A field guide to today's AI toolsJuly 2026

The AI
Toolbox.

What exists beyond the chat window — what it can do, and what it costs

chat agent toolbox
2
The hook

You know the chat window. That's maybe 10% of the story.

What most people ask

"Write me a post about our new savings account." → you get a paragraph of text to copy-paste, format and finish yourself.

What you could ask

"Here's the brief and last year's campaign folder — build the deck, the landing page and a voiced video draft." → you get finished files to review.

Often it's the same AI underneath. What changed is the tools around it — and this talk is a guided tour of those tools, price tags included.

3
Agenda

What we'll cover

The core idea

Chatbot vs agent, the loop, and why the model forgets — in five slides.

The big three

Anthropic, OpenAI, Google — their full toolboxes, plan by plan.

What you can do

A week of real tasks — decks, podcasts, pages, research.

The wider landscape

Video, voice, avatars, images, decks, automation — the deep dive.

Five wild stories

What agents actually did in the real world — good and bad.

Plugging it in

Excel, PowerPoint & co. — plugins, connectors and MCP.

Prices travel with the tools — every tool comes with its price tag, as we go.

4
Basics · pricing literacy

First, learn to read an AI price tag

1 · Subscription

Flat $/month, capped by usage limits. The labs: free → ~$20 → $100–200. Simple, predictable.

2 · Credits

Buy a bucket; each image/clip/minute burns some. Most creative tools. Hardest to budget — retries burn credits too.

3 · Per-seat

$/user/month for teams — adds admin controls, SSO, "no training on your data". Usually ~$20–40/seat.

4 · API (per token)

Metered pay-per-use for building software on top. No cap; scales with usage.

same bill every month each generation burns some a bucket that drains — retries too × $ per person, per month a meter that runs as you use it

And the trap to remember: free tier ≠ free for commercial use. More on that in "the fine print". All prices in this deck: July 2026 snapshot — verify before quoting.

5
Basics · chatbot vs agent

A chatbot answers. An agent acts.

Chatbot — the one you know

You ask → it tells you what to do → you open the files, do the steps, paste results back… and repeat.

Agent

You ask → it does the work itself — reads the files, takes the steps, checks the result → you review at the end.

chatbot — you run every step you chatbot open the file paste it back fix the format …every step round-trips through you, and repeat agent — one ask, finished files you agent one ask finished files you review at the end

Often it's the same model underneath. The difference is everything wrapped around it — and every tool in this talk is some flavour of that wrapping.

6
Basics · the loop

An agent is a loop

  • the model proposes an action → the tool runs it the result comes back. Repeat until ④ done.
  • Like a cook: read the recipe → do a step → taste → adjust → repeat.
  • This loop is the whole "magic" of agentic AI — no magic, just disciplined repetition.

Every agent product in this deck — Cowork, Codex, agent mode, all of them — is this loop in a different costume.

7
Basics · the brain

The model has no memory

  • A model is autocomplete on steroids — it only ever predicts the next word. Text in, text out. No database, no "you".
  • Think of the guy in Memento — the film hero who can't form new memories, living off notes & tattoos: each interaction, the model has no idea what came before.
  • So the tools around it must re-send everything it should know, every single time.

Keep this in mind — it explains why tools plug into your files, your brand kit, your inbox: they're the model's notes & tattoos.

8
Basics · cheat sheet

Words you'll keep hearing

TokenA chunk of text ≈ ¾ of a word. What models read & bill in.
PromptThe text you send the model.
ContextEverything the model sees in one call — its short-term memory.
AgentAn AI that does things in a loop, not just answers.
HarnessThe program wrapped around the model — its hands.
ConnectorA plug between an AI and another app (Gmail, Drive, Slack…).
MCPThe open standard those plugs are built on — "USB-C for AI".
CreditsPre-paid units most creative tools bill in — each generation burns some.
Generative AIAI that makes media — images, video, voice, music.
Deep researchThe AI browses for 5–30 min and returns a cited report.

Don't memorise these — each comes back with its own slide. This is just your map.

9
Basics · so what?

What agents are actually good for

Draft & research

Summaries, first drafts, comparisons, cited reports.

