Insights

A founder's week is mostly conversations: customer calls, investor meetings, field interviews, team check-ins. Most of what gets said there is lost by Friday. AI note-takers fixed the capture part some time ago. The real gain in 2026 comes from what happens next, once the notes are connected to the rest of your tools.

Bot or no bot

The tools record in one of two ways. Some send a visible bot that joins the call as a participant. Others record directly on your computer, with nothing added to the meeting. The second kind also works for calls outside the usual video apps, and some cover in-person meetings from a phone.

Prices are the ones listed in early October 2026. They move fast, so check before you buy. For a founder working across French and English, multilingual support matters more than the cheapest plan. If a client asks for European hosting, a French tool like Noota answers that question directly.

The gain is in the connection

A transcript on its own is one more file. Connected to an AI assistant, your calendar and your email, it becomes a production line. The connection now runs through MCP, the open protocol created by Anthropic in late 2024, adopted by OpenAI and Google in 2025, and managed since December 2025 by a foundation under the Linux Foundation.

• After a client call. The assistant reads the note and drafts the summary, the follow-up email and the next steps, in your format.
• After a sales call. It updates the prospect in your CRM: status, needs, date of the next follow-up.
• After a round of customer interviews. It groups every interview and pulls out the pain points, the exact words people used and the reasons they would buy. For a company still looking for proof of market, this is the most useful one.
• Before a meeting. It prepares a short brief: who you are meeting, what was said last time, what is still open.

Two more features are often missed. Scheduled tasks run on their own, like a Friday summary of the week's meetings and pending follow-ups. Reusable instructions turn one good format, a client summary or a diagnostic, into something every deliverable follows.

Where agents stand in 2026

AI has moved from answering to acting. It reads your tools, chains tasks and runs on a schedule. Microsoft's 2026 Work Trend Index, a survey of 20,000 workers in 10 countries published in May, found that 58% of AI users produce work they couldn't have done a year earlier. It also found that 65% fear falling behind if they don't adapt.

What works well today: research and synthesis, document processing, and recurring reporting. What stays fragile: autonomous decisions with high stakes, and strategy. The output there is usually correct but rarely different from what anyone else would get. The simple rule: the agent prepares, you decide. Checking the quality of what comes out is now a skill in its own right.

Keep the stack small

A small core beats fifteen subscriptions: one assistant, one note-taker, one automation tool. Then build skills in order of payoff:

1. Connect your tools to the assistant: notes, calendar, email. About an hour.
2. Write down your methods as reusable instructions, so every summary and every diagnostic follows the same format. A day or two.
3. Schedule recurring tasks: weekly summary, follow-ups, a watch on funding calls. Half a day.
4. Check the output, every time. Ongoing.
5. Automate between apps with a tool like Make or n8n, for example from a website form to your CRM. One to two weeks.

The first three happen inside the tool you already use. That's why they pay back first. These time estimates are rough.

In the first three days of September 2026, Anthropic, Google and OpenAI each shipped a new flagship model. Claude Fable 5.1 on September 1. Gemini 3.8 Flash on September 2. GPT-6 Astra on September 3. The same week, Nvidia agreed to buy Hugging Face. I went through the release notes and the independent comparisons with one question in mind: what should a founder actually do differently?

The top is converging, the price isn't

Fable 5.1 and GPT-6 Astra list at exactly the same price: $10 per million input tokens and $50 per million output tokens. Gemini 3.8 Flash costs $0.75 and $3.75, roughly thirteen times less, at an introductory rate that runs until December 31, 2026 and then doubles. On the independent indexes, the leaders are a few points apart. On cost, they're an order of magnitude apart.

Between the two most expensive models, the real difference now hides in the details. Cached input costs $0.25 per million tokens on Fable 5.1 and $1 on Astra, and some models finish the same task with far fewer tokens. For an agent that rereads the same long context all day, that line weighs more on the bill than the headline price.

Build so you can switch

Prices and rankings now move every few weeks: Google alone released three versions of its Flash tier in about six weeks. A company that hard-wires one vendor into its product inherits that vendor's pricing and roadmap. The safer design keeps the model behind a thin layer you control. It exposes your own tools through an open standard such as MCP, so the next model can use them without a rebuild.

Open source just got a very large owner

Nvidia agreed to acquire Hugging Face for about $12.9 billion, announced on September 3. Hugging Face is where most open models are published and shared. Nvidia has committed to keep the platform open and not to require its own hardware. Worth watching rather than worrying about. If your product depends on open weights, keep your own copy of the models you rely on, and track their licenses.

The EU rule that already applies

Since August 2, 2026, Article 50 of the EU AI Act requires that people be told when they're interacting with an AI system, unless it's obvious. The Digital Omnibus pushed the high-risk obligations back to December 2027, but it left this one in place. Fines can reach €15 million or 3% of worldwide annual revenue. If your product has a chatbot, an assistant or an agent that talks to users in Europe, the disclosure should already be in the interface. The assistant on this site presents itself as an AI assistant for exactly that reason.

