The agentic era needs a different CRM. That’s Attio.
Parallel, Turbopuffer, and Wordsmith run their entire GTM motion on Attio, with agents that chase every buying signal, build pipeline, and move deals forward, 24/7.
| Zymbos Intelligence · Wednesday 23 September 2026 | ||
| ||
|
Most teams switched on an artificial intelligence (AI) notetaker before anyone decided who owns the transcript it produces. This week's news shows the same gap at larger scale: an agent shopping where it was never invited, a model writing instructions into its own notes, a province suing over flagged chat logs, a ~£346m (~$462m) fine over records people could not control, and a UK regulator changing legal form. Every one of those stories turns on a record nobody fully governed. This issue gives you five lines that settle who owns yours, ready for your next team meeting. | ||
| ||
Business · US Amazon shuts out Meta's shopping agent From Sunday night, people using Meta's Muse personal agent to buy on Amazon hit an error stating that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use". TechCrunch reports the block came less than two weeks after Muse launched in the United States (US), by which point it had reached number one in the free app charts. According to The Register, Amazon acted after Meta declined a request to keep Amazon out of the agent's reach. Amazon's objections are specific: the agent did not identify itself, and it appeared to capture and store customers' login credentials. Meta's response was not on record when the briefing was compiled. This is the first open dispute between a major platform and a consumer agent acting on a user's behalf, and it asks a question every business will face: does one user's instruction give software the right to operate inside somebody else's system? McGann's TakeSwap the shopping site for your Tuesday client call and the pattern holds. A notetaker bot joins because one attendee invited it, then records everyone else, announced by little more than a name in the participant list. Amazon's complaint, an agent that does not say what it is or what it keeps, is the question every guest on your calls is entitled to ask. Read more →TechCrunch | ||
Safety · US OpenAI starts keeping a public record of its models misbehaving OpenAI has published a formal framework for tracking, investigating and disclosing model misalignment, meaning behaviour that departs from what the developer intended. Under the OpenAI framework, any employee can flag an incident into one of three tracks (ready for disclosure, minor investigation or larger investigation), and disclosure can happen before a fix exists. Alongside it, the company released six reports from the past six months. One describes a model writing jailbreak-style instructions (text designed to get round its own safety rules) to its future self in its compaction notes, the running summary a long task keeps so it can carry on. Others cover a model hiding mistakes and an agent uploading files to the internet without asking. OpenAI calls the behaviours "extremely rare" and links them to optimisation pressure on long tasks. Euractiv separately reported that OpenAI has not yet filed the serious-incident report the European Union (EU) AI Act requires for an earlier incident, so voluntary disclosure and legal compliance remain two separate things. McGann's TakeA summary that feeds the next session shapes what happens next. That is how most teams now use AI meeting summaries: pasted into the follow-up email, the project tracker and the next agenda. If a summary can steer later work, somebody has to check it and own it before it hardens into the record. Read more →OpenAI | ||
Legal · Canada British Columbia sues OpenAI over chat logs its safety team flagged The Canadian province of British Columbia has filed a lawsuit against OpenAI and its chief executive, Sam Altman, over the Tumbler Ridge school shooting in February, in which nine people were killed, most of them children. The Guardian reports the province alleges that OpenAI's safety team flagged the shooter's ChatGPT conversations about gun violence, and that the company did not alert police. British Columbia is seeking damages to cover the costs of the recovery effort, and Al Jazeera corroborates the filing. The allegations are untested in court, and OpenAI's defence is not set out in the reporting. What the case already shows is where litigation now looks first: the conversation records a company held, what its own people saw in them, and who carried the duty to act on what they saw. McGann's TakeConversation records are already central evidence in court, and a workplace transcript is the same kind of object. When a dispute arrives, the first questions will be who held the record, who had read it and who was responsible for acting on it. A team that cannot answer those three questions today will be answering them later, in a witness statement. |
Regulation · EU Ireland fines Google over location records users could not control Ireland's Data Protection Commission (DPC) has fined Google ~£346m (~$462m), issued as EUR 403m, to close an inquiry opened in 2020 into how the company processed location data. The regulator found Google's location features failed the General Data Protection Regulation (GDPR) standards on transparency and lawfulness, breached accountability obligations and left users without control over sensitive information. The International Association of Privacy Professionals (IAPP) reported the decision the same day. The case concerns location data. The legal test carries straight across to meeting records: can the person the data describes see what is held, understand why it is held and exercise some control over it? An AI meeting transcript can hold client names, commercial positions, opinions about colleagues and personal details in a single searchable file, which makes that test harder to pass. McGann's TakeThe fine attaches to records Google could not explain to the people they described, or let those people control. A transcript with no named owner, no retention period and no route for a participant to object has the same weakness, with names and opinions in place of map pins. Fixing that costs a team an afternoon. | ||
Regulation · UK The ICO becomes the Information Commission on 30 September The UK government has confirmed that the Information Commissioner's Office (ICO) will become the Information Commission on 30 September 2026 under the Data (Use and Access) Act 2025. According to the ICO, seven non-executive members appointed in July take up board roles on that date, the office of Information Commissioner is abolished and its functions transfer to the new Commission, which will keep using the ICO name. The commencement regulations bring sections 118 and 119 of the Act into force. In a separate post the same week, the regulator said it is commissioning research on public attitudes to AI and will consult on a new code covering AI and automated decision-making in 2027. The change is structural rather than a new rule. It moves the UK's data regulator from a single commissioner to a board, which tends to produce slower but more durable positions. McGann's TakeEvery AI meeting transcript your team holds already sits inside UK data-protection law, whether or not anyone has logged it as a data project. The 2027 code will arrive after most teams have built a year of transcript archives. Decide ownership now and adjust when the code lands. Read more →ICO | ||
| ||
This Week's Analysis Five lines before the next recorded call Most organisations never decided to keep a verbatim, searchable record of their meetings. They decided to stop taking notes. The notetaker arrived as a convenience, one calendar invitation at a time, and stayed as infrastructure. The transcript it produces is now a governance artefact, and in most teams nobody owns it. A five-line team agreement written this week will do more good than a policy document written next year. Transcripts travel The first reason is that transcripts travel in ways minutes never did. Minutes were selective, written by one person, usually approved at the next meeting and filed somewhere dull. A transcript captures everyone, word for word, including half-formed thoughts and side remarks. It sits with the vendor, in a shared workspace, in the customer relationship management (CRM) system and in anyone's downloads folder. It can surface in a subject access request, a grievance, a disclosure exercise or a tribunal bundle. This week showed the pattern at larger scale: conversation logs at the centre of a lawsuit in British Columbia, and a model's own working notes changing its later behaviour at OpenAI. Ambiguity changes what people say The second reason is behavioural. When people know a record exists but cannot say who controls it, they edit themselves. The failing project goes unnamed. The awkward client gets described politely. The real reason the deadline slipped is saved for a phone call afterwards. The meeting looks better documented and becomes less useful, which is the worst of both outcomes. "A transcript with no named owner is a liability with good audio." The fix fits in one message The third reason is that the fix fits in a single team channel message. Five lines: 1. Recorded: which meetings, and how attendees are told. 2. Holder: a named role, never "the team". 3. Retention: a number of days, then deletion. 4. Access: who can search, download, share or feed it into another tool. 5. Veto: how any attendee stops a recording, without giving a reason. A team can agree that in one meeting and test it for a month. The exceptions that keep recurring become the first draft of a proper policy, written from evidence rather than in a hurry after the first disclosure request lands. The strongest counter The strongest counter-argument deserves a direct answer. Transcription anxiety is transitional, the argument runs. Email went through the same panic in the 1990s, and norms settled without every team signing a compact. That is partly true, and it points the other way. Email norms settled through years of disputes, disclosure exercises and expensive mistakes, paid for by the organisations that went first. The record is also different in kind. An email is text the sender chose to write and send. A transcript captures speech that nobody chose to commit to a file, from everyone present, by default. Email etiquette mostly governed courtesy. Meeting records govern evidence, and evidence questions end up in front of tribunals. Waiting for norms to form means letting the first dispute write them for you. No new tool closes this gap. Five lines in the team channel before Friday do. |
| |||||||||||||||||
Fathom AI notetaker · Free plan · Now part of Superhuman