Work over your documents

Ask questions across a whole folder, inbox or wiki.

Make real files

Genuine decks, spreadsheets, pages, audio — not just text.

Automate chores

Sort, rename, triage, fill in — the repetitive stuff.

Build pages & apps

A landing page or small tool without an IT ticket.

…with a human in the loop

You review & approve. A power tool, not a replacement.

01

The big three

Anthropic, OpenAI, Google. The same $20 entry fee — three very different toolboxes.

10
Big three · models

The brains: today's models

MakerFlagship lineup (July 2026)Known for
AnthropicClaude Fable 5 · Opus 5 · Sonnet 5 · Haiku 4.5coding, agents, long context
OpenAIGPT‑5.6 — Sol · Terra · Lunageneral reasoning, Codex
GoogleGemini 3.6 Flash · 3.5 Flash‑Lite (3.5 Pro still unreleased)multimodal, large context

The same tiers repeat everywhere: large (smartest, priciest) · medium (the sweet spot) · small (fast & cheap) — plus a frontier tier on top lately. It reshuffles every few months — treat any list as a snapshot. For today, the products matter more than the model names —

11
Big three · anthropic

The Claude family

Claude app

The chat assistant — web, desktop, mobile. Drafting, analysis, files, web search.

Claude Code

The coding agent, for developers — terminal & IDE.

Claude Cowork

The agent for everyone else — point it at a folder of files. Next slide.

Claude Design

Prompt → decks, pages & visuals in your brand. Two slides on.

Claude in Chrome

A browser extension that reads pages, clicks and fills forms with you.

Claude for Microsoft 365

Side panel inside Excel, PowerPoint, Word & Outlook — see "plugging it in".

One subscription, not six products: everything here is included from Pro at ~$20/mo (free tier = chat only). Prices as of July 2026.

12
Big three · the coding agents

Claude Code & Codex — the coding agents

Claude Code

Anthropic's agent for developers: lives in the terminal or editor, reads a whole codebase, writes & runs code, fixes its own mistakes. Included from Pro (~$20/mo).

Codex

OpenAI's equivalent, on GPT‑5.6 — terminal, IDE and a cloud version that works on tasks while you're away. Bundled into ChatGPT plans (full access from Plus, ~$20/mo).

Why you should care even if you never code: these transformed programming first — Anthropic says over 80% of its own production code is now written by Claude. Then someone asked: what if we removed the code part?

13
Big three · claude cowork

Claude Cowork — an agent for your files

  • Point it at a folder, describe an outcome — it reads what's there, makes a plan, and works through it step by step.
  • It creates & edits real files — documents, spreadsheets, decks — and shows you each step; you approve the sensitive ones.
  • Keeps working while you're away; can run on a schedule (weekly report, every Monday).
  • Generally available since April 2026 · included in every paid plan (from Pro, ~$20/mo).
Claude Code,
without the code

Anthropic saw developers using its coding agent for non-coding work — organising files, planning, reports — so it removed the terminal. Cowork was built in under two weeks, mostly by Claude Code itself.

14
Big three · claude design

Claude Design — describe it, see it

  • Describe what you want → Claude builds a first version → refine by chatting, inline comments, or editing directly.
  • Makes decks, one-pagers, landing pages, mockups, marketing collateral — built for people without a design background.
  • Exports to PDF, PowerPoint, HTML, a live URL — or into Canva.
  • Launched April 2026 (Anthropic Labs, research preview) — 1M+ users in week one. Included in paid plans, drawing on plan usage.
It knows
your brand

The headline feature: import your company's design system & brand files, and everything it makes comes out in your visual identity — not generic AI-style slides.

15
Big three · openai

The ChatGPT family

ChatGPT app

The chat assistant — Free · Go ~$8 · Plus ~$20 · Pro $100–200/mo.

Codex

The coding agent — bundled in; full access from Plus.

Agent mode

Plans & executes multi-step tasks — browses, runs code, hands back real decks & spreadsheets.

Images 2.0

Image generation & editing in-chat — 2K output, consistent characters across a batch.

Custom GPTs

A reusable, pre-configured ChatGPT for one job ("our tone-of-voice editor") — shareable with the team.

⚠ Watch out

Free & Go now show ads. And the famous Sora video app? Gone — story in a moment.

Same ladder as Claude: ~$20/mo (Plus) is where the real toolbox unlocks. Prices as of July 2026.

16
Big three · google

The Gemini family

Gemini Notebook ⭐

(ex-NotebookLM, renamed July 2026.) Upload your documents → cited answers from only those sources — and Audio Overviews: your PDFs as a two-host podcast. Core features free.

Gemini app

The chat assistant — Free · AI Plus ~$5 · AI Pro ~$20 · Ultra $100–200/mo.

Veo 3.1 + Flow

Text-to-video (with synced dialogue) + a filmmaking front-end. The mainstream AI-video product — paid in credits via the plans (Pro ≈ 1,000/mo).

Gemini in Workspace

Built into Gmail, Docs, Sheets, Slides, Meet — now bundled in base plans (from ~$14/user/mo), no separate add-on.

Gemini Notebook is the classic first "wow": drop a 40-page report in, get a podcast out — on the free tier. Prices as of July 2026.

17
Big three · the $20 question

Same $20 — different toolbox

Claude Pro · $20ChatGPT Plus · $20Google AI Pro · $20
Agent that does workCowork (your files)Agent mode
Visuals & decksClaude DesignImages 2.0image gen
Office / docs integrationExcel · PowerPoint · Wordconnectors + add-insGmail · Docs · Sheets
Digest documentsprojectsDeep ResearchGemini Notebook + podcast
AI video— (Sora gone)Flow · ~1,000 credits
Coding agentClaude CodeCodexAntigravity · Jules

Honest advice: start with the ecosystem you already live in — and use the free tiers to find out which one that is. July 2026 snapshot.

18
Big three · the plan ladder

The ladder is the same everywhere

TierClaudeChatGPTGoogle AI
Freechat, limitedchat, limited (+ ads)chat + Notebook basics
EntryPro ~$20Go ~$8 · Plus ~$20Plus ~$5 · Pro ~$20
PowerMax ~$100 / $200Pro ~$100 / $200Ultra ~$100 / $200
Team~$20–25/seatBusiness ~$20–25/seatWorkspace ~$14–22/user

Higher tiers buy more usage & early access — not a different product. The team tiers are where admin controls, SSO and "no training on your data" live — for work accounts, that's the row that matters. Prices July 2026 — they change often.

19
Big three · a warning label

All of this happened in about a year

Sora ☠

OpenAI's video app, shut down Mar 2026 — reported costs of $1M–15M/day vs ~$2M lifetime revenue

Atlas ☠

OpenAI's AI browser, launched Oct 2025 — retiring 9 Aug 2026

Gemini 3.5 Pro ⏳

announced for June 2026 — still unreleased, delayed over coding performance

NotebookLM →

renamed Gemini Notebook, July 2026 — same product, new name

Freepik →

rebranded to Magnific, Apr 2026 — your subscription moved with it

Tome ☠

2023's most-hyped AI deck tool — dead Apr 2025, pivoted away

Two lessons: verify before quoting anything AI-related older than a quarter — and never build a process on one tool without a fallback.

02

What can you actually do with this?

One ordinary week of tasks — running example: a marketing person. Every task translates to other jobs.

20
In practice · a week of tasks

A week of real tasks

Mon · the review deck

"Results CSV + last quarter's deck → this quarter's 12-slide review." (Cowork / agent mode)

Tue · the podcast

"This 40-page study → a 10-min podcast for the commute." (Gemini Notebook)

Wed · the formats

"One key visual → every social format, on brand." (Canva)

Thu · the microsite

"A landing page for the promo — today, without a ticket." (Claude Design / Lovable)

Fri · the research

"Competitors' pricing → a cited comparison table." (Perplexity / deep research)

Any day · the voice-over

"This script in three languages — same voice." (ElevenLabs)

None of this is futuristic — every tile is a tool from the last section doing its day job. Each returns in detail in a minute.

