Business & founders · The Operator
Put AI to work in your business, without the hype.
Your first three moves
- Rewrite product pages from your own specsGive the model your spec sheet, return policy and three real customer questions per product, and ask for a description, five bullets and a short FAQ. A person approves every page. Check claims about materials, sizing and safety: a false claim is your liability whether a person or a model wrote it.
- Draft replies for your 20 most common support questionsExport a month of tickets, group the repeats (where's my order, returns, sizing), write approved answers, then let AI draft replies from that library plus live order data. Run it as drafts that an agent reviews and sends for the first month, and measure handle time before and after.
- Summarize real reviews; never generate fake onesA "What customers say" summary built from your actual reviews, negatives included, helps shoppers. AI-written reviews attributed to people who don't exist are banned under the FTC's fake-reviews rule, in effect since October 21, 2024.
ROI calculator
Be conservative with the share AI can handle. Field studies show real gains vary a lot by task and by person.
Business news
Playbooks by team
Customer support
The best-studied use of AI at work. Start with drafted replies from approved answers, measure, then automate the simplest questions.
Low effortFirst steps
- Export 30 days of tickets and list the 20 most frequent questions.
- Write or approve one canonical answer for each, and load them as the AI's knowledge source.
- Run AI-drafted replies that agents review and send; track handle time, reopen rate and CSAT against the previous month.
- After four clean weeks, let the AI answer only the top three question types on its own, with a visible "talk to a person" option.
Watch out for
- Never let it promise refunds, credits or policy exceptions; route those to a person.
- Watch for confident wrong answers on rare questions, which is where models fail most.
- Tell customers when they're talking to an automated assistant; in the EU it's a legal requirement from August 2, 2026.
Evidence: In a field study of 5,172 support agents, AI suggestions raised issues resolved per hour by 15% on average. Less experienced agents improved both speed and quality; the most experienced saw small speed gains and small quality declines. Brynjolfsson, Li and Raymond, Generative AI at Work
Sales
Use AI for research and prep, where errors are cheap, before letting it near outbound messages.
Low effortFirst steps
- Before each call, have AI summarize the account's website, recent news and your CRM notes into a one-page brief.
- Generate call follow-up emails from your notes, with a rep approving each one.
- Have AI clean and enrich CRM records in batches, with a sample checked by hand.
- Draft personalized first lines for outreach, capped at a volume reps can review.
Watch out for
- Mass AI-personalized email can breach anti-spam rules and burn your domain's reputation; keep volumes and opt-outs honest.
- Check every fact about a prospect before using it; invented details kill trust instantly.
- Recording and transcribing calls needs consent where state law requires all-party consent.
Marketing
Drafting and repurposing content is a proven win. Brand voice, claims and disclosure need guardrails.
Low effortFirst steps
- Write a one-page brand voice guide with five real examples, and include it in every prompt.
- Repurpose one strong asset (a webinar, a report) into posts, emails and a summary page.
- Set a claims-check rule: every statistic, comparison and product claim needs a source before publishing.
- Measure time per asset and performance against human-only assets for a quarter.
Watch out for
- Never publish AI-generated reviews or testimonials; the FTC's rule bans them.
- Substantiate any "AI-powered" claim about your own product; the FTC says there is no AI exemption from the laws on the books.
- Check image rights and likeness: don't generate images of real people without permission.
Evidence: In a controlled experiment with college-educated professionals doing mid-level writing tasks, ChatGPT access cut time taken by 40% and raised output quality, as rated by independent evaluators, by 18%. Noy and Zhang, Experimental evidence on the productivity effects of generative AI, Science (2023)
Finance
Good at explaining, drafting and spotting anomalies. Keep it away from final numbers without a check.
Medium effortFirst steps
- Use AI to draft variance commentary from the actual reports, with an analyst verifying every figure.
- Have it explain unfamiliar contract terms in vendor agreements and flag renewal dates.
- Ask it to write and explain spreadsheet formulas, tested on a few rows before use.
- Add a callback-verification rule for any payment change request, whatever channel it arrives through.
Watch out for
- Models make arithmetic slips; recompute anything that goes into a report or a payment.
- Deepfake payment fraud is real: Arup lost about US$25 million after a video call with a fake CFO and colleagues.
- Unreleased financials stay out of any tool not approved for confidential data.
HR & people
Great for writing and policy work. Hiring and firing decisions are where AI is regulated, so keep people in charge.
Medium effortFirst steps
- Use AI to draft job descriptions, then check for exclusionary language and required pay-transparency details.
- Turn your employee handbook into a searchable assistant that cites sections.
- Write the company AI-use policy and roll it out with a 30-minute training.
- Inventory every tool that screens, scores or ranks candidates or employees, including features inside your applicant tracking system.
Watch out for
- NYC requires a bias audit within one year before using an automated employment decision tool and notice to candidates 10 business days ahead.
- Illinois has required notice to workers about AI use in employment decisions since January 1, 2026; Colorado's SB26-189 adds duties for automated decisions starting in 2027.
- Never paste résumés or employee records into an unapproved chatbot.
Operations
Documentation, scheduling and process mapping first. Autonomous agents only for steps you can check.
Medium effortFirst steps
- Record a walkthrough of one messy process and have AI turn it into a written SOP with a checklist.
- Pick one weekly report built by copy-and-paste and automate it with a human sign-off.
- Use AI to summarize vendor and customer email threads into status updates.
- Pilot one automation tool (Make, n8n or Zapier) with an approval step before any external action.
Watch out for
- Agent reliability falls fast with task length: METR found the best public agents finished about 12-hour software tasks half the time but only about 1.5-hour tasks at 80% reliability.
- Every automation needs an owner and a log; silent failures cost more than slow manual work.
- Connecting AI to email and file systems widens access; grant the minimum.
Legal & compliance
Contract review and policy drafting save real time. Every citation and conclusion still gets human sign-off.
Medium effortFirst steps
- Use AI to extract key terms (renewal, termination, liability caps, data clauses) from your contract backlog into a table.
- Draft an AI-use policy and a vendor AI questionnaire (training on data, retention, sub-processors).
- Map which laws apply to your AI uses: EU AI Act if you serve the EU, state hiring and privacy laws in the US.
- Set up a log of AI systems in use, with owner, purpose and data types.
Watch out for
- Fabricated citations have led to court sanctions; verify every case and quote against the source.
- Privilege and confidentiality can be waived by careless tool use; use approved, contract-covered tools only.
- Laws are moving monthly; date-stamp every compliance memo.
IT & security
Give people approved tools quickly; blocking everything just moves usage to personal accounts.
High effortFirst steps
- Pick approved AI tools on business plans with SSO, admin controls, data-retention settings and contract terms on training.
- Publish the approved list and a one-line rule for data that never goes into AI tools.
- Review browser extensions and SaaS connectors that request access to email, files or calendars.
- Add AI-specific scenarios (deepfake voice, prompt injection, data leakage) to security training and incident response plans.
Watch out for
- Prompt injection: agents that read email or web pages can be hijacked by hidden instructions; require confirmation before agents send, pay or delete.
- API keys pasted into chats or committed to repos leak; scan for them.
- Phishing no longer has typos; train people to confirm unusual requests through a second channel they start themselves.
