Pebbles Ai - AI sales platform for modern B2B teams

@omri_ben_shoham1 You're absolutely right. Most of these tools overlap, and the category is crowded with point solutions that do one or two things very well.

But none of them focus on reasoning. They only focus on execution.

People think outbound is a sending problem. It's usually a reasoning problem, and it breaks into two formulas most tools never touch.

The first is the sequencing itself. A message that actually lands isn't a template with a {first_name} slot. It's at least 5 disciplines working together:

• Persuasion science: the established principles of influence and buyer psychology (reciprocity, social proof, commitment) and knowing which to use at which moment

• Communication science: how the message is structured and paced to be understood. Clarity, framing, timing, and the netiquette that makes it read like a human wrote it

• Hyper-personalisation: real profiling of who this person is and what they care about, so you connect on a human level despite being total strangers. Not a dumb variable field like job title

• Heuristics: the judgment rules for each situation. When to push, when to nurture, when to stay quiet, distilled from the combined wisdom of top operators

• Lexical semantics: precise word choice and sentence construction, because the same point lands or dies on phrasing and tone

Miss one and the sequence underperforms. Which is why so much outbound gets 1% reply rates and blames the list.

But here's the part almost nobody does. All of that only works if the upstream formula fired first. Before a single message goes out, you need:

1. Geolocation (e.g. Greater London)

2. Industry (e.g. financial services)

3. Persona cluster (ICP and archetypes)

4. Common denominators (historical data)

5. Market trends (the ones that actually move you)

You can do all of it on the platform. Great sequencing on bad upstream work is just a beautifully written message, sent to the wrong person, with the wrong value proposition.

That's the whole reason we built Pebbles as one neurosymbolic engine, not another point tool. It reasons over both dimensions: strategy upstream, execution downstream.

And this is the real problem with the point-solution pile. Those tools were built for a pre-AI world. They aren't AI-native, nor domain-specific, so they physically can't reason at this level.

There's a deeper issue too. Dig into the research and you'll find most GTM point solutions are built exclusively by engineers with no background in strategy, marketing, or sales.

So ask yourself: how do you solve something you've never lived? It's like our tream trying to fix a problem in corporate law. We don't know how paralegals work with associates, how associates work with solicitors, how solicitors work with partners, how partners work with the board, and how each of them handles clients through completely separate workflows. Hand on heart, personally I wouldn't know where to start.

Don't get me wrong, these engineers are brilliant. Seriously smart, and they've built genuinely great products. But each one covers one or two stages of a funnel that has at least ten.

Because go-to-market is no different from building software. Teamwork is everything. Back-end works with front-end, front-end with the product designer, everyone with QA and DevOps, and a solution architect sits across all of it. Engineering has Jira for exactly this.

Nobody had built the Jira for commercial teams. We're the first. It's called the Go-To-Market Operating System, and it could only exist with extraordinary intelligence beneath it.

Think of it like iOS. Each app (feature) has its own tech stack, its own intelligence, its own knowledge base, its own mapped-out GTM science. It's like having multiple startups in one place. And every app talks to the others.

Point solutions can only bolt on an "AI feature," which is usually just a free ChatGPT or Gemini model in a sexy trench coat (UX). Execution works, so you gain speed.

But it doesn't move the needle on pipeline, revenue, or margin. You and your team just do busywork faster. Three months later you look back, and nothing material has changed on the financials.

The meme says it perfectly. You think the fix is one more point solution. It isn't. It's more busywork, plus a fresh learning curve for yet another tool that most people never fully learn.

That's exactly the problem Pebbles solves, with a completely new approach.

@nhanife56159 This made my day, thank you! The tab-switching was the exact thing that drove me mad as a founder, so hearing that the strategy-to-outreach flow cut it down for you is the best feedback we could ask for.

That "one workspace" feeling is the whole reason we built it, so it means a lot that it is received well. If you hit anything confusing or have a wishlist as you go deeper, I'm all ears. We're shaping it around real users and we ship fast.

@tarqiya_forgah Great question! Let me break it down for you.

ChatGPT: replaced completely.

For GTM specifically, this is the cleanest swap. Generic AI gives you fluent text with zero memory of your company, your ICP, or your positioning, so you spend more time briefing it than it spends adding tangible value back to you.

Pebbles' Marketing Assistant and Strategy Assistant cover it, and actually make an impact on your pipeline, revenue, and margins. Not to mention it gives you energy back, because you don't need to learn so many tools or keep so many tabs open that never inform each other.

Clay and Apollo: largely absorbed.

This is where the patchwork usually collapses into one flow. Clay's real value is its enrichment waterfall and AI research, and Apollo's is its contact database plus sequencing.