21
In practice · worked example

Monday's deck, step by step

you › "build the Q2 review deck from this folder, same structure as Q1" …the agent, on its own: ✓ read the folder 14 files — results CSV, Q1 deck, brand template ✓ proposed a structure 12 slides, mirroring Q1 — waits for your OK ✓ built Q2-review.pptx charts from the CSV, your template & fonts ⚠ flagged two gaps "no June numbers for channel X — placeholder" you › add the June export… review, tweak two slides, done ✓

Notice what it didn't do: invent the missing numbers. And notice your role — you moved from making the deck to directing & reviewing it.

22
In practice · deep research

Deep research: one question, a full report

  • Ask one question → the AI browses dozens of sources for 5–30 minutes → returns a multi-page, cited report.
  • Shipped by all three labs (and Perplexity) under roughly this name; included from the ~$20 tiers.
  • Ideal for: market landscapes, competitor teardowns, "what does the research actually say?"
Looks right ≠ is right

These reports look authoritative and still contain errors. Treat them as a well-organised starting point with a reading list — never a finished document you forward.

one question browses dozens of sources 5–30 min, on its own ¹ ² a multi-page report, every claim cited — you check the citations
23
In practice · the rule

AI drafts. You sign.

Review scales
with risk

An internal brainstorm needs a skim. Anything public, legal or financial needs a real review. The tool drafts fast precisely so that you can afford to review well.

  • Every tool in this deck assumes a human checkpoint — that's what makes them safe to adopt.
  • The failure mode isn't "AI made a mistake" — it's "nobody looked."
  • So: which tool for which job? That's the rest of this talk →
skim proper look full review internal draft customer email public · legal · financial risk →
03

Beyond the labs: the wider landscape

Ten categories, one slide each: what it does · who leads · a real use case · the price. Tool names churn — the categories stay.

24
Landscape · the map

Ten categories, four jobs

MAKE MEDIA video — Veo + Flow avatars — Synthesia voice — ElevenLabs music — ElevenLabs Music images — Canva MAKE DOCUMENTS decks — Gamma copy — the $20 assistant BUILD & AUTOMATE apps from a prompt — Lovable automation — Zapier FIND THINGS OUT research — Perplexity the $20 assistants Claude · ChatGPT · Gemini

Names churn; the four jobs are stable. One leader per box — July 2026. The cheat sheet at the end of the section holds the full table.

25
Landscape · prompt-to-app

Describe an app, get an app

Lovable

The leader (~$500M annual run-rate, June 2026). Chat → a real, hosted site with forms & a database. Free · Pro ~$25 · Business ~$50/mo, credit-based.

v0 · Vercel

The design-quality pick — polished pages from a prompt or a Figma file. Free · Premium ~$20/mo.

The rest

Bolt.new (friendliest free tier) · Replit (real backends, ~$20) · Figma Make (inside Figma seats).

  • Use case: a campaign microsite, a savings calculator, a quiz — or a clickable prototype that aligns stakeholders in a day instead of a month.
  • Caveats: credits burn fast when you iterate — and anything public-facing still needs a security review: this is real code handling real data.

Prices ~July 2026, credit-based — verify before budgeting.

26
Landscape · ai video

Type a sentence, get a shot — not a film

Veo 3.1 + Flow

The leader. Best fidelity, the only one with synced dialogue. Via Google plans: ~$20/mo ≈ 1,000 credits (≈ 10 top-quality clips).

Higgsfield

The aggregator — 15+ models under one subscription, preset cinematic camera moves. Tiers ~$15 / $39 / $99/mo.

Runway

The editor's tool — generation + a real editing suite. $15 / $35 / $95/mo. Standard ≈ 4–6 finished 10-sec clips/mo.

Klingbest value, strongest motion — from ~$7/moChinese vendor — check data policies
Pikaplayful effects for social — from ~$10/moscroll-stoppers, not polish
Lumasmooth, dreamlike motion — from ~$30/momood loops & abstract brand visuals

Reality check: 5–10-second shots, credits burn on discarded takes too. Use for b-roll, social clips, animating an approved still. Prices ~July 2026.

27
Landscape · hyperframes

HyperFrames — video as code

  • Open-source (by HeyGen): you build a video like a web page — and it renders exactly what you designed, every time.
  • Use cases: 200 personalised variants from one template · monthly data videos (numbers change, design doesn't) · localisation by swapping text.
  • Brand compliance, solved: fonts, colours, logo are locked in code — reviewers approve the template once, not every output.
  • Price: the framework is free (Apache 2.0, render on your own machine); optional cloud rendering billed per minute by HeyGen.
Generative dreams.
This one obeys.

Veo & Runway imagine footage — you can't fully control the result. HyperFrames is a factory: same template, new data, a thousand on-brand videos. Ask a technical colleague to set it up once.

28
Landscape · avatars

A presenter who never books a studio

Synthesia — corporate

Script → presenter in 160+ languages. Training & explainers you re-render, not re-shoot, when things change. Free 10 min/mo · Starter $29 · Creator $89/mo.

HeyGen — marketing

Its party trick: video translation with lip-sync — the original speaker appears to speak another language, mouth matched. Free 3 videos · from ~$29/mo.

D-ID is the budget option (~$16/mo): animates a still photo into a talking head. Rule of thumb: Synthesia for training, HeyGen for marketing & localisation. Prices July 2026.

29
Landscape · voice

ElevenLabs — the voice everyone uses

  • Text-to-speech that genuinely sounds human · voice cloning (a synthetic copy of a consenting speaker) · dubbing into 29+ languages that keeps the speaker's vocal character.
  • Use cases: re-record a voice-over instantly when one word changes · one consistent brand voice everywhere · audio versions of articles & reports.
  • Rule of thumb: ~1,000 credits ≈ 1 minute of audio.
$5 buys the licence

Free: 10 min/mo — no commercial licence, watermarked, attribution required. Starter ~$5/mo: commercial licence + instant voice cloning. Creator ~$22 (adds professional cloning) · Pro ~$99/mo. The cheapest professional "wow" in this deck.

Prices July 2026 — credit-based, shared across speech, dubbing & effects.

30
Landscape · music & editing

Soundtracks & the sleeper hit

Suno

Describe a song → finished track with vocals. Pro ~$10/mo incl. commercial rights — but training data is in litigation: legal check before paid media.

Udio

Settled with the labels; rebuilding as a licensed "walled garden" — downloads currently disabled. A tool that can't export isn't a tool yet.

ElevenLabs Music

Licensed catalogue (Merlin/Kobalt, royalties paid) — the legally safest AI music route.

Descript — edit video like a document

It transcribes your recording; delete a sentence in the text and it's cut from the video. One-click "um" removal, studio-sound cleanup. ~$16–24/mo — the highest ROI per hour of learning in this entire deck.

Prices ~July 2026.

31
Landscape · images & design

Images: pick by job, not by hype

Canva Magic Studio

The daily driver. AI inside the tool teams already use — layouts, image gen, magic resize, brand kit enforced. Free tier genuinely usable · Pro ~$12–18/mo.

Ideogram

The one that gets text inside images right — posters, banners, packaging with real copy. From ~$8/mo, commercial use on paid tiers.

Adobe Firefly

The one legal signs off: trained on licensed content, enterprise indemnification. In Photoshop too. From ~$10/mo.

Midjourney

Highest aesthetic ceiling — concepts & mood boards. $10–120/mo, no free tier — and cheaper tiers are public: never paste unreleased work.

Footnote: Freepik is now Magnific (Apr 2026) — stock + AI + a best-in-class upscaler, from ~€6/mo. Prices ~July 2026.

32
Landscape · presentations

Decks from a prompt

Gamma

The leader. Topic or document in → designed deck out; exports to PowerPoint, publishes as a website. Free to try · paid ~$10–25/mo.

Canva presentations

Already included if you have Canva — better brand control, good presenter mode.

Beautiful.ai

"Smart slides" that keep layouts tidy — but $40/user/mo, no free tier. Hard to justify vs the other two.

Also in this race: Claude Design (slide 14) and ChatGPT's agent mode. And Tome — 2023's most-hyped deck tool — is dead: the churn lesson, again. Prices ~July 2026.

33
Landscape · copywriting

The category the chatbots ate

Jasper

The old category king — revenue roughly halved. Survives on brand-voice consistency across big teams. ~$39–69/mo.

Copy.ai

Left the category — pivoted to sales automation. Agents from ~$249/mo. The pivot tells the story.

Writer

The governance play: flags copy that breaks brand or compliance rules before human review. ~$39/user · enterprise.

Notion AI

AI over your own wiki — "summarise every campaign retro mentioning X." Now needs Business, ~$20/user/mo.

The takeaway: for writing itself, the $20 general assistant is now as good. Pay for these only for governance at scale. Prices ~July 2026.

34
Landscape · research & search

Perplexity — search that shows its sources

  • Ask a question → an answer in prose with citations to live sources — every claim verifiable, not ten blue links.
  • Pro ~$20/mo: unlimited pro search + 20 deep-research runs a day. Enterprise ~$40/seat.
  • Comet — its browser with the assistant built in — went free worldwide in October 2025 (before that, gated behind the $200 Max tier).
Careful where it clicks

An AI browser acting on pages you're logged into is a documented prompt-injection risk — a hidden instruction on a page can steer it. Keep agentic browsers away from sensitive logged-in systems.

Prices ~July 2026.

35
Landscape · automation

The glue: when X happens, do Y

Zapier

The leader — 8,000+ apps; describe the automation in plain English. Free 100 tasks/mo · from ~$30/mo.

Make

Visual drag-and-drop canvas, ~3–5× cheaper per operation (~$10–34/mo, per workspace). Steeper curve.

n8n

Open-source, self-hostable — the answer when data must not leave your own infrastructure. Free to self-host.

Relay.app · Lindy

Relay: built-in human approval steps — AI drafts, pauses for sign-off. Lindy: "AI employees" for a whole inbox. From ~$19–20/mo.

form submitted when X happens… AI drafts the reply ⏸ pauses here until you approve send the email log it to the CRM post to the channel

The pattern to copy: the pause in the middle — AI drafts, a human signs, then the machines fan out. Prices ~July 2026.

36
Landscape · the fine print

Before you use any of it for real

① Licence

Free tier ≠ commercial use. ElevenLabs, Suno, Ideogram: the commercial licence starts at the first paid tier.

② Privacy

Cheap tiers are often public by default — Midjourney below Pro shows your prompts & images to everyone. No unreleased work.

③ Provenance

Where did the training data come from? Firefly & ElevenLabs Music are licensed — the defensible picks. Others are in litigation.

④ Budget

Credit pricing is slippery — retries burn credits. Start on fixed plans, auto-top-up off.

And once more: every price in this deck is a July 2026 snapshot. Check the vendor's page before you commit a budget.

37
Landscape · cheat sheet

The whole landscape on one slide

JobLeaderAlso look atEntry price
Everyday assistant + agentClaude · ChatGPT · Geminifree → ~$20/mo
Page or app from a promptLovablev0, Bolt, Replitfree → ~$25/mo
Generated videoVeo + FlowHiggsfield, Runway~$15–20/mo
Templated / data videoHyperFramesfree (open source)
Avatar presenterSynthesiaHeyGen, D-IDfree → $29/mo
Voice & dubbingElevenLabs~$5/mo (commercial)
MusicElevenLabs Music (safest)Suno (check legal)~$10/mo
Images & designCanvaIdeogram, Firefly, Midjourneyfree → ~$12/mo
DecksGammaClaude Design, Canvafree → ~$10/mo
Cited researchPerplexitylab deep-research modesfree → $20/mo
AutomationZapierMake, n8n, Relayfree → ~$30/mo

Screenshot this one. Names churn; categories are stable — this is how you re-orient in six months. July 2026.

04

Five wild stories

True, documented, recent — what agents actually did out in the world. Good and bad.

38
Stories · 1 of 5

Claude ran a shop — and went broke

  • Anthropic's Project Vend: Claude ran a real snack shop in the office — buying stock, setting prices, chatting with customers on Slack.
  • Month one: −$200. Employees sweet-talked it into discounts, it hallucinated inventory — and once ordered tungsten cubes because someone asked.
  • Phase two — proper tools, a CRM, guardrails, an "AI CEO" — and the losing weeks largely disappeared.
Powerful
and naive

The raw model was both at once. What turned it from a pushover into a business? Tools and guardrails — the thesis of this entire deck.

