The Week AI Companies Started Selling Deals, Not Software

Week of May 4–11, 2026 | Issue 5 | For smart professionals catching up on AI

This week

Two of the biggest AI companies in the world stopped acting like product companies this week and started acting like consultancies. Anthropic signed a $1.5 billion deal with Blackstone and Goldman Sachs to embed Claude directly inside mid-market firms. OpenAI launched a $10 billion joint venture called The Deployment Company with TPG that guarantees investors 17.5% annual returns. Meta did the opposite and pulled back from open weights entirely with a new assistant called Muse Spark. And the EU quietly pushed its AI Act deadline out by a full 16 months. If your company has been "figuring out AI" for a year, this is the week vendors started selling you a bigger answer.

🗞️ News you should know

1. Anthropic and OpenAI turned themselves into consulting firms overnight

Anthropic signed a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to put Claude inside the daily work of mid-market firms — think regional banks, 500-person law firms, 1000-person insurance brokerages. The same week, OpenAI unveiled a $10 billion JV called The Deployment Company, anchored by TPG and 19 other investors, structured as a partnership that guarantees investors a 17.5% annual return. Neither is a product launch — both are vehicles for a salesperson to show up at your office and re-sell your company AI for a quarterly retainer. What this means for you: if you work at a firm with 200 to 5,000 employees, expect a procurement conversation in the next 90 days about "our Claude partnership" or "our OpenAI deployment." Ask your manager who's going to own it. The answer is usually a title no one held six months ago.

Source: techcrunch.com | Published: 2026-05-04 Corroboration: fortune.com

2. Meta killed its open-weights strategy and named the replacement Muse Spark

Meta launched Muse Spark, a fully proprietary AI assistant now rolling out inside WhatsApp, Instagram, Facebook, and Messenger. No open weights, no downloadable model, no Llama-style research community release. Investors rewarded the pivot: Meta stock jumped 9% on the announcement. The reversal matters because Meta spent three years telling developers that open weights would win — and then switched the moment it had a product to sell. Bonus trouble: the Authors Guild, Scott Turow, and a group of publishers filed suit the same week over the copyrighted books Llama was trained on. What this means for you: if you've been waiting for "free, open Meta AI" to catch up with ChatGPT, stop waiting. The open-weights era from the biggest lab is over. If your company has been making "we can self-host" a condition of AI adoption, your list of credible vendors got shorter this week.

Source: cnbc.com | Published: 2026-05-07 Corroboration: cbsnews.com

3. The EU AI Act got kicked down the road by 16 months

The European Commission confirmed on May 7 that the compliance deadline for high-risk AI systems has moved from August 2, 2026 to December 2, 2027 — a 16-month delay. The official reason was "industry readiness." The unofficial one was a lobbying win for American hyperscalers and European manufacturers. The law kept its hardest clauses — "nudifier" apps that undress real people are now a hard ban across the bloc — but the rest of the compliance machinery is on ice. What this means for you: if your company spent the last six months building an AI Act compliance checklist, it's not wasted work, but the clock moved. If you were about to hire an AI compliance officer for July 2026, that role lost its urgency for a year. If you sell software to European enterprises, your buyers have 16 more months to keep putting off the AI governance line item in their RFPs.

Source: reuters.com | Published: 2026-05-07 Corroboration: theregister.com

4. Claude released 10 named agents for bankers, insurers, and loan officers

Anthropic released 10 purpose-built Claude agents for financial-services firms on May 5 — each one tied to a specific daily task. Pitchbook drafts. Credit memos. KYC screening. Earnings-call summaries. Client-meeting prep notes. Each agent is a one-button routine, not a chatbot — you click once and get a first-pass draft in under a minute. Microsoft announced a parallel partnership with Moody's the same day. What this means for you: if you work in banking, insurance, or any regulated financial-services role, this is the week your IT department started getting pitched "AI for your job specifically." If you work anywhere adjacent — accounting, consulting, compliance — the same vertical-agent pattern is coming for you within six months. Start writing down the three tasks you do every week that are 80% template. Those are the first ones that will be automated out of your calendar.

Source: fortune.com | Published: 2026-05-05 Corroboration: finance.yahoo.com

5. OpenAI's new voice features let you talk to any document in any language

OpenAI launched a voice intelligence API on May 7 that folds transcription, translation, and conversational chat into one call. The interesting part for non-developers: the features showed up the same week inside ChatGPT's free mobile app. Point your phone at a PDF and you can now ask it questions out loud in one language and get answers in another. Translation quality on pricing agreements, medical forms, and rental contracts was measurably better than Google Translate in independent tests. What this means for you: if you've avoided ChatGPT's voice mode because the old version was awkward, give it 20 minutes this week. The next time a supplier sends you a contract in a language you half-read, open ChatGPT on your phone, point the camera at page one, and ask "what's the payment term and what happens if I'm late?" You'll get a faster answer than your usual routine.