Both run natively inside Pebbles, on our own multi-provider enrichment waterfall, feeding straight into reasoned email and LinkedIn sequences.

So the sourcing, the enrichment, and the outreach stop being 3 tools you stitch together by hand. It's all connected now.

On top of that, we use specialised profiling that reads each prospect from both a marketing angle and a sales angle. Pebbles doesn't just hand you leads. It tells you who they actually are, and recommends how to build a relationship with them.

HubSpot: keep it.

Your CRM is your system of record, and it should stay that way. It's wired into your reporting, your deal tracking, and your support.

We are not asking you to rip out the ledger. Pebbles sits on top as the reasoning and execution layer, and the CRM stays the source of truth. We'll be integrating with HubSpot, Notion, Salesforce, and Pipedrive soon.

Cost

There's a cost angle too. Those four as a combined stack, across seats, is not cheap. We priced Pebbles to come in under Clay alone at every tier, and that's before you count the three or four other subscriptions it folds in.

So the short version: keep the CRM as your record, retire the patchwork on top of it. We'll have integrations soon with Notion, HubSpot, Salesforce, and Pipedrive.

@shivangit26 Love that you want the detail. Let me lay it out step by step.

How we read deliverability

We do not publish a sender score. Our proxy is simpler, and we think more honest: replies of any kind. A yes, a blunt no, even an unsubscribe, all of them prove one thing.

The email reached a human inbox. Silence is the sound a dead domain makes in the spam folder. So the volume of real human responses is our live read on domain health, and it is the number we actually trust.

Our best performing account

A 3-person SaaS startup, three founders running their own outbound, no SDR. Before Pebbles they were paying for HubSpot, Apollo and Sales Navigator, with a rotating cast of ChatGPT, Claude and Grok on the writing.

The setup: 5 domains, a few 100 hyper personalised emails per domain, capped there on purpose. Pebbles is built for a few hundred hyper-effective personalised emails, not a volume blast, so the domain never takes the reputation hit you described. By design.

No cliff to fall off if you never climb the volume ladder. Doest hat make sense?

The result, and treat this as a ceiling rather than a promise since it is our strongest account: a 42% reply rate, with 23% of everyone contacted replying with genuine interest. By our own proxy, response at that level means the domains are healthy and the mail is landing.

Positve reply rate means that the variables within the formula are correct:

Y (outbound success) = Strategy × Positioning × Targeting × Value Proposition × Offer × Netiquette × Communication Sciences × Persuasion × Persona-centric Writing

Why it holds at that volume

Two reasons. Volume stays capped and every email is hyper personalised. And the word and pattern avoidance updates continuously from our signals into Google, Microsoft and Yahoo, so the copy stays on the right side of the filters instead of tripping them.

Besides, we live in a trust economy now. Quality over quantity. We see the evidence across all channels within Earned, Owned and Paid.

Great question Patrick!!! Love when technologists go deep.

For small B2B teams, full autonomy is mostly a pitch-deck fantasy today, and the reason is arithmetic, not opinion.

Agent errors compound multiplicatively across steps. At 95% reliability per step, a 10-step task lands around 60% overall (0.95^10). Even at an optimistic 99% per step, a 20-step chain only succeeds ~82% of the time.

Empirically it's worse: in Carnegie Mellon's TheAgentCompany benchmark, even the strongest model completed only ~30% of realistic, multi-step office tasks end-to-end (earlier runs landed at 24%).

And the market is pricing it in. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing unclear value and inadequate risk controls, and notes today's models can't reliably follow nuanced instructions over long horizons.

People are frustrated, and frankly, when I talk to customers, angry!

So the winning pattern isn't more autonomy. It's selective autonomy: aggressive automation on low-stakes steps, hard human gates on high-stakes ones, with escalation paths that cap the blast radius. Which is precisely the handoff quality you're describing.

You framed these four around the sales agent, but they're really the design spec for the whole platform, strategy, marketing and sales alike. So here's how each one works at that level:

What changed?
Whatever the surface, a market shift the strategy core detects, a competitor move, a new buying signal, an inbound reply, the platform isolates the delta and leads with it. You get the one thing that moved and why it matters to YOU, not a raw feed you have to reverse-engineer. Same behaviour whether it's your positioning or your pipeline.

What evidence supports it?
Every recommendation carries its reasoning, a strategic call, a beachhead segment, a line of outbound copy. This is where neurosymbolic earns its place over a raw LLM: the symbolic layer keeps each inference traceable back to the rule or data point that produced it, so a conclusion arrives with its "why" attached rather than as an unexplained assertion from a black box. Inspectable reasoning like that is what the research points to as the precondition for trustworthy human-in-the-loop oversight.