Most founders I talk to want to automate something. The tools are no longer the obstacle: a workflow tool, a model API and an afternoon are enough to connect a form to an email sequence. The real question is what to automate first, and what to keep manual on purpose.

Three places to automate, three very different returns

In a small company, automation usually lands in one of three places: finding customers, turning visitors into contacts, or delivering the work. They aren't worth the same.

• Acquisition. An asset that already exists, a guide, a demo, a calculator, that nobody receives unless they ask for it. Wiring a form, an automatic send and a short follow-up sequence takes a day or two and starts paying back the same week.
• Prospecting. Real upside, real risk. A bad first message sent to your whole network can't be unsent.
• Delivery. The most tempting and usually the least ready. You can't standardize a process you have run four times for four very different clients.

My order is almost always the same: acquisition first, prospecting second, delivery last.

Don't industrialize a process you have done four times

An agent amplifies whatever process you give it. If the process is still changing, you get the average of a few special cases, produced faster, and you spend more time correcting the output than you saved. A reasonable threshold is eight to ten comparable runs before encoding anything. Before that, the useful work is lighter. A structured intake form. A folder with your best past deliverables. And one reusable set of instructions for the single step that never changes.

What an agent actually is

A chatbot answers. An agent pursues a goal. It decides on an action, calls a tool and reads the result. Then it loops until the job is done or something tells it to stop. The building blocks are simple: instructions, a few tools, a loop, some memory, stop conditions, and a short list of things it must never do. Start with the smallest version that works, add tools one at a time, and test each tool on its own before connecting it.

Test it before you trust it

If an AI system talks to your prospects under your name, it needs a test set. Thirty questions, ten of them traps, each with the answer you expect. Rerun them after every change to the instructions. Add a monthly spending cap on the API key and a rate limit on the endpoint. Two settings, and they're what separates a tool you can rely on from a demo.

Prospecting: protect the network

In France, business-to-business email prospecting without prior consent is allowed under four conditions. The address is professional. The message relates to the person's role. The sender is clearly identified. And unsubscribing is easy. You should also be able to say where each contact came from. Beyond the law, the practical rules are simple. Segment before sending and clean the list. Test on fifty contacts before five hundred. Warm up a new sending domain gradually. And open with something useful rather than an offer.

One line I wouldn't cross: tools that automate LinkedIn invitations and messages break the platform's terms and put the account at risk. For a founder whose credibility partly lives on that profile, the downside is far larger than the gain. Automate the writing, not the sending.

I read the Stanford AI Index 2026 the way I read most large reports: looking for the two or three things that actually change how I'd advise someone. Here's what stuck.

The central finding

AI capability is advancing faster than the systems built to evaluate, govern and integrate it. That's the through-line of the whole report. AI is already inside the economy, education, medicine and research. The mechanisms for measuring and controlling it are lagging behind.

For anyone building or advising a company, that gap is where the opportunity and the risk both sit.

Capability isn't reliability

This distinction runs through the entire report and it's the one most people skip. A model can be exceptional on benchmarks and still fail badly in a real workflow.

The questions that actually matter when you deploy:

• Can it solve the problem?
• Can it solve it consistently?
• Can it use the tools it needs correctly?
• Can you verify the output?
• What happens when it's wrong?

This is what people mean by "jagged intelligence": remarkable performance on some tasks, surprising failures on others. Benchmarks don't capture it.

Agents are the next real shift

The report tracks a fast progression in systems that don't just answer but execute sequences of actions. The difference is structural:

Classic LLM	Agent
Question → answer	Goal → plan → tools → actions → verification → result

Performance on environments like OSWorld is improving quickly but remains imperfect. Agents still fail on a meaningful share of tasks. So the near future probably isn't full autonomy. It's supervised agents doing narrow, well-defined work where failure is cheap and verifiable.

What this means for a founder

If you're building on AI, the moat is drifting away from the model and toward the system. Your proprietary data, your context engineering, your evaluation layer, your feedback loops. Anyone can call the same API. Not everyone can build the surrounding infrastructure that makes the output trustworthy.

Where the value actually moved

The interesting shift isn't that models got smarter. It's that the cost of turning an idea into something functional collapsed. A year ago, wiring up a scored, deduplicated CRM with an outreach layer would have meant hiring a developer for months. I built mine in an afternoon and I maintain it myself. That compression, from "I need a tool that does X" to having it, is the real story. Most founders are still pricing their roadmaps as if it had not happened.

What it means for a deeptech company

Two things. First, the busywork around fundraising and GTM, the drafts, the trackers, the first-pass analysis, is now cheap enough that no founder should be drowning in it. Second, and less comfortable: if a task can be fully automated, it's no longer a moat. The durable value moves to the parts that can't. The judgment about which market to enter. The relationship that unlocks a round. The read on whether an investor is actually interested. That's where I spend a founder's time and mine.