Compliance checklist
- EU AI Act: the dates that matter. If you sell into the EU or your AI's output is used there, the AI Act applies. Bans on prohibited practices took effect on February 2, 2025, and duties for general-purpose AI model providers on August 2, 2025. Transparency duties and most remaining provisions applied from August 2, 2026. The AI Omnibus (Regulation (EU) 2026/1744, in force since July 27, 2026) moved high-risk obligations to December 2, 2027 for uses listed in Annex III, such as hiring, credit and education, and to August 2, 2028 for AI built into regulated products. It also softened the AI-literacy duty to taking measures that support staff AI literacy. Regulation (EU) 2026/1744, the AI Omnibus (EUR-Lex)
- EU transparency: tell people it's AI. Since August 2, 2026, Article 50 requires that people be told when they're interacting with an AI system unless it's obvious, that deepfakes be labeled, and that AI-generated text published to inform the public on matters of public interest be disclosed unless a person reviewed it and holds editorial responsibility. Providers of generative systems must mark output in a machine-readable way; systems already on the market before August 2, 2026 have until December 2, 2026. Breaches can cost up to EUR 15 million or 3% of worldwide annual turnover. Regulation (EU) 2024/1689, the AI Act (EUR-Lex)
- New York City Local Law 144: automated hiring tools. Employers and employment agencies using an automated employment decision tool for candidates or employees in NYC need a bias audit completed within one year before use, a public summary of the results, and notice to candidates at least 10 business days before the tool is used. Enforcement began July 5, 2023. Check whether features inside your applicant tracking system count. NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools
- Colorado and Illinois: AI in employment and other big decisions. Colorado repealed and replaced its 2024 AI Act with SB26-189, signed May 14, 2026, which covers automated decision-making technology used in consequential decisions including employment, housing, lending, insurance and health care. It requires notice at the point of use, a plain-language explanation within 30 days after an adverse outcome, a path to human review and correction, and three years of records; developer documentation duties start January 1, 2027, and the attorney general is due to adopt rules by then. In Illinois, since January 1, 2026, employers must notify workers when AI is used in employment decisions and may not use AI that has a discriminatory effect, including by using zip codes as a proxy. Colorado General Assembly, SB26-189 Automated Decision-Making Technology
- FTC: fake reviews, including AI-written ones, are banned. The FTC's Consumer Reviews and Testimonials Rule, in effect since October 21, 2024, bans fake reviews attributed to people who don't exist (AI-generated ones included), buying or selling reviews, compensation conditioned on positive or negative sentiment, undisclosed insider reviews and review suppression. Courts can impose civil penalties for knowing violations; the FTC's December 2025 warning letters cited up to $53,088 per violation. FTC, Final rule banning fake reviews and testimonials (Aug. 14, 2024)
- FTC: back up every AI claim you make. "There is no AI exemption from the laws on the books," the FTC said when it announced Operation AI Comply in September 2024. DoNotPay paid $193,000 over its untested "robot lawyer" claims, and several "AI-powered" online-store income schemes were shut down by courts. If you market AI features, have evidence for performance claims, avoid implying your product replaces a licensed professional, and never promise earnings. FTC, Operation AI Comply (Sept. 25, 2024)
- Data privacy basics for AI tools. Before staff put personal data into an AI tool: list what data each tool receives; confirm in the contract whether the vendor trains on it and how long it's kept; sign a data processing agreement, and a HIPAA business associate agreement if health data is involved; update your privacy notice if AI processes customer data. California's privacy regulator has added rules for automated decision-making technology used for significant decisions, with compliance required by January 1, 2027. Free and consumer tiers often have different data terms from business tiers, so read the tier you actually use.
- Your employee AI-use policy: what to put in it. Keep it to one or two pages: 1) approved tools and how to request new ones; 2) data that never goes into AI tools (customer personal data, health data, credentials, confidential financials, client files) unless the tool is cleared for it; 3) human review before anything leaves the company; 4) disclosure rules for customers and clients; 5) no AI-only decisions about hiring, firing, pay or promotion; 6) who owns AI-assisted work product; 7) how to report a mistake or leak, with no penalty for prompt reporting; 8) a review date. Train on it: only about 1 in 10 small-business workers say they were offered formal AI training. US Chamber of Commerce Foundation and Ipsos, Main Street AI Monitor (June 17, 2026)
Founder corner
- Where the money is going. AI startups raised $102 billion in Q3 2026, 64% of the $159 billion invested globally. A record 27 companies raised rounds of $1 billion or more, and those 27 took around a third of the quarter's venture capital. Half of all Q3 capital went to companies founded since 2022. Aerospace, robotics, data centers, semiconductors and energy each drew $10 billion or more. Read it as a barbell: giant rounds for model labs and infrastructure, and a crowded field for everyone else, where traction beats narrative. Crunchbase News, Q3 2026 global venture funding (Oct. 5, 2026)
- Free cloud credits worth applying for. Google for Startups Cloud Program: AI startups can get up to $350,000 over two years (up to $250,000 in year one and up to another $100,000 in year two), with verifiable pre-seed to Series A funding; third-party models such as Claude on Vertex AI are billed separately and not covered. AWS Activate: up to $200,000 on the Portfolio tier through an accelerator or investor partner for pre-Series B companies founded within the last 10 years, or up to $5,000 self-funded. Microsoft for Startups: up to $150,000 in Azure credits, starting with $5,000 after business verification. Credits expire; architect so you can move when they run out. Google for Startups Cloud Program
- The common AI business models. 1) Seats with AI features: predictable revenue, but heavy users can cost more than they pay. 2) Usage-based: tracks your costs, harder for buyers to budget. 3) Outcome-based: charge per resolved ticket, booked meeting or processed claim; easiest to sell, and you carry the risk when the model fails. 4) AI-enabled services: sell the finished work, like bookkeeping or claims processing, and use AI to run at software-like margins. 5) Vertical workflow software: own a regulated or messy workflow end to end. 6) Developer infrastructure: tools for the people building all of the above. Most durable companies combine a workflow, proprietary data and distribution.