39
Stories · 2 of 5

Ten years of research. Two days.

10 yrs → 2 days

Imperial College London spent a decade on how superbugs acquire resistance. Google's AI co-scientist, given the same question, ranked the correct — unpublished — answer first in two days.

  • It also proposed a hypothesis the team had never considered. The researchers themselves went on record — this isn't vendor marketing.
  • The honest caveat: the AI generated the hypothesis. Humans had spent the decade doing the lab work that proved it.
  • The point: it didn't replace ten years of science — it means the next question might not need them.
40
Stories · 3 of 5

Gold medals, in minutes

CFA Level III

The hardest finance exam — essays included. Models passed a full mock in minutes (best ~79%). Humans train ~300 hours per level.

Maths Olympiad gold

Official IMO gold medal, certified by the organisers — 5 of 6 problems, standard conditions. Only ~1 in 10 human contestants got gold.

human candidate ≈ 300 hours of study — per level the models minutes — full mock passed (best ~79%)

The reframe: the exam-shaped part of professional work — recall, procedure, structured answers — is now cheap. Judgment, accountability, and knowing what to ask are not.

41
Stories · 4 of 5

The same skill, both directions

🛡 Defence

Google's Big Sleep agent found a flaw in software on billions of devices — one only attackers knew — and it was patched before they could strike. A first.

⚔ Offence

Anthropic disclosed the first largely AI-orchestrated espionage campaign: a jailbroken coding agent ran ~80–90% of the attack autonomously (Nov 2025).

It's the same loop from slide 6 doing both jobs. Which one it does is decided by the humans and guardrails around it — which is why the fine-print slides in this deck aren't decoration.

42
Stories · 5 of 5

The best AI ad of 2026 was made by humans

  • Anthropic's Super Bowl debut — "Ads are coming to AI. But not to Claude." — was made the old-fashioned way: human agency, human director, real actors.
  • Result: daily active users +11%, Claude into the top-10 free iPhone apps.
  • Meanwhile Coca-Cola's AI-generated Christmas ad reportedly took ~70,000 generated clips and ~100 people to assemble two finished spots.
Floor, not ceiling

AI raised the floor — what a small team can produce. It hasn't raised the ceiling. Craft, taste and a real idea still win.

05

Plugging it into your tools

A chat window is a silo. The real power arrives when AI reaches Excel, PowerPoint, mail and your drives — safely.

43
Plugging in · mcp

MCP = "USB-C for AI"

without MCP — N×M glue with MCP — connect once MCP
  • Before: a custom connector for every AI × every tool. MCP is the standard plug: build once, works everywhere.
  • Anthropic open-sourced it (Nov 2024) → OpenAI, Google & Microsoft adopted it (2025) → donated to the Linux Foundation's new Agentic AI Foundation (Dec 2025). 10,000+ servers exist.
  • Every "connector" or "app" on the next slides is this, under the hood. You never see it — it just makes the plugs fit.
44
Plugging in · connectors

A connector is a key you grant

  • A connector links your assistant to one app — Gmail, Drive, Slack, SharePoint, the calendar — so it can search and act there.
  • You grant each one yourself from the app's connector directory — and revoke it just as easily.
  • Once granted, "find the latest contract in Drive and draft the reply" is one request — no copy-paste shuttle.
  • Under the hood, every plug is MCP (last slide) — why all three assistants can reach the same tools.
your AI assistant you approve each plug not granted — out of reach Gmail Drive Slack CRM calendar it can only reach what you've plugged in
45
Plugging in · claude × office

Claude inside Excel, PowerPoint & Word

  • A side panel inside the real apps — not copy-paste from a chat tab.
  • Excel: builds & fixes models without breaking formulas, tracks every change per cell · PowerPoint: decks in your template · Word: tracked changes · Outlook: beta.
  • Install from the Microsoft add-in marketplace · included in any paid Claude plan (~$20+) · needs a work/school Microsoft 365 account.
change it in Excel… Excel PowerPoint Word Outlook · beta …the deck already knows Claude one conversation

One thread, four apps — the side panel shares a single conversation across all four.