Source: techcrunch.com | Published: 2026-05-07 Corroboration: arstechnica.com

6. Google quietly rolled Gemini 3.2 Flash into the iOS app — and it costs 40% less

Google pushed Gemini 3.2 Flash into the iOS Gemini app and AI Studio on May 5 with no marketing, no keynote, no press event. The new model runs at $0.25 per million words of input — 40% cheaper than 3.1 Pro — and, in informal tests from people who use these things all day, responds faster on long-document summaries. Google also announced Gemini Agent, a 24/7 digital partner that lives inside Gmail and Google Docs, and expanded Gemini inside Google Home with camera controls. What this means for you: if your company uses Google Workspace, the AI features you were told to wait for are live. Open Google Docs this week, click "Help me write" on something real you're writing, and judge for yourself whether it saves you 15 minutes. If it does, that's the moment to start asking your Workspace admin about the paid Gemini tier.

Source: 9to5google.com | Published: 2026-05-06 Corroboration: arstechnica.com

💡 5 prompts you can paste today

Short, practical prompts you can copy into ChatGPT, Claude, or Gemini free tier — no paid plan needed.

1. Figure out whether your job gets a Claude agent next

When to use: Any time a vendor mentions "vertical AI agents" or your company rolls out an AI pilot. This prompt tells you which of your tasks are at risk and which aren't. Works best in: claude, chatgpt, or gemini

I work as a {job title} at a {company size} {industry} company. My main weekly tasks are:

1. {task 1 — be specific}
2. {task 2}
3. {task 3}
4. {task 4}
5. {task 5}

For each task, tell me:
a) Is this an 80% template task that an AI agent could realistically draft in under a minute in 2026? Score 1–5.
b) What makes it agent-resistant (if anything) — judgment calls, external relationships, data only I have access to?
c) If my job was 30% automated away within a year, which two of these five tasks are the most dangerous to lose?

Keep answers under 50 words per task. No hedging. If you don't have enough information, ask me one clarifying question first.

What you get back: A blunt risk map of your own week. You'll see which tasks to stop volunteering for and which ones are your actual value. Validation: ✅ Claude Opus 4.7 on 2026-05-11 · ⚠️ Other models untested

2. Translate a vendor pitch into a plain procurement question

When to use: When your procurement or IT team forwards you an AI vendor deck and asks what you think. Works best in: claude or chatgpt

I received a sales deck from {vendor name} promising {their headline claim}. Here's the key slide content:

{paste the slide text or summary, or the email thread}

In plain English, give me:
1. The one question that would force them to prove the claim with a number.
2. The one question that would expose the hidden ongoing cost — seats, usage, custom integrations.
3. The one question that tells me whether a competitor (free tier of ChatGPT, Claude, Gemini, or Microsoft Copilot) would get us 80% of the value for 10% of the price.

Frame each question as something I could say in a meeting without sounding hostile. Under 25 words each.

What you get back: Three meeting-ready questions that move the conversation from "cool demo" to "show me the number." Saves you the usual three-week pilot detour. Validation: ✅ Claude Opus 4.7 on 2026-05-11 · ⚠️ Other models untested

3. Turn a contract in a foreign language into your three decisions

When to use: Any time you're handed a supplier agreement, NDA, or rental contract in a language you half-read. Works best in: chatgpt (the new voice + camera feature from item 5 makes this faster than ever on mobile)

Here's a contract I received: {paste text, or upload the PDF}

In English, give me exactly three things:
1. The payment terms — amount, currency, and what happens if I'm late.
2. The termination clause — how either side ends the agreement, and what notice is required.
3. The one clause that's unusual for this kind of document. If there's nothing unusual, say so.

Then list three questions I should ask {the other party} before signing. Each under 20 words.

Don't summarize the whole contract. Don't give me legal advice. Give me the three decisions and three questions.

What you get back: Six lines you read in 90 seconds. You still send it to a lawyer if it matters — but you walk into the call knowing what to ask. Validation: ✅ Claude Opus 4.7 on 2026-05-11 · ⚠️ Other models untested

4. Spot when your AI tool is pretending to know something

When to use: Every time a chat tool gives you a confident answer to something specific — a statistic, a date, a person's title, a quote. Works best in: claude, chatgpt, or gemini

Earlier I asked you {your original question} and you told me {paste the model's answer, or the specific claim you want to check}.

Now, as a separate conversation:
1. What's your confidence level on each specific fact in that answer? Score 1–5.
2. Which facts are you drawing from memorized content vs. inferring from context?
3. If I had to independently verify one thing before quoting this in a meeting, which one is most likely to be wrong?
4. What's one web search I could run in under a minute to check it?

Be honest. I'd rather you say "I'm not sure" than sound confident and be wrong.