What needs human judgment?
Whether it's a strategy decision, a targeting call, or an outbound message, when confidence is low (we have "judges") or the call is genuinely ambiguous the platform surfaces that explicitly and routes it to you, instead of papering over uncertainty with a confident-sounding guess. Overconfidence in failure is one of the documented ways agents break, so we treat "I'm not sure, your call" as a feature, not a bug.

What should never be sent without approval?
Anything irreversible or externally facing, a published campaign, an outbound send, an investor or sales asset going out the door, is gated by default. The platform can research, reason, write, sequence and prepare the whole thing, but the irreversible action stays with a human until you explicitly choose to loosen it.

Notice the common thread: every one of those four depends on reasoning you can actually trust, not a confident guess. That trust is the exact thing the arithmetic up top says nobody has yet, it's why 40% of these projects get canceled. So the real question is whether our engine is any different. Here are the numbers.

How does our neurosymbolic actually perform?

A base LLM is inherently a chatbot system. It predicts the most probable next word, brilliantly, but that is not enough for complex domains such as B2B GTM, Corporate Law, and Human Resources. Nothing in it stops to ask "is this true, and does it obey the domain rules".

Neurosymbolic AI is 3-step systems working together, with a check between them:

• The neural half reads language and context, the way any strong LLM does

• The symbolic half applies explicit logic, rules and a structured knowledge base

• A verification step sits between that and the output, catching anything that breaks

So you get the fluency of an LLM with a reasoning and fact checking layer bolted underneath. Outputs are accurate, reproducible and explainable (even auditable) rather than a confident black box.

It is also rare: academic interest went from 112 papers in 2015 and 2016 to over 9,000 in 2025 and 2026 [Google Scholar], yet real production systems are almost nowhere, because building one needs machine learning, formal logic, knowledge engineering and domain science in the same room at once.

Let's look at some numbers. First, let's compare Claude Opus on the Max tier with a single instruction against the full Pebbles pipeline:

• Accuracy: 33% vs 87%

• Precision: 57% vs 91%

• Sales Efficacy, MQL to SQL: 15% vs 85%

And at the architecture level:

• 82% lower error rate than LoRA fine tuning, because the architecture is structurally accurate rather than nudged

• 3x better gross margin than wrappers, because the reasoning is not rederived from scratch on every call

• ~2% hallucination on rule bound queries, versus 31.4% across real world use HalluScore benchmark

But don't take my word for it: Claude Opus sits around 33% factual hallucination on the public HalluScore benchmark, while generally neurosymbolic methods approach ~100% accuracy on rule based tasks [arXiv 2502.01657]. We are closing up at 85-98%.

Cost is where it gets almost silly. To rebuild one reply with a raw model:

• Around 12 prompts per reply, each re-sending 25,000 to 35,000 tokens of context

• Roughly $5 to $7 per usable output, and that's not even top 1% reasoning

• About $5,000 to $7,000 a month at that volume, before 500 hours of human prompting

THIS MEANS WE CAN GIVE MUCH MORE AI ALLOWANCE TO OUR USERS!!! 🫶🏻❤️📈

Finally, the neurosymbolic reasoning is what lets it carry complex, multi-faceted, and cross-functional B2B work all the way through. These are examples you can build and execute on, which is impossible with base-models or tools with wrappers:

• A full go to market strategy, grounded in your ICP, positioning and live market signals, then turned into the campaigns that run it

• A beachhead strategy, picking the wedge segment worth attacking first, sizing it, and sequencing the entry instead of guessing

• Industry trend analysis read across macro, meso and micro signals, so you see the shift before it hits your pipeline

• Investor decks and enterprise sales assets, two pagers, RFPs and proposals that hold up when a sharp reader pushes on them

• An omnichannel outbound engine, from fresh leads to reasoned email and LinkedIn sequences to replies captured and qualified in Smartbox, built and run end to end

And this is only the current stage. We are making the first steps toward a true Jarvis for go to market: a system that can safely, securely and reliably run the work fully autonomously, with no human in the loop.

I hope that makes sense.

@michael_shollenberger Thanks man, and great questions!!! From what I can see, surface level, you're building something similar just for a different domain.

Firstly, a well-prompted base model LLM (e.g. Claude Opus Max) writes an acceptable message. Fluent, personalised on surface data. If writing quality and personalisation were the bar, you wouldn't need us. And everyone would be driving Maseratis 😁.

