“Which AI is best?” posts age within a month: models are renamed, overtaken and re-priced faster than any comparison can keep up. A more durable map groups tools by the job they do — answering, researching, automating, analysing, building, creating media, presenting — because the governance question depends on the job. A chat assistant raises a data-retention question; an automation platform raises a permissions question; an image generator raises a rights and deepfake question. Every tool below is one an employee can start using in minutes, which is exactly why each is also a possible source of shadow AI.
1 · Twenty-Two Tools by Job
| Job | Tools | Good for | Ask before using with work data |
|---|---|---|---|
| General assistants | ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Microsoft Copilot, Grok (xAI), DeepSeek | Drafting, analysis, coding help, summarising long documents | Are we on an enterprise plan where inputs are not used for training and retention is under our control? Where is the data processed? (Several governments restrict DeepSeek’s app on official devices.) |
| Research | Perplexity, NotebookLM (Google) | Answers with cited sources; Q&A over documents you upload | Do the citations actually say what the answer claims? Who can see the uploaded sources? |
| Automation & agents | n8n, Zapier | Connecting apps and AI steps into workflows | What credentials and OAuth scopes does each workflow hold, and who owns it when its creator leaves? |
| Data & analytics | KNIME, Alteryx, Excel with Copilot | Data preparation, analysis and models without heavy coding | Does the sensitive dataset leave our environment? Is the output checked before it reaches a decision? |
| Build | Lovable, Hugging Face | Generating working apps from prompts; finding and hosting open models and datasets | Has the generated app had a security review? Are downloaded models and datasets from a trusted source, with acceptable licences? |
| Media generation | Sora (OpenAI), Veo (Google), Runway, invideo, Stable Diffusion, Adobe Firefly | Video, images and marketing assets | Do we have rights to commercial use? Is AI-generated content labelled? Could it impersonate a real person? |
| Presentations | Gamma | Turning notes into slide decks and documents | Are confidential board or client materials being uploaded to a personal account? |
Tools listed by job, not ranked. Several span more than one job, and most add features monthly. Vendor names are given only to identify products.
2 · Five Checks Before Any Tool Sees Work Data
- Account type — an enterprise or team plan under a company contract, signed in with company SSO; never a personal account.
- Training and retention — inputs are not used to train models, and stored data can be deleted on request.
- Data location and processors — where prompts and files are processed and stored, and which sub-processors see them (a DPDP and contract question).
- Permissions — for tools that connect to email, drives or other apps, the minimum scopes, reviewed regularly.
- Output rules — a human checks anything sent, published or acted on; AI-generated media is labelled.
▸ A tool that passes all five can go on an approved list. Publish that list: people use whatever is easiest, so the approved option has to be easy to find and easy to get.
3 · How to Choose Without a Leaderboard
| Question | Why it matters more than a benchmark |
|---|---|
| Does it fit where the work already happens? | An assistant inside the email, documents and chat your teams already use gets used; a separate one gets copy-pasted into |
| Can it be governed? | SSO, audit logs, admin controls and data-retention settings decide whether security can approve it at all |
| Does it do your task well? | Test on twenty real, non-sensitive examples from your work; public benchmarks measure something else |
| What happens when it is wrong? | Low-stakes drafting tolerates errors; legal, financial and security outputs need checking built into the workflow |
| Can you leave? | Prompts, workflows and outputs should be exportable, so switching tools later is possible |
Writing better prompts for whichever tool you choose: Prompt Frameworks, Compared. The terms: AI Glossary.
Bottom line
- Map by job, not by leaderboard — the jobs stay stable while the rankings churn.
- The job decides the risk: retention for assistants, permissions for automation, rights for media.
- Five checks before work data: account, training and retention, location, permissions, output rules.
- Publish an approved list and make it easier than the alternatives.
- Test on your own work, and keep a human accountable for what goes out.
Sources & method. Tool descriptions from each vendor’s public product information as of November 2026; capabilities and terms change frequently, so confirm current plans and data terms before adoption. Grouping, governance questions and checks are the author’s analysis. No vendor relationship or endorsement is implied. Original analysis; no third-party graphics or text reproduced.
← Back to Cyber security