The B2B go-to-market playbook has changed more in two years than in the previous decade. New terms appear almost weekly (AEO, context engineering, MCP, FDE, token economics) and each one signals a real shift in how B2B companies acquire and grow customers.

This is my synthesis of Kyle Poyar's 2026 field guide, filtered through what I actually see working with founders.

Pick one motion

The six B2B GTM motions:

• Inbound, content that generates leads
• Outbound, email, social, phone prospecting
• Product-led (PLG), self-onboarding, fast time-to-value, viral loops
• Account-based (ABM), high-value accounts, personalized campaigns
• Paid digital, media buying for attention
• Partnerships / ecosystem, distribution through other companies

The rule that matters: focus on one primary motion at a time. Excelling at several simultaneously is close to impossible, and most early-stage teams that try end up mediocre at all of them.

The AI layer is now infrastructure, not a feature

GTM teams are building AI into the operating layer: enrichment, signal detection, routing, sequencing, forecasting. The teams that do this well aren't using AI to write more emails. They're using it to know which twenty accounts are worth contacting this week.

Signals over volume

Outbound based on volume is dying. Outbound based on signals (hiring patterns, funding events, tech stack changes, product usage) is what still works. The infrastructure to detect those signals is now cheap enough that a small team can run it.

Most rebranding projects start in the wrong place: new logo, new colors, new website, new tagline. The design work is visible, satisfying and fast. It's also usually premature.

A distinctive brand is a strategic system before it's a visual one. It defines a clear position in the market, a specific way of expressing that position, and a coherent experience across every touchpoint.

Start with the gap

The first step isn't creative. It's listening. You gather perceptions from customers, prospects, founders, sales, marketing, product, support, HR, sometimes partners and investors.

What you're looking for is the distance between two things:

• What the company thinks it is
• What the market thinks it actually is

That gap is frequently large. A company sees itself as "innovative, premium, close to its customers." The market sees it as expensive and slow to respond. Both descriptions can be honest. Only one drives buying decisions.

Foundations before expression

The methodology splits into two phases. First, define the foundations: understand the business, the market, the customers, and what should genuinely make it different. Second, turn those foundations into reality: deploy the brand internally, in communications, and across the full customer experience.

Skipping the first phase is why so many rebrands look better and change nothing.

RevOps unifies the three revenue-generating functions: marketing, sales and customer success. Historically these worked in silos, each with its own tools, metrics and view of the customer. RevOps removes those walls and builds one continuous system from first marketing touch through renewal and expansion.

The stated objective is straightforward: make growth fluid, repeatable and predictable by removing operational friction.

The four pillars

• Process and alignment, harmonized customer journey, clear lead statuses (MQL, SQL, PQL), inter-team SLAs, routing rules, handoff management
• Tools and stack, CRM administration and integration (Salesforce, HubSpot), prospecting automation, conversational intelligence, billing
• Data and intelligence, a single source of truth, data warehouse modeling (Snowflake, BigQuery), predictive reporting and forecasting
• Enablement and strategy, team training, admin simplification, compensation plan design, change management

Why the fractional model works

A Head of RevOps costs €100k to €150k per year internally. Senior profiles are rare. So startups and scale-ups increasingly use fractional consultants: operational from week one, budget flexible, and carrying benchmarks from dozens of other companies.

Market rates, France and Europe, 2026

Level	Daily rate	Monthly retainer
Confirmed (2-5 yrs)	€500 – €650	~€2,000 (2 days/month)
Senior (5-8 yrs)	€650 – €850	~€3,500 (1 day/week)
Expert / Lead (8+ yrs)	€850 – €1,200	€6,000 – €8,000 (2-3 days/week)

Some of the most useful entrepreneurial lessons don't come from launching a company. They come from working on problems where the answer isn't obvious and assumptions have to be tested.

Reviewing papers for a finance conference in Nice turned out to be one of those.

What people say versus what people do

One paper I reviewed, Sustainable Investing in the Changing Environment, analyzed 13,396 Finnish individual investors. A striking finding: 73% of respondents reported not knowing about the European SFDR regulation.

But the number wasn't the interesting part. The review process raised a more fundamental question: how much can we really infer from what people say they do?

The study relied heavily on survey responses. The natural next step would be comparing those answers against actual portfolio and transaction data.

Don't confuse interest with demand

The strongest validation is behavioural:

• Did they actually use it?
• Did they pay?
• Did they come back?
• Did they change an existing process to accommodate it?
• Did they allocate budget or headcount?

Customer interviews are useful for understanding why. Usage, payment and retention are the evidence for whether.

Statistical significance isn't business significance

The second lesson from the same review: with a large enough sample, small effects become statistically significant. But a statistically significant relationship isn't necessarily economically meaningful.

The same trap exists in startups. A 2% conversion lift can be statistically real and commercially irrelevant. The question is never just "is this effect real?" It's "is this effect large enough to change a decision?"