In partnership with
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
| Zymbos Intelligence · Wednesday 20 May 2026 | ||
|
||
|
|
Most AI procurement in 2026 looks like 2010 software-as-a-service procurement at peak fragmentation. This week's intelligence puts numbers on why, and the editorial below sets out the four questions every buyer should ask their next AI vendor.
|
|
|
Pricing · Vendors
AI's all-you-can-eat era ended on 14 May
Anthropic announced that on 15 June it will split Claude subscriptions into two pools. Interactive usage (a human at the keyboard) stays inside the existing limits. Automated and agentic usage (coding tools and harnesses that route round-the-clock workloads through Claude) moves to a separate credit meter. OpenAI countered within hours: two months of free Codex for any developer who switches, plus a one-click migration tool that moves prompts, skills and Model Context Protocol (MCP) configurations across. Anthropic re-countered the same day with a 50 percent capacity boost on Claude Code valid through 13 July. Three days earlier, GitHub announced all Copilot plans move to usage-based billing on 1 June. The flat-fee era at the three largest AI coding vendors ends inside the next thirty days. Uber's Chief Technology Officer (CTO) Praveen Neppalli Naga said the company burned through the entire 2026 AI budget by April. ServiceNow disclosed the same problem. Goldman Sachs surveys put enterprise AI overruns at orders of magnitude. McGann's TakeAnything you contracted for last quarter is being repriced this quarter. Audit your AI bills before your Chief Financial Officer does, because the July renewal cycle will be the first one that bites.
Read more in State of Brand →
·
Read more on the GitHub Blog →
|
|
Governance · Deployment
Sinch: 74 percent of AI customer-service rollouts get rolled back
The AI Production Paradox study from Sinch surveyed more than 2,500 AI decision-makers and found that 74 percent of live AI customer-service deployments have been rolled back or shut down. The figure rises to 81 percent at firms the report classifies as having mature governance guardrails. The headline is striking; the more useful finding is the inversion. Better-governed organisations roll back more, not because the technology is worse there, but because they detect failure faster. Where governance is absent, broken AI quietly continues serving customers and the rollback never happens because the failure was never escalated. The study is single-vendor-funded and should be read for direction rather than magnitude, but a corroborating Register piece from the same day reports executives now value human workers more after deploying AI, an unusual reversal of the productivity narrative. McGann's TakeA 74 percent rollback rate is not a tool problem, it is a buying problem. You want to be in the 81 percent that caught the failure, not the 26 percent that never noticed it was failing.
Read more in The Register: rollback study →
·
Read more in The Register: executives value human workers less →
|
|
Regulation · UK
European Commission opens consultation on draft high-risk AI classification
The European Commission released draft guidelines on 19 May that operationalise the high-risk classification regime under the Artificial Intelligence (AI) Act. The text gives providers, deployers and competent authorities a workable test for deciding whether a system falls into the Annex III high-risk band. Systems that do are subject to conformity assessment, post-market monitoring, and human-oversight obligations. Consultation runs to 23 June 2026 and the final guidelines are expected in the second half of 2026. This is the first major non-general-purpose AI (GPAI) rulemaking instrument since the Act entered enforcement. United Kingdom firms selling into the European Union are in scope, regardless of where they are headquartered. Industry coalitions including DigitalEurope are expected to push back on scope; civil-society groups will argue the carve-outs are too broad. Watch the consultation register. McGann's TakeIf you sell anything into the European Union, treat the Annex III work as homework, not background reading. The consultation closes 23 June and the final guidelines will land before most firms have staffed the workstream. Now is the cheapest moment to run the classification dry-run.
Read more on the European Commission draft guidelines →
·
Read more on the EU AI Act regulatory framework →
|
|
Vendors · Market Share
Anthropic overtakes OpenAI on business AI spend share for the first time
The May Ramp AI Index places Anthropic at 34.4 percent of business AI spend in April, against OpenAI at 32.3 percent. It is the first time Anthropic has crossed OpenAI on this metric. Claude Code is named as the leading driver of the shift. Ramp's index measures spend across Ramp's own corporate-card customer base, which skews toward United States technology firms and the middle market, so the figure should be read as a trend signal rather than a global market share. The Microsoft and OpenAI camps dispute the methodology. Three structural risks to the lead remain: pricing pressure from open-weight challengers (DeepSeek, Qwen, Mistral), OpenAI's enterprise sales push following the Dell partnership announced 18 May, and unit economics that neither side has yet shown to be profitable. McGann's TakeTwo-point-one percentage points sounds small until you remember it moved in a single quarter. The next twelve months will reset vendor share faster than most enterprise contracts can react. Multi-year exclusivity clauses are now a luxury, not a default.
Read more in VentureBeat →
|
|
Strategy · Procurement
CIO names the next divide: AI owners against AI renters
A CIO essay published 19 May frames the strategic choice underneath the procurement noise: firms that run AI on controlled infrastructure with proprietary intelligence loops, against firms that run AI on someone else's platform. Renting moves faster. A paid seat is live in minutes, a vendor handles the upgrades, the buyer can switch providers when the contract permits. Owning is slower and more expensive at the start but compounds over time: predictable cost, data sovereignty, and a learning loop the vendor cannot remove. The piece is opinionated rather than empirical, but it lands the question every Chief Information Officer (CIO) will have to answer in 2027 budget cycles: which AI workloads are core enough to own, and which are not. McGann's TakeFor most firms the honest answer is renter, and that is fine. The unforgivable position is renting without knowing you are renting, because that is the position with the worst price discovery. Name your core workloads, own those, and rent everything else with eyes open.