But it requires fundamentally more to commercially move your company. Here is our formula:
Y (success) = (Geo × Industry × Persona × Common Denominators × Market Trends) × (Persuasion × Communication × Personalisation × Heuristics × Lexical Semantics)

Allow me to show you live on a call, concretely, the difference between our generation of strategy, campaigns, content, marketing materials, and sales assets, and that of a base model.

Would you like that?

@nikitaeverywhere The main interface today is the web app. We also have desktop apps that live in your dock on iOS (mac) and Windows, so it sits right alongside your other daily tools rather than buried in a tab.

No installation required. Super light.

On Claude specifically: we don't integrate with Claude, and that's on purpose. Pebbles runs its own reasoning models, our cores, which are purpose-built for GTM rather than general-purpose.

MCPing Claude would mean inheriting a general model's guesswork for rule-bound, domain-specific complex work, which is exactly what we set out to avoid.

Instead, we make switching effortless. Think of it like changing mobile providers and keeping your number. The system does the heavy lifting, and you're moved over in about 3 minutes without losing your most up-to-date company info.

You don't rebuild your progress that you've carefully built within Claude (or ChatGPT, Gemini, etc.). You just move it over to Pebbles Ai within 3 minutes, and run it on much smarter models.

Lastly, every customer gets a custom API that allows to connect the Pebbles Ai neurosymbolic brain trained exclusively on your company to your existing techstack.

Think of it as the primary brain of your existing techstack. You can wire it into your website to answer inbound inquiries, power the chatbot on your site, and plug into the other automation tools you already run.

In short, you get both a sophisticated multi-purpose platform + a connectable brain that makes your other tools smarter too.

Hope that make sense.

@jasnoor_singh_oberoi There's good news. The £1,499 isn't the entry point. That's the Organisation tier, built for a 10-seat commercial team, often for lower-mid to mid-market organisations.

The way in is Pro, at £49 a month. Even better, the Team subscription at £349 a month has the most value. The most bang for your buck.

We wanted to make the value stupidly high. Like a NO-BRAINER. Something that would absolutely flabbergast new users. How is it possible to add so much value at such a low price.

(Spoiler alert: it's possible because our system is built on neurosymbolic AI)

So the exact person you're describing, the solo founder buried in fragmented tools with no sales team, isn't an afterthought for us. They're arguably who we built this for first. Don't forget, we were also in that position.

In our experience, solving for smaller teams has proven to be much more difficult than for larger teams, ironically. So we did it the hard way first. Naturally haha.

On the positioning itself, you're right. When we say "enterprise-grade" or "McKinsey-level," we mean the quality of the reasoning and judgment, not the size of the customer or the price of the plan. We should be more clear.

Our mission is David versus Goliath. Give a two-person startup the same GTM firepower a big company pays top-tier consultants and 30 premium tools for, at a price a founder can actually afford.

In short, you only scale up when your team does.

Best way to judge it is to run it yourself against the mechanic and overwhelming tool fragmentation problem you described.

Start free, no credit card needed. I'd recommend trialing the Team subscription.

You won't regret it. 😊

@omri_ben_shoham1 Short answer: the upstream reasoning works on day 1. It does not need to sit and watch you for a month first.

Here's why. The intelligence isn't starting from zero and waiting to learn from your behaviour. It already carries the encoded GTM knowledge, the strategy models, the persona logic, the market-signal reads. Y

And telling it who you are takes about 3 minutes. Think of it like switching phone providers. You don't rebuild the phone number and data, you just port your number across and it works.

Onboarding walks you through your company, your product, and your ICP, and the cores light up immediately with that context.

So it's not "use it for three weeks and it might get good." It's already INSANELY SMART in the first session, and it compounds from there.

  @seffal_abdelaziz Great question, and yes, it can. If a campaign is underperforming, you can take it into the Strategy Assistant, use the Audit workflow, and reason through exactly that: is the weak link the ICP, the segment, the offer, the messaging, or the channel. Or, something else.

The engine can work through those variables with you and pinpoint where it's breaking.

What's coming (now in development) is the automated version. We're building a feature that runs that whole diagnosis for you, isolating which variable failed across the full formula:

Y = (Geo × Industry × Persona × Common Denominators × Market Trends) × (Persuasion × Communication × Personalisation × Heuristics × Lexical Semantics)

Because it's multiplication, not addition, one weak input collapses the whole result. So the diagnosis isn't "the campaign was bad," it's "this specific factor was the zero."

The system pinpoints the exact variable, then tells you how to fix it.

That's a VERY complex analysis, tracing a poor result back through every input to find the one that broke, and it's a reasoning problem the neurosymbolic engine is built for.

Short version: today you can diagnose it with you Strategy Assistant. Soon it'll do the full root-cause breakdown automatically and hand you the fix.