Build an Intent Agent Without the ABM Tax (2026)
Skip the $20K/mo ABM platform. A lightweight intent agent costs $1,025/mo, routes 2,000 signals at $0.51 each, and uses fixed rules under 100ms. Start with five signals tied to clear actions.
Build an Intent Agent Without the ABM Tax
At 2,000 routed signals monthly, a $20,000 contract costs $10 per routed signal. A lightweight build can cost about $0.51 per routed signal using the assumptions below.
An intent-routing agent watches buyer activity and chooses the next action. That action might assign a rep, create a task, post in Slack, or draft an email.
The action matters more than the signal.
Step 1: Choose Signals That Predict an Action
Most intent programs collect too much data.
They track topic surges, page views, funding events, hiring, job changes, and ad clicks. Then they dump everything into a dashboard nobody checks.
Start with five signals tied to clear actions:
1. A target account visits your pricing page. 2. A known contact returns three times within seven days. 3. A former customer views a product page. 4. An open opportunity shows new website activity. 5. A qualified account submits a sales form.
You can add G2, LinkedIn, Clay, or reverse-IP data later. Don't buy more signals until the first five produce meetings.
Clay added Intentsify data covering 1.1 trillion monthly signals. Those signals cover 4.2 million accounts and 33,000 topics.
The commercial terms weren't disclosed.
That's common with intent data pricing. Vendors share big volume numbers before they share a usable price.
Volume isn't value.
A pricing-page visit from an open opportunity matters. A topic surge from an account outside your market doesn't.
Create one standard record for every incoming signal:
- Account domain
- Contact email, when known
- Signal type
- Signal timestamp
- Source
- Current account owner
- CRM stage
- Target-account status
- Previous activity count
- Consent or suppression status
Store the raw event before changing it. You'll need that record when a rep asks why they received the alert.
Tools: Your website analytics, CRM, product database, and licensed data feeds.
Expected outcome: One clean event stream with enough context to make a routing decision.
Step 2: Keep the Fast Decision Layer Boring
Don't place a large language model inside every routing decision.
That may sound strange coming from someone who builds AI agents. It's still the right call.
Your first decision layer should use fixed rules. Fixed rules are fast and cheap. They're also easy to audit.
A simple score might look like this:
- Pricing-page visit: 30 points
- Open opportunity: 40 points
- Target account: 20 points
- Repeat visit: 10 points
- Existing customer: minus 20 points
- Suppression flag: immediate stop
Route accounts scoring 70 or higher. Send scores from 40 through 69 into an AI review queue.
Everything below 40 gets logged without creating rep work.
This local rules layer can finish within 100ms on basic cloud infrastructure. It doesn't need model inference.
Be careful with claims about the OpenAI Decisions API.
The supplied research doesn't include official OpenAI pricing or a sub-100ms promise. A cited Realtime-Bench paper measured OpenAI Decisions at 1.126 seconds average latency.
That isn't a product service-level promise. It also isn't under 100ms.
Use OpenAI Decisions only if it's available in your account. Confirm its current price, data location, and latency in OpenAI's official documentation.
The API can choose between approved routes. It shouldn't invent new actions.
Give it four allowed results:
- `priority_route`
- `standard_route`
- `research_only`
- `suppress`
Reject anything outside that list.
Tools: A small API endpoint, Postgres, Redis, or n8n's data store.
Expected outcome: Most signals route without AI. Unclear cases receive structured AI review.
Step 3: Separate Judgment From Execution
The model decides what should happen. Your workflow decides what actually happens.
This keeps a model from sending emails or changing account ownership on its own.
For a priority signal, the agent might trigger four fixed tasks:
1. Assign the account using CRM ownership rules. 2. Create a task due within 15 minutes. 3. Post the signal and reason in Slack. 4. Draft an email for rep approval.
The model shouldn't choose random Slack channels. It shouldn't rewrite CRM ownership. It shouldn't send an email without checking suppression rules.
We use n8n for this work at StoryPros. It gives us more control over retries and lets us self-host.
Set one n8n branch for each approved result. Keep model output away from direct database commands.
Every action also needs an idempotency key. Use the account, signal type, and time window.
That stops one pricing-page visit from creating six tasks.
Add retries for temporary failures. Stop after three attempts and alert a person.
Check outputs before acting on them.
An n8n research workflow tested 3,408 controlled failures. Retries recovered 234 temporary failures. Output checks rejected another 234 structural defects.
That's what a working agent looks like.
OpenRouter reportedly uses the same split with its Rasp sales agent. Rasp drafts, tags, and routes work. Salespeople keep sensitive decisions.
Rasp reportedly costs about $30 daily and saves 600 hours monthly. It also reduced demo processing time from 103 minutes to 44 minutes.
Tools: n8n, your CRM API, Slack, and an email drafting model.
Expected outcome: Approved decisions create predictable actions without duplicate tasks or surprise emails.
Step 4: Calculate Cost Per Routed Signal
Don't judge a vendor by signal volume.
Focus on cost per useful action instead.
Use this formula:
Monthly system cost ÷ successful routed signals = cost per routed signal
Here's an example budget. These are planning assumptions, not vendor quotes.
| Cost item | Monthly assumption |
|---|---|
| n8n hosting | $50 |
| Database and logs | $25 |
| AI decision budget | $100 |
| Contact enrichment budget | $250 |
| Four maintenance hours at $150 | $600 |
| Total | $1,025 |
Assume the agent receives 10,000 raw signals monthly. Fixed rules and suppression checks route 2,000 useful signals.
That produces this result:
$1,025 ÷ 2,000 = $0.51 per routed signal
Now test larger software contracts against the same volume.
| Monthly cost | Routed signals | Cost per routed signal |
|---|---|---|
| $1,025 | 2,000 | $0.51 |
| $5,000 | 2,000 | $2.50 |
| $10,000 | 2,000 | $5.00 |
| $20,000 | 2,000 | $10.00 |
The $20,000 figure is a sensitivity test. It isn't a published 6sense or Demandbase price.
The supplied sources don't provide public prices for 6sense, Demandbase, or Intentsify through Clay. Any honest intent data pricing comparison must label them as quote-based.
Get the actual quote. Add setup fees, data fees, seats, and required add-ons.
Then divide by successful routed actions.
Don't divide by raw signals. A billion ignored signals are still worth zero.
Tools: A spreadsheet, n8n execution logs, and CRM task records.
Expected outcome: A monthly number you can defend without vendor math.
Step 5: Measure Decisions, Not Activity
A routed signal isn't automatically a good signal.
Track what happened after routing.
Your scorecard should include:
- Raw signals received
- Signals blocked by suppression rules
- Signals routed
- Duplicate actions prevented
- Median signal-to-task time
- Rep acceptance rate
- Meetings booked
- Opportunities created
- Cost per routed signal
- Cost per accepted signal
- Cost per meeting
Review every rejected route during the first month. Review a sample of accepted routes too.
ElevenLabs improved its Dom AI SDR from 78% to 92% qualification accuracy. The team used a documented, multi-signal waterfall instead of one model judgment.
Dom reportedly reached a four-minute median qualification time. Human qualification had taken two days.
The system also created more than $1 million in qualified pipeline during one month. ElevenLabs reported 180 hours saved during that period.
Don't take this as a reason to buy an AI SDR.
V1 wasn't good enough.
ElevenLabs added website research and other checks when live conversations were unclear. That feedback loop raised accuracy by 14 percentage points.
Your routing agent needs the same treatment.
Start with human approval for email drafts. Remove approval only after measuring accuracy and complaints.
StoryPros builds working agents around this pattern. Strategy picks the signal, message, and route. AI handles the judgment that rules can't cover.
The workflow carries out the action.
Tools: CRM reports, n8n logs, and a weekly routing review.
Expected outcome: Better routing each month without giving the model unchecked control.
FAQ
How much does an AI automation cost?
A lightweight intent-routing agent can run near $1,025 monthly using the sample budget above. At 2,000 successful routes, that equals about $0.51 per routed signal.
Build cost depends on data sources, CRM complexity, and approval rules. Licensed intent data can cost more than the agent itself.
How do I operationalize buyer intent signals?
Send each signal into one event format. Add CRM context, apply fixed rules, and send unclear cases through structured AI review.
Each approved route should trigger a fixed action. Common actions include assigning a rep, creating a task, posting in Slack, and drafting an email.
How do I automate a signal-based outbound motion?
Start with one high-value signal, such as target accounts visiting pricing pages. Connect that signal to n8n, apply scoring rules, and create approved CRM tasks.
Keep sending behind human approval at first. Track meetings and accepted signals before adding more sources.
Is the OpenAI Decisions API required?
No. A local rules engine can handle most routing decisions faster and cheaper.
Use the OpenAI Decisions API for unclear cases if your account supports it. Confirm official pricing and latency before placing it in a critical workflow.
Should I buy 6sense or Demandbase instead?
Buy 6sense or Demandbase when their proprietary data improves revenue enough to cover the contract. Don't buy either platform just to create tasks and Slack alerts.
Ask for the full annual cost. Divide it by accepted routed signals and meetings booked.
Related Reading
How much does it cost to build an AI intent-routing agent?
A lightweight intent-routing agent costs about $1,025 per month using n8n hosting, a database, AI decision budget, enrichment, and four maintenance hours. At 2,000 successful routes, that equals $0.51 per routed signal. A $20,000 monthly ABM platform at the same volume costs $10.00 per routed signal.
How do I route buyer intent signals without an expensive ABM platform?
Apply fixed scoring rules locally: pricing-page visit scores 30 points, open opportunity scores 40, target account scores 20. Route accounts at 70 or higher immediately. Send scores between 40 and 69 to an AI review queue. This layer finishes in under 100ms without model inference.
What actions should an intent agent trigger automatically?
A priority signal should trigger four fixed tasks: assign the account using CRM ownership rules, create a task due within 15 minutes, post the signal in Slack, and draft an email for rep approval. The model decides what happens. A workflow engine like n8n carries out the action.