Read more in CIO →
|
|
|
This Week's Analysis
The Four-Question Vendor Cut-Through
Vendor decks rhyme. Every AI tool claims to be the most accurate, the most secure, the most integrated, the most enterprise-ready. The most useful thing in a vendor evaluation is to ignore the deck and ask four questions. The four questions
Question one. Where does the model run, and who has access to the prompts and outputs? The answer reveals whether the data leaves the buyer's boundary, whether the vendor is reading it, and whether those terms can change unilaterally. Answers vary by tier and by region. Press for written confirmation, not the slide. Question two. When the underlying foundation models change next quarter, what changes in this product, and who decides? The answer reveals whether the buyer is paying for a product or a pass-through interface, and how much of the price is value-add against simple markup. The 14 May pricing war is the live example. Anthropic and OpenAI did not change the underlying models. They changed the billing. The product on the user's desk did not change. The contract did. Question three. If we cancel in 18 months, what comes with us? The answer reveals lock-in. Prompt libraries, fine-tuned models, accumulated context, evaluation data: these are the assets that should belong to the buyer, not the vendor. Most vendors will hand back transcripts and structured data but withhold the fine-tunes and the embedding indexes. Ask which specifically come with you, in what format, by what date. Question four. What specifically have you stopped doing for customers in the last 12 months? The answer reveals product discipline. Vendors who can name something they removed are vendors who can prioritise. Vendors who claim no removals are vendors with feature debt. The four questions take 20 minutes per vendor call. They will surface differences a year of feature comparisons will not.
The strongest counter-argument is that mature vendors will rehearse these answers. Some will. Rehearsed answers reveal preparation; unrehearsed answers reveal candour; both are useful signals. The questions still work. The recommendation: pick three vendors currently in your AI stack and run the four questions on each before your next procurement review. The vendor that fails question two is the vendor you most need to renegotiate. |
|
|
Apollo.io
Prospecting · Enrichment · Outreach
What it is
Apollo.io is an end-to-end platform spanning outbound prospecting, inbound qualification, data enrichment, and deal execution, powered by a business-to-business (B2B) contact database the vendor cites at 270 million plus. Native customer relationship management (CRM) integrations with Salesforce and HubSpot let a workflow automation engine chain research, enrichment, and outreach without leaving the tool. A Chrome extension and the Apollo AI assistant complete the stack. What it does badly
Data quality varies by region: European and Asia-Pacific records are noticeably weaker than North American ones, and several user reports flag stale phone numbers outside the United States. Feature breadth can also overwhelm new users; expect roughly two weeks of learning curve before the workflow engine starts paying back the cost. Fit for this issue's theme
For a procurement-themed issue, Apollo sits awkwardly. It is a tool you would buy through the exact procurement maze the editorial describes, and it does pass the four-question test: model location declared, foundation-model dependency limited, exportable contact data, recent feature retirements published. If you sit in revenue operations or sales leadership, it earns the eight. If you sit elsewhere, the editorial framework matters more than this specific review. Ratings
Verdict
Strong end-to-end platform for revenue operations and sales teams. Regional data variance and feature density are real, but neither breaks the case. An 8.0 is the top of the defensible band. Some links are affiliate. Zymbos AI may earn a small commission at no extra cost to you. |
|
|
Run the four-question vendor cut-through
Procurement · Vendor evaluation · Works in Claude or ChatGPT
A 20-minute exercise for procurement leads and budget owners. Open Claude or ChatGPT, paste the prompt below, and feed it the vendor's website, sales deck, and most recent contract. Out will come a structured assessment against the four questions, plus the specific follow-ups to put back to the vendor on the next call.
You are a procurement adviser helping me evaluate an AI vendor. I will share the vendor's website, sales deck, and current contract (if I have it). Your job is to assess them against four questions, then give me a follow-up question list for the vendor.
For each of the four questions below, give me: - The vendor's likely answer based on the material I share. - Your confidence in that inference (high, medium, low). - The specific follow-up question I should put to the vendor. - A red flag, if any. The four questions: 1. Where does the model run, and who has access to the prompts and outputs? 2. When the underlying foundation models change next quarter, what changes in this product, and who decides? 3. If we cancel in 18 months, what comes with us? Specifically: prompt libraries, fine-tunes, accumulated context, evaluation data, in what format, by what date. 4. What has the vendor specifically stopped doing for customers in the last 12 months? End with a single-page summary: vendor strengths, weaknesses, the three follow-up questions to send before the next call, and one recommendation (proceed, renegotiate, or walk). |
|
|
Closing Perspective
Who would notice if your vendor doubled the price?
The firms winning AI procurement in 2026 have already accepted that vendor relationships will not look the same in six months as they did at signing. The 14 May pricing war is the warning shot, not the end state. Three predictions for the next two quarters. First, at least two of the major coding-tool vendors will rename or repackage their enterprise tier before October, because the current branding implies a flat-fee deal they can no longer deliver. Second, the median enterprise will sign at least one usage-cap clause into a new AI vendor contract by the fourth quarter, because the alternative is the Uber surprise. Third, AI ownership in the procurement function will become a named role at FTSE 250 scale before the second quarter of 2027. The question for your team this week: who would notice if your largest AI vendor doubled the price tomorrow? If the honest answer is nobody, that is the role you are missing. If you ran the four questions on one of your live vendors after reading this, hit reply and tell me what came back. I read everything. John McGann
Founder, Zymbos AI |
|
Zymbos Intelligence
zymbos.ai
You are receiving this because you subscribed at zymbos.ai
© 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |