In partnership with
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
| Zymbos Intelligence · Wednesday 22 July 2026 | ||
|
||
|
|
Most artificial intelligence (AI) pilots overrun because they start before the conditions are set. KPMG's data this week shows agentic AI adoption is already widespread across UK organisations, and Gartner confirms the budget is arriving, with AI platform spending forecast to grow 63% in 2026. Yet scaling beyond pilots stays rare. The five stories below share one through-line: the organisations getting value narrow the scope, settle permissions and governance before kickoff, and measure from day one. The ones that stall discover those conditions mid-flight, one sign-off at a time.
|
|
|
UK · Enterprise Adoption
KPMG finds agentic AI everywhere, and stuck at pilot stage
KPMG's Make AI Scale report, released this week through techUK, finds adoption of agentic AI across UK organisations is already broad, while moving from pilot to production demands changes to operating models, governance and data access that most organisations have not made. The report's author, KPMG's Paul Henninger, frames the blockers as fear, focus and friction, and argues scaling is an organisational redesign problem rather than a tool rollout. It is the clearest evidence this week that the bottleneck has moved from what the models can do to whether the organisation is ready to use them. For anyone planning a pilot this quarter the finding is oddly encouraging. If the constraint is organisational rather than technical, it sits inside your control, and most of the work happens before the first line of code is written.
McGann's TakePilots stall because organisations are not ready, not because the models are not. That is a solvable problem, and the solution starts before kickoff.
Read more at techUK →
|
|
Global · Market Forecast
Gartner puts AI platform spending up 63%, and CFOs split on why
Gartner forecasts worldwide end-user spending on AI models and platforms will reach approximately £47 billion ($64 billion) in 2026, up 63.4% from around £29 billion ($39 billion) in 2025, with generative AI model spending growing faster still. A companion survey of 204 finance leaders found 45% of chief financial officers (CFOs) say their AI investments lean toward productivity gains, while only 20% prioritise decision quality. The forecast largely predates the week's open-weight price pressure, so rising spend against falling unit prices means volume is doing the work. The CFO split is the number worth sitting with. Money aimed at productivity with no decision-quality target buys activity, not evidence, and activity is exactly what a stalled pilot produces for eight months before anyone asks whether it worked.
McGann's TakeSpending is not the constraint and never was. A pilot funded without a decision-quality target has no way to prove it earned its budget.
Read more at Gartner →
|
|
UK · Governance
techUK publishes a framework for scaling agentic AI
techUK published its industry brief on scaling the responsible adoption of agentic AI on 21 July, setting out how UK organisations can move from isolated pilots to enterprise-wide deployment. It names four interdependent success factors: technical readiness, organisational structures, workforce preparedness and governance frameworks. Read alongside the KPMG data, it looks less like guidance to file and more like a pre-kickoff checklist. Each factor it lists is a condition that is cheaper to fix before a pilot starts than to discover halfway through. The brief lands the same week the UK moved AI to cabinet level, so the governance signal is arriving from industry and government at once. UK organisations that work through the four factors before they begin are not being cautious. They are being fast.
McGann's TakeTreat this as a checklist to run before kickoff, not hygiene to bolt on after deployment. Every item on it is cheaper to fix before the budget is committed.
Read more at techUK →
|
|
Security · Autonomous Agents
A Hugging Face breach was run end-to-end by an AI agent
Hugging Face disclosed a security incident in which an intrusion into internal infrastructure was driven end-to-end by an autonomous AI agent, reaching internal datasets and service credentials, with no evidence of tampering with public models. The detail that should worry every security team came next. When the company investigated, its usual AI tools refused parts of the analysis, because safety filters on the large commercial models blocked the security questions its own defenders needed to ask. The team switched to an open-weight model it could run in-house. Stratechery framed the resulting guardrail asymmetry, attackers using unrestricted models while defenders are blocked by their own, as a structural problem for Western cyber defence. Agent-speed attacks are no longer theoretical, and your incident response may be constrained in ways your attacker's is not.
McGann's TakeThis is the conditions argument applied to security. If your approved tooling has never been tested against an agent-speed attack, that gap is a planning failure waiting to surface at the worst moment.
Read more at Hugging Face →
|
|
US · Defensive AI
Google ships defensive AI as a gated government pilot
Google DeepMind launched Gemini 3.5 Flash Cyber on 21 July, a lightweight model fine-tuned to find, validate and patch software vulnerabilities, released first as a limited-access pilot for governments and trusted partners through its CodeMender programme. Google says the model beats larger general models on the CyberGym benchmark and on scanning Chrome commits, though those are vendor-reported results and independent validation will lag while access stays gated. The shape of the rollout is the point. One narrow use case, controlled access, measured claims, and a pilot frame even where the use case is plainly operational. It lands days after the Hugging Face incident sharpened demand for defensive tools that will actually engage with offensive material, and it reads as an early move in a race to give defenders purpose-built models rather than guardrail-limited general ones.
McGann's TakeWhen the company with the most compute on earth still ships this as a gated pilot, copy the method. Narrow the use case, control the access, measure the outcome before you scale.
Read more at Google DeepMind →
|
|
|
This Week's Analysis
The eight-month pilot is a choice
The median enterprise AI pilot in 2026 runs for eight months. The top quartile delivers in three. Nothing about the technology explains that gap. The eight-month figure is a planning artefact, the cost of starting before the conditions are set, and this week's KPMG Make AI Scale report says the same thing from the other direction: adoption is broad, scaling is rare, and the blockers sit in the organisation rather than the model. Four conditions, set before kickoff
The first is data permissions, scoped before kickoff rather than discovered after it. Slow pilots begin technical work, then hit a governance process nobody mapped, and every week after becomes a waiting room for legal and security sign-off. Fast pilots make signed data access a pre-condition of project approval. No agreement, no kickoff. The second is a quantified success metric, locked before day one. A pilot chartered to "improve knowledge management" has no exit criteria, so it runs until someone loses patience. A pilot chartered to cut average query resolution from twelve minutes to four by week eight can stop the moment the evidence lands, in either direction. A pilot without a metric is an expensive conversation. The pilot was not slow. The conditions were not set.
The third is scope. Horizontal capability pilots need consensus from every function and end up owned by none of them. The top quartile gives the pilot one workflow, end-to-end, with one named accountable owner. Ownership compresses time because decisions stop travelling. The fourth is cadence. Weekly steering, one decision item per session. A monthly committee turns every blocked decision into a four-week delay, and AI pilots generate blocked decisions in clusters, technical, legal and operational arriving together. Weekly review forces the calls while they are still cheap. The strongest counter-argument deserves naming. In regulated industries the data permissions step alone can take months, and no amount of process discipline erases that. Correct, and it sharpens the point rather than blunting it. If permissions are the long pole, they belong at the front of the schedule, resolved in parallel while the build runs on synthetic or first-party data, never discovered mid-pilot after half the budget is spent. The top quartile in regulated sectors does not wait. It sequences. So the recommendation is one page long. Before any AI pilot budget moves, require a conditions sheet: data access signed, metric quantified, workflow isolated with a named owner, steering set to weekly. Each line needs an owner, a date and a decision rule. If any line is blank, the check has not delayed the pilot. The pilot was never ready to start. |
|
|
Glean
Enterprise Search · Knowledge Grounding · Permissions-Aware
What it is
Glean is an enterprise search assistant that grounds AI answers in your organisation's own files, chats and tickets. It fits knowledge workers, project leads and operations heads who need governed answers from internal systems without waiting on engineering. Its strength is not that it writes prose. It is that the answer cites documents you can open and check. Where it fits
This issue's argument is that the strongest 2026 pilots are knowledge-grounding pilots, and Glean is the category leader for exactly that. Native connectors span Google Drive, Microsoft 365, Slack, Jira, Confluence, SharePoint, GitHub and Salesforce, so a knowledge-grounding pilot can own one workflow end-to-end. Access is permission-aware: a user cannot surface content they could not already open, which keeps the data-permissions condition intact rather than working around it. Watch-outs
Deployment still needs IT involvement to wire up connectors and confirm permissions, so day-one setup is not self-serve. And pricing is opaque. Glean publishes no rates on its site, so a buyer cannot size the cost without a sales call. Grounding quality also depends on how clean and well-permissioned your underlying content already is. Ratings
Verdict
Glean is strong on exactly the conditions that decide a knowledge-grounding pilot: permission-aware grounding in sources your team already uses. The one material drag is pricing opacity, which turns a build-versus-buy decision into a sales call. Scope it as a single end-to-end workflow, and settle the pricing before you settle the plan. Pricing checked on the Glean site as at 22 July 2026: no pricing is published; all tiers are enterprise quote-based via demo request. |
|
|
Pilot Pre-Mortem
Pilot Planning · Governance · Works in Claude or ChatGPT
Use this before the pilot starts, not after it stalls. Describe the pilot you are about to run and it returns the four upstream conditions that decide its speed, the order to fix them, and the single one most likely to sink it. It is this issue's argument made executable. One tactical note on reading the output: treat the answer as a gate, not a suggestion. If it surfaces a condition you cannot put an owner and a date against today, that is the pilot telling you its real start date. Run it again after each condition closes; the ranking should move, and if it does not, the sponsor conversation has not happened yet.