- Pitfall: the thin wrapper. If your product is mostly a prompt and a nice interface on someone else's model, the model vendor can ship your feature in a release note. Test yourself: when the next model is twice as good, does your product get better or get replaced? Defensible products own something the model doesn't: proprietary or hard-won data, deep integration into a workflow with switching costs, distribution, regulatory approvals, or an eval suite that proves you're best at one job.
- Unit economics: do the token math before you set a price. Worked example at Anthropic's list prices. A support reply that reads 3,000 tokens and writes 500 costs about $0.011 on Claude Sonnet 5.5 ($2 / $10 per million tokens) and about $0.00055 on Claude Haiku 5.5 ($0.10 / $0.50). An agent run that re-reads a growing context over 40 steps can easily total 2 million input and 60,000 output tokens: about $9.20 on Claude Opus 5.5 ($4 / $20). Three levers change that math: prompt caching (cache hits on Opus 5.5 cost 5% of the base input price, $0.20 per million), the Batch API (50% off for non-urgent work), and routing easy steps to a smaller model. Price per user on your heaviest 10% of users, never the average. Anthropic, Pricing
- Pitfall: demo reliability vs. production reliability. A demo needs one good run; a customer needs it to work every time. METR measured the best public agents completing about 12-hour software tasks half the time, but only about 1.5-hour tasks at 80% reliability. Build evals from real customer tasks before launch, publish your accuracy honestly, and design the product so a person can catch the failures cheaply. METR, Frontier Risk Report (May 19, 2026)
- Model prices fall under you, so price on value. Claude Opus 5.5 lists at $4 / $20 per million tokens, below the $5 / $25 of Claude Opus 5, and Anthropic made Claude Sonnet 5's introductory $2 / $10 price permanent instead of raising it to $3 / $15 as scheduled. Pricing your product as cost plus a markup hands every price cut to your customers. Price on the value of the outcome, and treat falling model costs as margin. Anthropic, Pricing
- Selling to businesses: pass the security review. Mid-market and enterprise buyers will send a security questionnaire before they sign. Have ready: a data-flow diagram showing which model providers see customer data, contract language on training and retention, SSO, audit logs, a data processing agreement, and a SOC 2 report or a timeline for one. If you serve EU customers, document how you meet the AI Act's transparency duties.
Startup news
Is your business ready?
- Do you have a written rule for which AI tools staff may use, and what data can go into them?No rules yet · Informal rules people mostly know · A written policy · A written policy that staff were trained on and that we enforce
- Can you name one task that eats at least five hours a week and produces output someone can check?Not really · I have a hunch · Yes, one clear task · Yes, and we've measured how long it takes today
- Where does the information that task needs live?Paper, inboxes and people's heads · Scattered spreadsheets · One main system (CRM, accounting, ticketing) · Systems with clean exports or APIs
- Who would own an AI pilot?Nobody in particular · Whoever has spare time · A named person, part-time · A named owner with a goal, a budget and a deadline
- How well do you know what customer and employee personal data you hold?Not sure what we hold · We know roughly, with few controls · We know, and access is restricted · Documented, with vendor contracts (DPAs, BAAs) in place
- How much AI training have your people had?None · Self-taught · One session or a shared guide · Ongoing, with a shared library of prompts and examples
- How would you know if a pilot worked?We'd feel it · We'd ask the team · Before-and-after on one metric · A baseline, a target and a review date set in advance
An interactive version with a personal roadmap is coming. Questions? Read how we work.
Some tool links may earn us a commission. It never changes our recommendations, and we only say we tested something when we did.