What you get back: A self-audit of the earlier answer. Use this especially before quoting an AI-provided statistic in an email, a deck, or a client call. Validation: ✅ Claude Opus 4.7 on 2026-05-11 · ⚠️ Other models untested

5. Write a "here's what I did this week" note in under 90 seconds

When to use: Friday afternoon, when your manager or client wants a weekly update and you haven't started writing. Works best in: claude or chatgpt

Here's my rough week in bullet points:

{paste 8-15 bullets, no structure, whatever's in your head — meetings, calls, documents you sent, numbers that moved, things that slipped}

Turn this into a weekly update email to {manager / client name}. Rules:
- Under 150 words total.
- Three sections: "What got done," "What moved the number," "What's stuck and what I'm doing about it."
- Lead each bullet with a number or a name. No vague "progress on" or "working on" language.
- Honest about what's stuck. If nothing's stuck, say so.
- Sign-off matches how I normally sign emails: {"Cheers, {name}" / "Thanks, {name}" / "Best, {name}"}.

What you get back: A draft you edit in 60 seconds and send. The discipline of "lead each bullet with a number or a name" makes it land better than a wordy update by itself. Validation: ✅ Claude Opus 4.7 on 2026-05-11 · ⚠️ Other models untested

🎯 Use case of the week

How a three-person boutique law firm replaced a $900-per-month document review service with a 20-minute Claude routine

The situation: A three-person contract law firm in Madrid used to outsource first-pass review of supplier agreements to an outside legal-ops vendor at €900 per month. The vendor's turnaround was 48 hours. The firm's clients wanted 24. The partner doing most of the final review was working Saturday mornings to close the gap.

The AI version: The firm now runs the same steps every Monday. Paste the contract into Claude. Ask for the three-decisions summary from prompt #3 above. Flag any clause the model rates unusual. Then do a full human review only on the flagged clauses. Turnaround for clients dropped from 48 hours to 6. Vendor cost went from €900 per month to zero (Claude free tier handles 90% of their volume; they top up with a single €20 Pro subscription for long PDFs). Recovered time: roughly 8 hours per week per partner, or the equivalent of one additional junior associate billable capacity.

The catch: The first three weeks produced two minor misses — Claude initially flagged a late-payment clause as "standard" when it was unusually punitive for their jurisdiction. The partner now reviews every flagged-as-standard clause in 30 seconds instead of trusting the label. Net time saved still stands at 7 hours weekly per partner. Also: they never paste documents that contain client names or identifying details — they redact first. Privacy discipline is non-optional.

How to try it yourself:

1. Open Claude at claude.ai — free tier handles documents up to about 20 pages. For longer, use ChatGPT free on mobile with PDF upload.
2. Redact anything client-identifying in the document first. Find-and-replace "[CLIENT NAME]" is fine.
3. Paste the redacted document and run prompt #3 above. Walk through the three decisions and three questions. Keep the human review for the flagged clauses only.

Source: Reader submission, anonymized.

🛑 Don't pay for this

The claim: There's a new category of "AI Deal Advisory" consultancies launching this week, piggybacking on the Anthropic-Blackstone and OpenAI-TPG news from item 1. Pitches range from "we'll help your board evaluate your AI vendor contract" to "AI procurement strategy sessions" priced at €3,000 to €15,000 per engagement. LinkedIn is full of them.

The reality: Most are two-person shops with a PowerPoint template and a ChatGPT Plus account. The advice they deliver to a €5,000 client is the same three questions any competent procurement lead already asks: what's the real unit cost at our volume, what's the termination clause, what's the vendor's runway. None of this is worth €5,000.

Better alternative: Your CFO and your head of IT, together, in a room, for one hour. Give them prompt #2 above for prep. Have them draft your three questions before any vendor demo. If your company is big enough to need outside help, hire a named procurement consultant with a fixed-price engagement and references — not a two-person "AI deal advisory" with a six-week-old website.

📖 One skill to practice this week

The skill: Naming your "AI unknowns" out loud before every meeting.

Why it matters: The biggest AI-related career risk in 2026 is not that your job gets automated. It's that someone in the room knows more than you about a vendor, a model, or a capability, and you defer on a decision you should have owned. Naming what you don't know — out loud, in advance — flips the dynamic. You stop being the person who got surprised; you're the person who asked the right question.

How to practice (about 45 minutes, spread over the week):

1. Before your next three meetings, spend 60 seconds writing down the three AI-related terms, tools, or vendors you expect to hear. Don't Google them yet.
2. Score yourself 1 to 5 on each. "I could explain this to a smart colleague" is a 5. "I've heard the name" is a 2.
3. For every term at 3 or below, spend 60 seconds in Claude or ChatGPT with this prompt: "Explain {term} in plain English in under 80 words. What's one question I could ask in a meeting that would show I understand it? What's one question that would expose a vendor who's overselling it?"
4. Walk into the meeting and, when the term comes up, ask the exposing question. Not combative — curious. "Interesting — how does {specific implementation question} hold up with a company our size?"
5. Note who in the room didn't have a good answer. That's your signal for who to trust on future AI decisions.

Signs you've got it: You stop leaving meetings thinking "I should Google that later." You start leaving meetings with other people thinking that. Two weeks in, a colleague asks you to join a vendor call specifically because "you ask the right questions." That's the whole skill.