You are an AI pilot governance advisor. I am about to start an AI pilot and I want a pre-mortem before any technical work begins. Assume the pilot will overrun unless the upstream conditions are fixed first.
Here is the pilot: [describe what it does, which team it serves, what data it needs, what success looks like, and the target timeline]. Assess it against these four conditions: 1. Data permissions: is access to every required data source scoped, approved and signed before kickoff? 2. Success metrics: is there one quantified outcome with a number, a baseline and a deadline? 3. Workflow ownership: does the pilot own one workflow end-to-end, with one named accountable owner? 4. Steering cadence: is there a weekly steering meeting with one decision item per session? Return: 1. For each condition: what "ready" looks like for this specific pilot, what usually breaks it, and one question to ask the business sponsor to expose hidden risk. 2. The order to fix the four, with a one-line rationale for the sequence. 3. The single condition most likely to sink this pilot if unresolved, in one sentence. 4. A start or delay recommendation, plus the one thing that would change it. Be specific to the pilot described. No generic advice. The most useful question this prompt asks back is the sponsor question under data permissions. Hidden vetoes live there, so find out who can quietly stop the pilot before you find out mid-flight.
|
|
|
When the sandbox doesn't hold
OpenAI's models broke out of their own safety test the same week Hugging Face's defenders were blocked by their own guardrails. My read on whether anyone was actually driving, and the blast-radius question for every team already running agents. Pairs with story 4 above.
Read on LinkedIn →
Most AI rollouts fail on Day 7
The day nobody owns, when the questions start and then stop. The 90-day AI Team Enablement Roadmap gives every stage a date, an owner and a decision rule. The rollout side of this issue's conditions argument.
Read on LinkedIn →
Meta stepped back. Your voice didn't.
Meta pulled its image feature after the backlash but left the audio reuse opt-out running. What to switch off, and why defaults buried several taps deep are a design choice. The permissions condition, playing out in public.
Read on LinkedIn →
|
|
|
Closing Perspective
The conditions become the KPI
The KPMG and Gartner numbers point the same way. Spending is accelerating, adoption is broad, and the conditions are the constraint. So two predictions, both checkable. First, by 31 December 2026, knowledge-grounding pilots will overtake content-generation pilots as the most common enterprise AI use case. The value sitting in what organisations already know is larger than the value of generating more words, and tools like Glean are the infrastructure for that shift. Second, by 31 January 2028, at least one FTSE 250 firm will report "pilot to production" time as a named board-level key performance indicator (KPI). The pressure to show return on investment (ROI) from AI budgets is too high for open-ended experimentation to survive another cycle. Both dates are specific enough to check, and I will mark them against reality here when they arrive. If you ran the pilot pre-mortem on a pilot this quarter, hit reply and tell me which condition was furthest from ready. I read every response. John McGann
Founder, Zymbos AI |
|
Zymbos Intelligence
zymbos.ai
You're receiving this because you subscribed at zymbos.ai
© 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |