See pricing
What Happens When Al Agents Can Spend Your Money?
The a16z ShowAI

What Happens When Al Agents Can Spend Your Money?

Jean-Denis Greze argues that AI is not stuck because models are too weak. It is stuck because most people are not tinkerers, so the winning products will discover useful work to automate for them, earn trust step by step, and stay flexible as the model layer keeps changing.

October 19, 2026 · 63 min listen · 16 min read · Jean-Denis Greze
0:00
–:––

Context

Jean-Denis Greze, former CTO of Plaid, explains why he shut down a promising AI tax startup and rebuilt Town as an AI assistant for email, calendars, and everyday business workflows. The episode matters to PMs because it argues that the main blocker to AI adoption at work is not model quality. It is discovery, trust, and product design for people who do not have time or skill to invent their own automations. It also gives a clear enterprise AI strategy: start with user delight, earn deeper access connector by connector, and keep the application layer flexible as models and agents keep evolving.

The Big Idea

AI at work will break out when products stop asking users to invent their own automations and start discovering, proposing, and running useful tasks for them. The winning application layer will feel consumer-simple, prove value before asking for sensitive integrations, and stay neutral across model providers as the frontier keeps moving.

Greze's core argument is that intelligence is already good enough for many business tasks. The harder product problem is packaging that intelligence so non-tinkerers can trust it, understand it, and use it in the flow of work.

Key Insights

Good product, wrong market

Town's first product, AI tax prep for SMBs, worked technically and was growing, but Greze killed it because the market mechanics were wrong for venture scale. Year one revenue was a little below $1M, and year two looked like $2.5M to $3M, which he said would have looked strong in a pre-AI company. The problem was structural:

  • Tax is bought only once a year.
  • It creates weak natural word-of-mouth.
  • Incumbents with distribution can add automation to an installed base faster than a startup can win massive share.

The PM lesson is that product quality is not enough. Workflow frequency, distribution loops, and how fast incumbents can catch up all shape whether a market can support breakout growth.

Discovery, not intelligence

Greze's central claim is that many businesses could already get major value from today's models, but most of that value is trapped behind a discovery problem.

  • The winners today are usually tinkerers, people willing to connect tools, write prompts, and keep iterating.
  • A generic chat box nudges people toward "better search," not toward durable automation.
  • The real product job is to show users what AI can do for their specific work, then make the first win almost automatic.

His example was a business renting surfboards and small beach buggies. Someone there used Town to inspect photos after rentals and decide whether to return a deposit. The insight is not that this use case is special. The insight is that many valuable automations already exist, but users often need the product to reveal them.

Email and calendar first

Town started with email and calendar because they are common, high-signal systems for knowledge work and require less initial trust than core systems of record.

That lets an assistant prove value on jobs like:

  • preparing for meetings
  • reminding someone what a meeting is about
  • drafting answers to repetitive emails
  • helping someone start the day prepared

Greze's sequencing matters. Once users feel that early value, the product has earned the right to ask for CRM or other deeper systems. For enterprise customers, that second step then has to come with audit logs, privilege boundaries, and cost visibility.

Consumer UX beats enterprise-first

Greze's Dropbox lesson is that a world-class consumer experience can be adapted for enterprise. The reverse is much harder. If a team starts from compliance and admin requirements, it often bakes in a product experience that never becomes delightful.

That is why Town resisted building for the CISO first. Greze's concern was that if the buyer is not the user, the team can drift into building a product that makes the check-writer comfortable without making the product meaningfully better for the person doing the work. The iPhone versus BlackBerry framing is the point: once leaders feel a much better personal experience, they often push the organization to make that experience work inside the company.

Horizontal first, tailored later

Town did not begin as AI for one vertical. It launched broadly, then used customer behavior and interviews to learn where repeated value was hiding.

Greze described the playbook:

  • Look at aggregate behavior to find power users.
  • Interview a cluster of similar users, such as restaurant owners.
  • Find repeated workflows across that segment.
  • Turn those workflows into starter routines and recommendations.

