Putting AI to Work · Resource
The tools, as of August 2026.
Alicia Dahling · Dahling Consulting
If you do nothing else, pay for one general-purpose assistant at the twenty-dollar tier — ChatGPT, Claude, or Gemini — and use it daily for a month. That covers most of what a business owner, board member, or nonprofit leader actually needs. Everything else on this page is situational: excellent for a specific job, none of it a prerequisite. Every price below was checked in August 2026 and will change; several vendors block automated price checks, so where a figure could not be confirmed on the vendor’s own page, the text says so. For the companies behind these tools, Forbes publishes an annual AI 50 list.
The everyday assistants
| If your task is | Use | Cost | The honest caveat |
|---|---|---|---|
| Everyday writing, brainstorming, deep research reports | ChatGPT | Free; Go $8; Plus $20; Pro from $100/mo; Business $20 per user monthly billed annually | The plan ladder changed twice this year and will change again. Deep Research citations should be clicked, not trusted. |
| Long documents, reports, board memos — where writing quality matters. Also Excel, via the Claude for Excel add-in | Claude | Free; Pro $17/mo annual ($20 monthly); Max from $100/mo; Team $20 per seat annual | No native meeting transcription, and real-time news is its weakest lane. |
| Work that already lives in Gmail, Docs, Sheets, and Meet | Gemini | Free; Plus $4.99; Pro $19.99; Ultra from $99.99 monthly | Its research mode sometimes answers from memory instead of live sources — add “cite all sources” and click two before anything client-facing. |
| An organization standardized on Microsoft 365 | Microsoft Copilot | Business add-on $18 per user monthly annual — first-year promotional pricing through September 2026 (list $21) | Meetings are Teams-only, and you are paying for the integration; the model underneath is OpenAI’s. |
| Research questions that need cited, clickable sources | Perplexity | Free; Pro $20/mo; higher tiers for teams | The citations make answers look verified; they are not. A Columbia Journalism Review audit found roughly a third of citations misattributed. Click through on anything you will repeat. |
| Monitoring X, breaking news, live sentiment | Grok | From $10/mo; standard tier $30/mo | The weakest of the six on reliability in independent 2026 testing. News pulse, yes; anything legal, financial, or client-facing, no. |
What you type may be training data. The business tiers above do not train on your content by default; several free and individual tiers do unless you change a setting. Plain rule: do not paste client lists, donor records, health information, or anything privileged into a consumer chatbot — US or Chinese. The business tiers exist for exactly this reason.
The Chinese models
By May 2026, Chinese open-weight models accounted for roughly 61 percent of all tokens processed on OpenRouter, a major model marketplace, by OpenRouter’s own published data. They are real, they are cheap, and several are genuinely good. Whether you should use one depends less on the model than on how you access it.
The three access lanes, safest last:
- The vendor’s consumer app — free, governed by the vendor’s own privacy terms.
- The vendor’s API — cheapest per token; terms vary widely.
- The open model weights run on US infrastructure — costs roughly double the vendor’s token price, keeps your data on US soil, and is the only lane that fully answers the concern below.
The caveat, stated plainly:
Under China’s 2017 National Intelligence Law, Chinese organizations can be compelled to assist state intelligence work without notifying users. That is the stated basis for the federal defense-contracting restrictions on DeepSeek and for bans in several countries. No US law restricts private-sector use today, but businesses serving federal customers should assume restrictions flow down. Never put client or regulated data into a Chinese vendor’s first-party app or API. Open weights on a US host involve no data flow to China at all.
DeepSeek. The budget benchmark the market prices against; open weights under the MIT license, and the sharpest of the group. Its own privacy policy stores first-party app and API data on servers in China indefinitely. Use it only via US hosts.
Qwen (Alibaba). The widest catalog and the cleanest enterprise story, with a US cloud region and a stated no-training policy. Verify in writing which region your workload routes through — the default international endpoint is Singapore.
Kimi (Moonshot). Strongest for very long documents and agent work. The sharpest terms caveat of the group: its own platform terms say user content — including paid API traffic — may be used for training. Open weights on a US host for anything confidential.
GLM (Zhipu). The best-known low-cost coding alternative, with open weights. The company was added to a US Commerce Department restricted-entity list in 2025 — not a ban on using the model, but paying the vendor directly deserves a conversation with counsel.
MiniMax. Cheap long-context agents plus the strongest video and voice portfolio of the group. Its benchmark claims are vendor-reported; do not repeat them as independent fact.
On your own laptop — or in your pocket
You can run a capable model entirely on your own machine, free, with nothing leaving your computer. That last part is the whole reason to do it.
The laptop apps. LM Studio is the shortest path for a non-technical person: point-and-click, with a hardware check that warns you when a model will not fit. Ollama is the free open-source alternative if a tech helper is involved — but avoid its newer cloud models (names ending in :cloud); those run on servers, which defeats the purpose.
What to run. On a 16GB laptop: OpenAI’s gpt-oss-20b, which OpenAI says performs near its own early-2025 mini cloud model — that is the honest yardstick, not today’s frontier. If you need it to read scans and charts: Google’s Gemma 4 12B. On a 32GB machine: Alibaba’s Qwen3.6-27B is the most capable thing that fits — and since a locally run model sends nothing anywhere, its origin stops mattering.
On your phone. The free open-source app PocketPal AI (iPhone and Android) or MLC Chat (Android) runs a small model — Qwen 1.7B or Gemma 1B class — entirely on the device: download the model once, then it works in airplane mode. Good for private drafting and summarizing your own notes; not a research tool. Your phone may also already have one built in — recent Androids ship Gemini Nano on-device, and recent iPhones run Apple’s on-device model.