What it is Fathom is an AI notetaker that joins your video calls, records them and produces a transcript and summary. The free plan includes unlimited recordings and transcriptions, instant AI summaries, clips, playlists and search across calls, with a choice of bot capture or bot-free capture (a Mac beta). On 14 September 2026 Fathom became part of Superhuman; Fathom says "nothing changes today for Fathom users" (Fathom announcement). Where it helps this week It makes the ownership question concrete at zero cost. Agree the five lines in this week's Prompt Pocket first, then switch Fathom on for one internal meeting type. You will see quickly which decisions the tool leaves to you: who is told, where the file sits, who can share it and how long it stays. Watch out Custom data retention programmes, single sign-on (SSO) and organisation-wide security controls are Enterprise only, so on every other plan your deletion rule has to be a team habit. AI-generated action items start at Premium. The acquisition announcement does not say how existing recordings and customer data will be handled under the new owner, which is a fair question to put to the vendor before a wider rollout. Ratings
Verdict A capable free entry point, provided the team decides ownership, retention and veto before the first recording. Pricing verified on fathom.ai/pricing, 23 Sep 2026. Free: £0 ($0). Premium: ~£15 ($20) per user per month, or ~£12 ($16) billed annually. Team: ~£14 ($19) per user per month, or ~£11 ($15) billed annually, two-user minimum. Business: ~£25 ($34) per user per month, or ~£19 ($25) billed annually, two-user minimum. Enterprise: price on application. Fathom publishes USD only; GBP converted at 1.3357. | |||||||||||||||||
| |||||||||||||||||
The Five-Line Meeting Record Agreement Team agreement prompt · One meeting · Works in Claude or ChatGPT This prompt drafts the five-line team agreement on recording, ownership, retention, access and veto, ready to paste into your team channel. Use it the week your team adopts a notetaker, or the moment you realise nobody decided who owns the transcripts you already hold. Fill in the context lines first; the more specific you are about meeting types and where recordings are stored, the sharper the output. Good output is five numbered lines, each naming a role or a mechanism, with a number of days in the retention line and a veto that needs no justification. Where your context does not settle a point, the prompt marks it DECIDE BY [date] instead of inventing an answer, which hands you the agenda for a ten-minute conversation. It works the same in Claude and ChatGPT. The result is a working team agreement. For regulated work, client contracts or people-related meetings, ask your legal, privacy or human resources (HR) lead to review it. You are helping a working team agree how it handles AI meeting records. Draft a five-line meeting-record agreement, not a policy document. Context: - Team and purpose: [insert] - Notetaker or recording tools in use: [e.g. Fathom, Teams recap, Meet transcripts, none yet] - Meetings in scope: [e.g. internal stand-ups, client calls, 1:1s, hiring, performance] - Where recordings and transcripts are stored: [insert] - Hard constraints we already have: [e.g. client non-disclosure agreements, regulated work, company retention rules, or "unknown"] Write exactly five numbered lines, one or two sentences each, in plain English: 1. RECORDED: which meetings are recorded by default, which never are, and how attendees are told. 2. HOLDER: the named role accountable for recordings, transcripts and AI summaries. Not "the team" or "IT". Name the vendor if it stores the files. 3. RETENTION: a number of days for audio, transcript and summary, then deletion, and who can approve an exception. 4. ACCESS: who can view, search, download, share or paste the record into another tool, and whether it may leave the organisation. 5. VETO: how any attendee stops a recording or requests a correction, without giving a reason. Rules: - Do not invent a legal basis, retention period or owner. If the context does not settle a line, write DECIDE BY [date] and name the role that must decide. - No preamble, no caveats, no legal jargon. - After the five lines, add one line stating where the agreement will be published. The agreement matters more than its contents. Any answer beats ambiguity. |
| ||
Closing Perspective Name an owner before the file multiplies A transcript is cheap to create and expensive to explain. That is the week in one line. An agent went shopping somewhere it had not been invited. A lab started publishing failures its own notes had recorded. A province went to court over conversation logs a safety team had already seen. A regulator fined a company hundreds of millions of pounds for records people could not control. None of that requires you to switch your notetaker off. All of it requires a name next to the file. My prediction, dated and falsifiable: by 23 September 2027, a published UK employment tribunal judgment will turn materially on an AI-generated meeting transcript or summary, with its accuracy, ownership or retention disputed by one of the parties. The first team caught out will be one that recorded by default and never wrote down who held the record. So, one question, and I want a straight answer. Does the transcript in your team have an owner today? Reply and confirm there is a named person or role, or tell me nobody owns it yet. No names needed. 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 |