Why does "AI-powered" no longer get replies?
Buyers in most B2B functions now hear from AI vendors regularly, and many have already tried a general-purpose assistant for the task you are selling. The category word has turned into noise. What still gets attention is a precise description of the work: "drafts first replies to tier-1 tickets inside your help desk" says more than "AI agent for support".
The second change is scrutiny. Teams that rushed into pilots learned to ask harder questions: what happens to our data, how often is it wrong, who checks the output, what does it cost at our volume. Your outbound has to anticipate those questions instead of saving them for the security review.
Who decides to buy an AI product?
Usually more people than for ordinary software, because the product touches company data and changes how a team works. Map the roles before you write a single email:
- The owner of the workflow: head of support, finance operations, RevOps, recruiting or legal operations. They feel the pain and become your champion. Write to them first.
- An AI or innovation lead, in companies that have one. They compare vendors, run pilots and may already have a shortlist.
- IT and security, who ask about data retention, access controls, subprocessors and whether customer data trains any model.
- Legal or compliance, more often in regulated industries or when the product makes decisions about people.
- Finance, especially when pricing is usage-based and the cost at full volume is hard to predict.
Which workflows should an AI company target first?
Not every function buys AI at the same speed. Two things set the pace of a deal: how easily the buyer can check the output, and how much harm a wrong answer does. Start where both are favorable, then move up the risk ladder once you have references and a solid security story. Here is our assessment of common workflows (a judgment call, not survey data):
| Workflow | Typical owner | How the buyer checks the output | Scrutiny to expect |
|---|---|---|---|
| Drafting replies to support tickets | Head of support | An agent reads every draft before it is sent | Moderate: customer data sits in tickets |
| Meeting notes and call summaries | Sales managers, team leads | The attendee reads the summary | Low to moderate: recording consent |
| CRM cleanup and enrichment | RevOps | Spot checks on changed records | Low to moderate |
| Invoice matching and accounts payable | Controller or finance operations | An exceptions queue reviewed by the team | High: payments and audit trails |
| Contract review and redlines | Legal operations or general counsel | A lawyer approves every change | High: confidentiality and privilege |
| Screening job applicants | Head of talent or recruiting | Recruiters review rankings | Very high: rules on automated hiring decisions |
How does an AI deal move from pilot to contract?
Problem conversation
Agree on one workflow and one number that matters: hours per week, tickets handled, documents reviewed or errors caught.
Pilot on their data
A short, time-boxed pilot on a real sample. Define success before it starts, including an acceptable error rate and who reviews the output.
Security and legal review
Expect questions on data storage, retention, training use and model providers. Buyers who follow the NIST AI Risk Management Framework will also ask how you measure and manage risks such as inaccurate output.
Rollout and pricing
Agree how usage will grow and what it will cost at full volume, so finance is not surprised in month three.
What should your trust sheet answer before the first call?
The security review is where many AI deals stall, and most of its questions are predictable. Write a one-page trust sheet before you start outbound, link to it from your second email and keep it current. It should answer, in plain words:
- Where customer data is processed and stored, and in which regions.
- How long inputs and outputs are kept, and how a customer deletes them.
- Whether customer data trains or fine-tunes any model, yours or a provider's. A plain "no", backed by the contract clause that says so, is worth more than a paragraph of reassurance.
- Which subprocessors and model providers handle the data, with a link to a list you keep up to date.
- Who reviews the output before it has an effect, and how a customer configures that review.
- How you measure accuracy: on which test set, how often, and what you do when it drops.
- Who on your own team can see customer data, and how that access is logged.
Which signals show a company wants to automate your workflow?
For an AI product, the best signal is a person describing in public the manual work you automate. Here is where those moments show up and what each one suggests:
| Signal | Where it appears | What it suggests |
|---|---|---|
| "Is there a tool that can do X automatically?" | Reddit communities for the function, threads on X | Active search: reply fast (tool requests) |
| A post about a backlog, overtime or a manual process | Reddit, X, founder posts | Pain without a chosen solution (pain-point posts) |
| Frustration with a legacy tool or an AI tool that underdelivered | Reddit, X, review roundups in search results | Open to switching (competitor complaints) |
| A job post for an AI lead, an automation engineer or an operations analyst | Search results, company career pages | An owner and a budget for automation (hiring signals) |
| A new product launch in your buyer segment | Product Hunt, startup directories | Early adopters who buy tools quickly (product launches) |
Reach companies with a reason to buy this week
Startories finds the buying signal, verifies the decision-maker and runs the outreach until they book a call.
Example: an AI support agent and a ticket backlog
A fictional example of the full chain. You sell an AI agent that drafts first replies to support tickets and hands complex cases to a person.
Who qualifies
- B2B software companies with 30 to 300 employees and a support team of 3 to 15 agents.
- A mainstream help desk with an API your agent connects to.
- English-language support with many recurring "how do I" tickets.
- Excluded: outsourced support teams, and healthcare or banking support, where the review is longer than your team can handle yet.
The signal and the opening
Signal: a support lead at a fictional 90-person logistics software company posts on Reddit that the team is "two weeks behind on tickets after a release" and asks whether anyone lets AI draft answers.
First line: "Your post about being two weeks behind after the release caught my eye. We draft replies to the repetitive tickets inside your help desk and leave sending to your agents, so nothing goes out unchecked. Happy to run it on 200 of your closed tickets so you can judge the drafts before deciding anything."
What does a sequence look like when the goal is a pilot?
Continuing the support example, here is a four-touch sequence that moves from the public post to a pilot. Each touch answers one question a support lead will ask, in the order they usually ask it.
Touch 1, day 0: the backlog
Subject: "two weeks behind after the release". Body: the opening above, plus one line on scope: "It handles how-to and account questions; billing disputes and bug reports go straight to your team." Saying what the product does not do makes the rest believable.
Touch 2, day 4: the data question
Subject: "what happens to your tickets". Body: "Before anyone asks: tickets are processed in the US, deleted after 30 days and never used to train a model. Our subprocessors are listed on one page, linked here." These details are placeholders for the example; use your own, and only ones you can back up. The email answers what IT will ask, before the support lead has to forward your pitch to them.
Touch 3, day 9: the pilot terms
Subject: "a test on 200 closed tickets". Body: "We take 200 tickets you have already answered, draft replies blind, and your agents compare them with what they actually sent. If fewer than 7 in 10 drafts are usable with light edits, we part ways. If more are, the price for your volume is agreed in advance." The threshold is an example. Writing one down turns a vague trial into a decision.
Touch 4, day 15: the next release
Subject: "before your next release". Body: "Backlogs tend to return with the next big release. Want me to check in two weeks before yours?" The follow-up is tied to the buyer's calendar, not to your quota.
How should an AI company price pilots and contracts?
Usage costs make AI pricing different from ordinary SaaS. Every task costs you compute, so a contract that looks large can lose money at full volume. Run the numbers before you name a price in an email.
A worked example with assumptions you should replace. Each drafted reply costs you about $0.05 in model and infrastructure costs. The prospect handles 8,000 tickets a month, and 60% are the repetitive kind you draft: 4,800 drafts, or about $240 a month in direct cost. At a flat $1,500 a month, your gross margin on the account is around 84%. At $99 per seat for their 6 agents, you earn $594 against the same $240, and every extra ticket shrinks your margin.
- Put a minimum volume in your ICP. A friendly team that sends 300 tickets a month makes a pleasant pilot and an unprofitable customer.
- Charge for the pilot, or fix the price before it starts. A fixed pilot fee credited against the first contract filters out teams that are only curious and pays for the work of preparing their data.
- Price the unit the buyer already counts. Tickets resolved, documents reviewed or invoices matched are easier to approve than a token meter finance cannot predict.
What objections will AI buyers raise?
"We can build this ourselves with an API."
Some can. Ask who will maintain prompts, evaluations, integrations and edge cases once the first demo works. Your answer is the maintenance they avoid and the accuracy you have already measured, not the model.
"What happens to our data?"
Answer in writing before they ask: where data is stored, how long you keep it, whether it is used for training and which subprocessors see it. Vague answers end deals in security review.
"We tried an AI tool and it was wrong too often."
Take it seriously. Offer a pilot with a measured error rate on their own examples and a human review step, and agree on the threshold that would make it worth paying for.
"Our current vendor will ship this as a feature."
Possibly. Show depth in the specific workflow (integrations, edge cases, controls) that a general feature is unlikely to match, and keep initial terms short so the decision feels low-risk.
"Which model do you use, and what if it changes?"
Answer factually and move the conversation to what you control: your evaluation set, your review step and how you test a model change before customers see it. Buyers want to know that an upgrade will not quietly change their results.
Which numbers tell you the outbound is working?
For AI products, a booked call is only the first gate. Track the whole path:
- Replies and calls per signal type; pain posts, tool requests and hiring signals rarely perform the same.
- Pilots started per call, and the time from first email to pilot start.
- Pilot-to-paid conversion, with the reason each stalled pilot stopped (accuracy, security review, no internal owner).
- Expected usage cost against price, so a large pilot does not turn into an unprofitable contract.
- How often prospects open the trust sheet before a call. If nobody reads it, move its key lines into the email itself.
How to start with one workflow and one funnel
Startories handles the part before the pilot. It monitors Reddit, X, Product Hunt, directories and search results for the moments in the signals table above, matches each to a company, checks your ICP, finds the workflow owner and verifies their business email. The first email references the post or event, and you can approve every message before it goes out. See how this works as AI lead generation, or as an AI SDR that also handles follow-ups and replies.
Begin with one workflow and one buyer role, for example support leads at B2B software companies. Starter runs one funnel for $99 a month, with your first project starting on a 3-day full-access trial for $1; Growth adds funnels and reply classification. See pricing. If your buyers are engineers, read lead generation for developer tools; if you sell automation as a service, AI and automation agencies fits better. For a first ICP draft, use the ideal customer profile template.
Frequently asked questions
How do AI startups find their first customers?
Start with one workflow and one buyer role, look for people describing that manual work in public, and offer a short pilot on their own data. Early customers buy because the problem is acute, not because the product uses AI.
Should an AI startup offer free pilots?
A short pilot with a clear success metric works better than an open-ended free trial. Consider a small fee, or agree on the post-pilot price in advance: both filter out companies that are only curious.
What security questions do buyers ask AI vendors?
Expect questions on where data is stored and for how long, whether it is used to train models, which subprocessors and model providers handle it, who can access it, and how you measure and limit inaccurate output.
How should an AI product be priced for B2B deals?
Start from your cost per task at the buyer's real volume, then pick a price that keeps a healthy margin as usage grows. Per-seat pricing is easy to sell but can lose money on heavy users; pricing per resolved task tracks your costs more closely.
Can Startories find companies that want to automate a specific task?
Yes, when people describe it publicly. Startories detects posts that ask for a tool, describe a pain point or complain about a current vendor on Reddit, X and other sources, then matches them to companies in your ICP.
Is outbound right for an AI product priced under $100 a month?
Usually not as the main channel. Outbound costs time and money per conversation, so it pays off when a customer is worth at least several hundred dollars. Low-priced self-serve products often do better with content, communities and product-led growth.