The restaurant example was supplier price monitoring. A restaurant owner might get monthly spreadsheets or PDFs from wholesalers and rarely compare price changes in a disciplined way. An AI routine can check those changes frequently, find when a supplier got too expensive, and surface savings that matter even if nothing interesting happens in most months. The broader bet is that AI products will win less through static feature lists and more through how precisely they adapt to each user's mix of roles, tools, and repeated jobs.

Neutral apps reduce platform risk

Greze's argument against tying an application to one lab is practical, not ideological.

  • A mid-market or enterprise buyer does not know which model provider will be best over the next 3 to 5 years.
  • The best provider may differ by task, such as coding, voice, slides, or video.
  • A model-tied product can become a quality risk or a pricing risk for the customer.

Town's answer is model neutrality plus data ownership. The business should own the memories, contacts, and context created inside the product, and should be able to move if a better solution appears. Greze also draws a hard line on training philosophy. He does not want a customer's operating context turned into model weights that could help someone else.

Agents shift value capture

Greze thinks agentic software changes where value sits on the internet. If an agent can watch prices, compare options, fill forms, or buy on a user's behalf, then pure supply aggregation becomes less valuable because human attention is no longer the scarce resource it used to be.

Alex Rampell's CamelCamelCamel example makes this concrete. People already use a tool to monitor Amazon prices when the savings are worth the effort. Agents extend that logic to many more people and many more decisions.

Greze adds an important caveat: not all friction is bad.

  • Browsing stores in Soho is friction some people enjoy because discovery is part of the value.
  • Reordering toilet paper is friction most people want removed.

The PM lesson is not "automate everything." It is to decide which friction is part of the experience and which friction is just toil.

Mental Models & Frameworks

Incumbent catch-up clock

A startup's AI advantage has a time limit. The question is not just whether a workflow can be automated. The question is whether the new entrant can build distribution and product pull before incumbents add similar automation to an installed base.

Use this model when evaluating a market:

  • How often do users buy or engage?
  • How much word-of-mouth is built into the workflow?
  • How quickly can incumbents copy the core automation?
  • Can the startup reach scale before that catch-up happens?

Town's tax product failed this test because the buying cycle was annual and the incumbents already owned distribution.

Token allocation machine

Greze said, "I think of our product as a token allocation machine." The idea is simple: the product should keep finding places where cheap model calls unlock meaningful value.

Use it to prioritize AI work:

  • Prefer tasks where token cost is tiny and user value is recurring.
  • Do not obsess over whether a task matters every time. A routine that finds value 1 month out of 12 can still be worth running continuously if the cost is near zero.
  • Focus on tasks humans would never staff manually because the economics were previously too bad.

Meeting prep for every meeting and monthly supplier price checks are both examples.

Earn the next connector

This is a staged trust ladder for enterprise AI products.

  • Start with lower-risk systems, such as email and calendar.
  • Deliver obvious value quickly.
  • Use that value to justify connecting deeper systems, such as CRM.
  • Add governance at each step, including audit logs, permission controls, and spend visibility.

This model helps PMs avoid a common mistake in enterprise AI, asking for maximum access before the product has earned any trust.

From apps to outcomes

Greze argues that "the unit of value in software has changed from the app to the computer." His deeper point is that AI can increasingly do anything a human can do with a computer, so value shifts from app-local features to end-to-end outcomes.

Use this model when scoping product work:

  • Define the job the user wants done, not the screen they must visit.
  • Cross tool boundaries when the user's outcome crosses tool boundaries.
  • Expect the value of isolated app features to compress over time.

This is a useful antidote to building AI features that stay trapped inside one interface while the real user job spans five tools.

Good friction, bad friction

Not every inefficiency should be removed.

  • Bad friction is work the user wants completed with minimal attention, such as price checks or repeat purchases.
  • Good friction is effort that creates discovery, taste, or enjoyment, such as browsing for inspiration.

Use this model before automating a flow. If the friction is part of why people use the product, removing it can make the product worse even if it makes the workflow faster.

Context as business asset

Greze treats the memories, contacts, preferences, and working context created by an AI system as a business asset, not just exhaust data.

