways:
- Before you buy, ask a vendor precisely what the AI does and for evidence it works.
- When you market your product as “AI-powered,” provide verifiable evidence to support the claim.
🤖 Google and Canva Tool Updates
Canva vibe-coding
Canva opened vibe coding to everyone, free tier included. Canva Code 2.0 builds websites, landing pages, and small interactive apps from a plain-language prompt, then lets you edit the result like any Canva design, swapping brand colours, fonts, and images with tools you already use.
Google search & NotebookLM
Google is turning Search into an assistant that acts. Google’s AI Mode can now link and use connected apps like Instacart, Canva, and YouTube, so Search completes tasks instead of only returning links. Now you can draft a restock list for your restaurant and it can load those items into your Instacart cart, or ask Canva to pull flyer templates for a weekend promo without leaving the search box.
NotebookLM is now Gemini Notebook. Mostly a name change, same research tool, now inside the Gemini family, with one real addition: notebooks can run code for data analysis, rolling out to Pro users over the coming weeks.
💼 How People Use Claude at Work
Two new data releases show where people use Claude at work
The June 26, 2026 Economic Index report classified Claude conversations. Looking only at work conversations, the biggest categories are:
- business documents and communication (27%)
- learning and guidance (9%)
- summaries and data analysis (6%)
These account for 42% of work-related tasks. The rest spreads across 25+ smaller types (marketing content, blog posts, database queries, scripts, presentations).
Interestingly, a median chat conversation producing a blog post takes 13 rounds of back and forth.
A separate July 7, 2026 analysis of 1.2 million Claude Cowork sessions across more than 600,000 organizations found the biggest usage categories were:
- business operations tasks (33.4%)
- content creation and copywriting (16.4%)
- data analysis (5.8%)
- sales operations (4%)
Business operations include pulling updates into reports, building checklists, and reconciling spreadsheets. Anthropic calls this “the work around the work”: the connective tasks every role carries but nobody was hired to do.
How Anthropic stands out
Anthropic’s is built for business: roughly 80% of its revenue comes from more than 300,000 business customers, while its closest rivals still earn most of their revenue from consumers.
Anthropic has also assembled an unusually high-profile and stable team. Nobel laureate Ben joined its independent oversight trust on July 9, 2026. Anthropic has recently recruited top talent from OpenAI, DeepMind, and Microsoft. SignalFire’s 2025 State of Talent Report found it keeps 80% of hires past two years, the highest retention of any major AI lab.
🌿 5 Tips for Greener AI Use
Ontario’s skies turned orange last week.
On July 15, Toronto briefly recorded some of the worst air quality in the world. Summers like this used to be rare. They are becoming the new normal.
AI runs on electricity and water, and globally about 57% of that electricity still came from fossil fuels in 2025 (Ember, 2026), the fuels behind climate change.
Here are some recommendations for more efficient use of AI:
- Batch your tasks. Do not open a new chat for every customer email, caption, and invoice question. Collect them and run one session per workflow. Draft a whole week of social content in one conversation, not seven.
- Match the model to the job. Using a flagship model to reformat a price list is like running your delivery van to drop off one envelope. Light models (Claude Haiku, GPT mini tiers, Gemini Flash) handle summaries, formatting, and FAQs fine, at a fraction of the cost.
- Stop re-explaining your business. Save your context and standing instructions once. Most tools now support this: Projects in Claude and ChatGPT, Gems in Gemini. Load your brand voice, your customers, your rules, and work from there.
- Cut visuals. Images and video are far heavier than text. Generating one image can use as much energy as charging a smartphone. Generate with intent.
- Edit visuals, don’t re-prompt them. Regenerating an image is one of AI’s heaviest tasks, and re-prompting to fix one detail often changes everything. Canva’s Magic Layers turns a flat image into editable layers, so you can fix the one thing without generating a new picture.
Thank you for reading today’s edition.
If this issue was valuable, pass it along to a fellow business owner. I’d love to hear your feedback at natalia@nataliabrattan.com.
See you next week!