Inbound form submissions are only valuable if the leads behind them are real. For a North American and EU manufacturing software company managing a diverse portfolio of products and brands, spam and vendor solicitations were quietly eroding rep productivity and distorting funnel metrics. Process Pro designed and deployed a HubSpot AI Spam Agent that automatically screens inbound submissions before they ever reach a sales rep, protecting pipeline quality and giving the team back hours of wasted time every week.
The client began noticing inconsistencies in MQL quality week over week. Leadership started drilling into the top of the funnel and identified that spam and vendor solicitations were getting through HubSpot's native filtering and making their way to sales reps. The impact was straightforward: reps were spending time on records that didn't deserve any time at all.
The problem was compounded by the complexity of their product portfolio. With brands spanning different buyer types, form languages, and inquiry styles, a one-size-fits-all spam rule wasn't going to work. A transactional product might receive short, low-context form submissions that look suspicious but are perfectly legitimate. A complex enterprise product might receive detailed outreach that turns out to be a vendor pitch. The client needed something smarter than a simple keyword filter.
The client came to the conversation with a head start: a historical list of contacts that had already been manually flagged as spam or solicitation by their team. That list became the foundation for everything that followed.
Audit and Discovery
Process Pro started by analyzing the client's historical spam data. Using the manually flagged list the client provided, our team assessed patterns and trends across record types, identifying common signals like reverse pitches, brand hijacking, celebrity name combinations, and gibberish in name and company fields. That analysis informed the initial prompt structure and gave the team a baseline to test against. The target from the start was an 85% or higher accuracy rate on known spam records.
Prompt Engineering
The spam agent runs inside a HubSpot workflow and evaluates each inbound submission against a structured prompt built specifically for this client. The prompt establishes a role for the agent, defines which contact properties to evaluate (first name, last name, company name, domain, and message field), and walks the agent through a series of gates in a defined order. The approach evolved significantly over time. Early versions gave the agent a list of things to evaluate without specifying an order, which led to inconsistent scoring. Structuring the evaluation as a step-by-step process with clear gates produced more consistent results across record types.
The prompt also accounts for the client's product complexity. Transactional, self-serve brands receive less punitive scoring for short or low-context form submissions, since those products attract a different type of buyer who may simply be raising their hand without a detailed message. Enterprise-facing brands are held to a higher standard.
A buy-intent override is built into the prompt logic. If a submission clearly indicates purchase intent, such as a request for pricing, a demo, or a quote, the agent is instructed not to disqualify it regardless of other signals. This ensures that genuine buyers are never blocked.
Scoring and Output
The agent assigns a spam score from 0 to 100 representing confidence that the record is spam, and stores a written rationale directly on the contact record. This transparency is a core part of the design. Sales leaders can see not just whether a record was flagged, but why. If the score meets or exceeds the threshold, the lead outcome is set to spam or solicitation, stamped as set by the bot, and the record is routed out of the MQL workflow. Records that fall into a middle range can be sent to a manual review queue where a human makes the final decision. That review period also provides valuable feedback for refining the prompt over time.
Workflow Integration
The agent sits inside the client's existing MQL workflow, triggered after a contact reaches their total lead score threshold. Before the agent evaluates a record, several bypass checks run first: contacts from uploaded event lists are skipped entirely, existing customers are routed out, and support form submissions are excluded. This keeps credit consumption controlled and ensures the agent is only spending resources on records that actually need evaluation.
Testing and Iteration
Getting the prompt dialed in took roughly two to three weeks for the first implementation. Our team worked through approximately 30 to 40 iterations, testing inside HubSpot's Breeze Assistant before going live in the workflow to avoid consuming credits during refinement. The biggest challenge was calibration: early versions of the prompt were either too permissive or too aggressive. Establishing a clear role for the agent and structuring the evaluation as an ordered set of gates brought the scoring into a more consistent, predictable range.
Ongoing Monitoring
In the early weeks after launch, the team reviewed flagged records on a weekly basis, checking the agent's written rationale against their own judgment to identify any gaps or hallucinations in the scoring logic. Segments built around the bot-set lead outcome and timestamp properties made this review process straightforward.
The Results
Over a six-month period with the most recent workflow version, the agent evaluated approximately 2,000 contacts and flagged 258 as spam or solicitation. Across a year and a half of running across workflow versions, approximately 15,000 contacts moved through the screening process.
Rep time saved is estimated at two to three hours per week based on the volume of records flagged and the average time a rep would spend investigating and dismissing a spam submission. Credit consumption remained well within the client's baseline Breeze credit allocation throughout.
The spam agent has become a standing part of the client's MQL workflow infrastructure. Reps spend less time on records that should never have reached them, funnel metrics are cleaner, and the team has a transparent, auditable record of every spam decision the agent makes. The prompt is designed to be refined over time as new spam patterns emerge, and the manual review band gives the team a built-in mechanism to keep improving accuracy without going back to square one.
Process Pro has since deployed a streamlined version of the same agent for a second client, using their implementation as the foundation and adapting the prompt for that client's specific inquiry patterns and industry context. The core structure, stepped evaluation gates, buy-intent safeguards, score-based routing, and Slack-friendly error handling, transfers across accounts with relatively minor customization.
For any B2B company generating meaningful inbound form volume, the spam agent is a practical, low-overhead way to protect rep time and keep CRM data clean.