At work, admins can block add-ins — expect this to start with a conversation with IT.

46
Plugging in · files from chat

Real files, straight from the chat

No add-in needed

Claude & ChatGPT both produce genuine .pptx / .xlsx / .docx in the chat: describe it, upload source data, download an editable file — not text to reformat.

Which one when?

Add-ins (last slide) work inside a file you already have. Chat produces a new file. Editing the Q2 model → add-in. First draft of the deck → chat.

"build the Q2 deck" + the results CSV, attached Claude · ChatGPT .pptx — the deck .xlsx — the model .docx — the memo real, editable files — download, open, finish

For many people this is the biggest single upgrade in this talk — and it needs nothing installed. Included from the ~$20 tiers.

47
Plugging in · the chatgpt side

The ChatGPT side of the plug

Connectors

Link to Drive, Gmail, SharePoint, Slack… — it searches your actual documents, not just what you paste.

Excel & Sheets add-in

Official OpenAI add-in: narrates each step, cites cells, asks before writing. Built for financial workflows.

Apps in ChatGPT

Third-party tools inside the conversation — design in Canva, search hotels — without leaving the chat.

Region check: several consumer-tier features have been limited in the EEA/UK; business tiers get broader access. Verify on your tier, in your region, before promising a demo.

48
Plugging in · the built-ins

Or use what's already in the building

Microsoft 365 Copilot

Agent mode in Word, Excel & PowerPoint — works on the document itself. ~$21–30/user/mo on top of M365. Fun fact: the Office agents run on Anthropic's models.

Gemini in Workspace

Now bundled into base plans (from ~$14/user) — Gmail, Docs, Meet. Star feature: =AI("prompt", A2) in Sheets — an AI prompt as a formula, filled down 1,000 rows.

The trade-off vs the add-ins: built-ins are the path of least resistance (and often the easiest compliance story) — the dedicated tools are usually better at the work itself. Prices July 2026.

49
Plugging in · the universal adapter

The universal adapter for everything else

  • No official connector for your tool? Zapier MCP: one connection exposes ~30,000 actions across 8,000+ apps to Claude or ChatGPT.
  • Setup is paste-a-URL — then just say it: "add this lead to the CRM and post a summary to the channel."
  • Make offers the same on its visual canvas. Both need their own (cheap) plan + a paid AI plan.
Expose the minimum

The safety rule: plug in only the actions you need. The AI can only misuse what you've connected — keep the adapter's keyring short.

50
Plugging in · which route

Which route when?

Your situationStart withEntry cost
You live in Microsoft OfficeM365 Copilot or Claude add-ins+$21–30/user · or ~$20/mo
You live in Google WorkspaceGemini — it's already bundledin the plan (~$14/user)
One agent for chat + files + design + OfficeClaude Pro (Cowork · Design · add-ins)~$20/mo
A tool nobody has a connector forZapier MCP / Make~$10–30/mo + AI plan

The pattern everywhere: business tiers carry the admin controls, data guarantees and regional availability. Consumer tiers get features first — workplaces get them safely. July 2026.

51
Close · three things

If you remember three things

1 · It's not a chatbox

The chat window is the smallest part of what exists. Agents do work and hand you files — decks, pages, audio, spreadsheets.

2 · $20 is the magic number

Each lab's ~$20 tier buys a different toolbox. Try what's included — Cowork, Design, Notebook, agent mode — before buying anything niche.

3 · Check the fine print

Before real use: commercial licence? public by default? where does the data go? Free tiers rarely cover work.

52
Summary

What to take away

  • Agents = chatbots with hands: they do the work and hand you files to review.
  • The big three sell different toolboxes for the same ~$20 — Cowork & Design, agent mode, Gemini Notebook & Flow.
  • The wider market has a leader per category — the cheat sheet (slide 37) is your map when names churn.
  • Connectors & MCP plug it all into Excel, PowerPoint, mail — where work actually happens.
  • Everything here: a July 2026 snapshot. Verify before you budget or quote.

Thanks! Questions?

David Budáč