That framing changes product decisions:

  • portability matters
  • training policy matters
  • admin control matters
  • switching costs should be explicit, not hidden

For PMs, this is a strong reminder that the memory layer in AI products is part of the core product, not a backend implementation detail.

Decision Principles

Principle: Exit weak markets

When: a product works and revenue is growing, but usage frequency, distribution dynamics, and incumbent advantages make breakout scale unlikely.

Why: good execution cannot fully compensate for a market where the buying cycle is too slow and the catch-up clock is too short.

Principle: Win users first

When: buyer requirements and end-user experience start pulling the roadmap in different directions.

Why: Greze believes it is easier to add enterprise packaging to a product people love than to retrofit delight into a product built mainly for procurement and compliance.

Principle: Earn deeper access

When: an AI product needs more permissions, more systems, or more autonomy to unlock the next level of value.

Why: trust compounds. A product that proves itself on email and calendar is far more likely to win CRM access than a product that asks for everything up front.

Principle: Stay model-neutral

When: the model frontier is changing quickly and customers cannot predict which provider will lead across tasks over time.

Why: model lock-in can become a product quality risk, a pricing risk, and a strategic dependency for the customer.

Trade-offs & Nuance

Consumer polish versus controls

Starting with consumer-grade delight creates stronger user pull, but enterprise customers still need governance. The point is not to ignore security, finance, or compliance. The point is to avoid letting those requirements define the first version of the user experience.

Breadth versus vertical depth

A horizontal product sees more use cases and can learn from a bigger surface area. A vertical product can message more clearly and solve one workflow more deeply from day one. Town is betting that starting broad, then learning segment-specific routines, creates a better long-term flywheel than choosing a vertical too early.

Neutrality versus switching costs

Model neutrality and portable context are attractive to buyers because they reduce dependency on one lab or one app. The trade-off is that this can weaken the vendor's lock-in. Greze is effectively betting that trust and product quality will matter more than artificial switching friction.

Autonomy versus auditability

The more agents act at AI speed, the more companies need guardrails that operate at human trust levels. Greze's examples include audit logs, permission boundaries, and cost controls. Faster autonomy creates more value, but only if the organization can understand and govern what the system is doing.

Efficiency versus enjoyable friction

Agents can remove a huge amount of toil, but they can also flatten experiences people like precisely because they are exploratory or social. PMs should treat automation as a design choice, not a universal good.

Common Mistakes

Mistake: Shipping a blank box

A blank prompt box assumes users already know what to ask for. Most do not. Greze's whole argument is that many users are leaving value on the table because the product makes them invent the workflow themselves.

The better approach is to infer likely jobs from role, tools, and behavior, then present a small set of concrete starting points.

Mistake: Selling to admins

If the roadmap is driven mainly by the CISO, legal team, or procurement team before the user experience is compelling, the team can build something enterprise-safe but user-weak. Greze sees this as one of the fastest ways to miss the real opportunity in AI for work.

Mistake: Locking one model

Picking one model provider too early can quietly hard-code product constraints into the experience. What looks simpler in the short term can become expensive or uncompetitive as different model providers pull ahead on different tasks.

Practical Application

Map inbox and calendar toil

Review two weeks of email and calendar data for one target persona. Tag the repeated jobs that happen around meetings, follow-ups, prep, summaries, scheduling, and status updates.

Those are good candidates for an AI assistant wedge because they are:

  • frequent
  • easy for users to recognize as painful
  • often valuable even before deep integrations

Show tasks, not possibilities

If onboarding still begins with "What would you like AI to do?", rewrite it. Replace the blank box with three to five concrete, role-relevant jobs inferred from the user's connected data.

For each suggestion, show:

  • what the agent will do
  • what data it will use
  • what the output will look like
  • whether it runs once or continuously

This follows Greze's core point that products should reveal useful automation, not wait for users to imagine it.

Interview segment power users

Pick one segment where usage is already strong. Greze mentioned talking to about 10 power users in a segment and looking for the repeated workflows they all share.

Turn the results into:

  • a starter library of routines
  • better onboarding suggestions for that segment
  • clearer product marketing that names the job, not the technology

Build a connector ladder

Define the order in which your product asks for access.

  • Start with the lowest-permission systems that still create visible value.
  • Write down exactly what new value each additional connector unlocks.
  • Pair each deeper connector with the governance features needed to make that step acceptable.

This makes connector strategy a product decision, not just an integration backlog.

Keep context portable

Separate the memory and context layer from the model layer as much as possible.

At minimum, define:

  • who owns stored memories and contacts
  • whether customer data is ever used for training
  • how a customer could export or migrate core context
  • how easily model providers can be swapped per task

In AI products, this architecture becomes part of the customer value proposition.

Questions to Consider

  • Where in your product are users still expected to invent their own automation instead of seeing a concrete job, drawn from inbox, calendar, or system data, already suggested for them?
  • If a CISO or procurement team vanished for one quarter, would your AI product become noticeably more useful for the end user, or is the current roadmap already user-led?
  • Which memories, contacts, operating context, and learned preferences inside your AI product would be painful for a customer to port elsewhere today, and is that pain creating trust or just hidden lock-in?
  • Which parts of your workflow are like reordering toilet paper, pure toil the user wants removed, and which parts are like browsing shops, where the friction is part of the experience?

Bottom Line

Jean-Denis Greze's argument is that AI at work is no longer waiting on smarter models. It is waiting on better products: products that surface the right tasks, earn trust in stages, and keep users flexible as the model layer changes. For PMs, that shifts the job from adding AI features to designing discovery, trust, and adaptive automation around real workflows.

Resources Mentioned

ResourceTypeWhy it was mentioned
TownWebsiteThe product discussed throughout the episode and the main example of a model-neutral AI assistant for work.
Alex Rampell on XSocial profileIncluded in the episode description for listeners who want more of Alex Rampell's thinking on software, platforms, and AI.
Jean-Denis Greze on XSocial profileIncluded in the episode description for listeners who want to follow Jean-Denis Greze and Town more closely.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
TownAI assistant across email, calendar, and business workflowsThe central product in the discussion and the vehicle for Greze's product strategy.
OpenAI and AnthropicFrontier model providersUsed as examples of why application companies may want model neutrality instead of tying themselves to one lab.
CanvaDesign softwareGreze used it as an example of a tool an agent could operate on a user's behalf to produce event materials.
CamelCamelCamelAmazon price tracking serviceAlex Rampell used it to show that users already delegate price watching when savings justify the effort.
MuseMeta's personal agent productDiscussed as an example of how agent monetization and data advantages may work differently from traditional ads.

People to Follow

Jean-Denis Greze

Founder and CEO of Town, and formerly CTO of Plaid. He is worth following for practical thinking on enterprise AI UX, model neutrality, and how assistants can move from chat interfaces to workflow automation.

Alex Rampell

General Partner at a16z and the interviewer in this episode. He is worth following for his writing and investing perspective on software platforms, marketplaces, and how agentic products may reshape commerce.

Experiments to Try

Experiment: Role-based onboarding

  • Hypothesis: new users activate faster when onboarding shows three role-specific automations inferred from email and calendar instead of starting with a blank prompt box.
  • Change: infer role from inbox and calendar signals, then surface suggested routines with example outputs.
  • Test: A/B test suggested-workflow onboarding against the current blank-start experience.
  • Measure: first useful action rate, first week retention, and percentage of users who create or enable an automation.

Experiment: Low-permission enterprise pilot

  • Hypothesis: enterprise teams will connect deeper systems sooner if an AI assistant proves value on email and calendar first.
  • Change: run a pilot that limits the initial setup to email and calendar, then prompt for CRM only after the user has experienced repeated value from prep, follow-up, or summary workflows.
  • Test: compare this staged rollout against a cohort that is asked to connect email, calendar, and CRM near the start.
  • Measure: connector completion rate, active usage after two weeks, number of security objections, and conversion to a deeper deployment.

Notable Quotes

"I think of our product as a token allocation machine." (Jean-Denis Greze)

"You don't need more intelligence. The intelligence is there. It's just how do you deploy it effectively at scale?" (Jean-Denis Greze)

"The unit of value in software has changed from the app to the computer." (Jean-Denis Greze)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free