What local is honestly good for. Summarizing and rewriting your own documents, first drafts, reformatting notes — all without anything leaving your machine. Not good for: anything requiring current information, verifiable citations (local models invent sources confidently), legal or tax analysis, or math on real financial data without checking.
A note on notes. If you want your notes themselves to stay private, Obsidian stores them as plain files on your own machine — no cloud, and no AI reads them unless you invite one.
For branding and marketing
Canva. The default all-in-one for a small business: social graphics, flyers, decks, and brand consistency, with AI generation in the same editor. All AI features draw from one shared credit pool, and daily image work can exhaust it.
Midjourney. The highest-end brand imagery, from $10 a month. Two caveats from its own terms: on the lower tiers everything you generate is publicly visible — client work effectively requires the $60 tier — and businesses grossing over $1M a year need the higher plans for commercial use.
Adobe Firefly. The commercially-cautious choice: Adobe states it trains only on licensed and public-domain content. From $9.99 a month; everything is credit-metered, and video burns credits far faster than images.
ChatGPT image generation. Best at doing exactly what you asked in plain English, included in plans most owners already have. A generator, not a design tool — fine text and exact logos drift between generations.
Ideogram. The pick when the graphic must contain legible words: banners, posters, signage.
Gamma. The fastest path from notes to a presentable deck, free tier available. If the deliverable must be an editable PowerPoint file, test the export early.
Higgsfield. Social-first product shots and short video from one subscription, aggregating several video models under one roof. The entry tier runs out of credits fast on video.
Your existing assistant as the copy desk. Save a one-page style guide into a Claude Project or ChatGPT custom instructions and you have covered most of what dedicated brand-voice tools charge for. Never let an AI post replies as you.
HeyGen. AI avatar video: record yourself once, then generate presenter-style videos — popular with real-estate agents for listing walkthroughs and multilingual versions. Two cautions: disclose AI-generated presenters wherever the video runs (platform rules and several state laws are tightening), and remember that the reason people hire an agent is the agent — an avatar spends trust, so use it for volume tasks, not for the introduction.
The logo warning. The US Copyright Office’s position, unchanged through 2026: work generated entirely by AI from prompts has no copyright protection. When an image tool’s terms say you own the image, that is a license statement, not a copyright guarantee — a competitor can legally copy an uncopyrightable logo. For flagship brand assets, have a human designer rework the AI concept and document the human contribution.
The real-estate photo warning. Realtor ethics rules require clear disclosure of virtually staged or AI-altered listing photos, MLS policies typically prohibit altering permanent features, and California’s AB 723 additionally requires access to the original image. Rules vary by MLS board and are tightening; check your board’s written policy.
Agent platforms
Agents are AI tools that take actions — browse, build, send — rather than just answer. With agents, the failure mode is the more important half of the review.
Manus. Best for delegating a whole open-ended task: research these ten competitors, pull their pricing, build me the spreadsheet. It runs in its own cloud computer and hands back finished files. Failure mode: unpredictable credit burn — similar tasks can cost very different amounts — and mid-task failures on long runs.
Perplexity’s browser and agent features. Best for cited research for people who will never build anything. Failure mode: prompt injection, and it is structural — security researchers have shown hidden text on a webpage can hijack a browsing agent. Never let an agentic browser act while logged into email, banking, or anything holding client data.
MindStudio. Best for packaging a repeatable workflow — an intake summarizer, a grant-letter drafter — as a button someone else can press, without code, with model usage billed at cost. Failure mode: it does nothing until you design the workflow; people expecting a just-do-it agent stall at the blank canvas.
Lindy. Best for email triage, scheduling, and follow-up from templates, for owners drowning in inbox volume. Failure mode: bill shock from the credit meter underneath the sticker price — and an email-sending agent with a bad instruction sends bad emails to real people. Pilot in drafts-only mode first.
Building your own. Best for one high-value, well-defined recurring workflow where nothing off the shelf fits. For most readers the realistic route is hiring it out: a few thousand dollars up front plus ongoing costs. Failure mode: maintenance — a custom agent is software, and an unmaintained agent silently rots. Do not build custom for anything a tool above does adequately.
If you would rather opt out
Declining is a legitimate strategy, not a failure to keep up. If you want less AI in your life rather than more:
Turn off the training toggles. In ChatGPT: Settings, Data Controls, turn off model improvement. In Gemini: turn off Gemini Apps Activity. Claude’s consumer tier does not train on your conversations by default. The working guide shows each switch.
Quiet your inbox. Gmail’s AI features live under Settings — turn off Smart features and personalization, and the smart replies, summaries, and nudges stop. Your mail still works exactly as before.
Truly private email exists. Proton Mail, run by a Swiss foundation, encrypts your mail so that not even the provider can read it — which also means no AI reads it. Free tier available.
Or go local. The laptop and phone sections above are the middle path: the usefulness without the cloud.
Your opt-out checklist
Tick these off as you go. Selections stay in this browser only, on this device. Nothing is sent anywhere.
Loading your saved progress
The judgment about where AI does not belong is the sharpest tool on this page — that list is on the AI page.
Every price and claim on this page was checked in August 2026 and will change. Where a claim is a vendor’s statement about its own product, the text says so. Nothing here is a legal opinion — the export-control, copyright, and MLS items are pointers to ask counsel. If you are reading this in 2027, treat every number as a starting point for your own check.
Questions, or a tool you think belongs here? alicia@dahlingconsulting.com