Support agents that route tickets, answer knowledge, and know when to step back.

The support inbox is the wrong place for a generic chatbot

The vendor demos look great. The reality of dropping a generic bot in front of your customers is something else — wrong answers given confidently, escalation paths that don’t exist, refund promises nobody approved, and a CSAT score that quietly tanks for a quarter before anyone connects it back to the bot.

Support is the highest-stakes place in your business to deploy AI badly. We treat it that way. Every integration we ship is designed around the assumption that the model will be wrong sometimes, and the system has to catch it before the customer does.

What we ship into your helpdesk

Ticket triage and routing

Every incoming ticket gets classified — by product area, urgency, customer tier, sentiment, and whether it’s a known issue. Routing rules use those classifications instead of brittle keyword matches. Tickets land in the right queue before a human ever opens them.

Draft answers from your knowledge base

The agent doesn’t reply to customers directly by default. It drafts an answer pulled from your knowledge base, your past resolved tickets, and your product docs — and drops it into the agent’s reply field. Your human agent reads, edits, sends. Response time drops; voice stays consistent; nothing wrong goes out unsupervised.

Intelligent escalation

The model knows when to stop. Frustration signals in the customer’s message, repeated contact in 24 hours, mentions of cancellation or legal, VIP account flags — all trigger immediate human escalation with full context already summarised for the agent picking up.

Sentiment and churn-risk signals

Every conversation gets a sentiment score that flows back to your CRM. Aggregate trends surface in dashboards. Account managers see the customers who are quietly getting unhappy before they show up in a churn report.

Helpdesks we integrate with

Active integrations on Zendesk, Intercom, Freshdesk, Help Scout, HubSpot Service Hub, Salesforce Service Cloud, Kustomer, and custom helpdesks built on tools like Linear, Jira Service Management, or in-house ticketing. The integration uses native macros, native tags, native triggers — no parallel system for your agents to babysit.

Human-in-the-loop by default

We ship three trust tiers, and you choose where each ticket class sits:

• Suggest: Model drafts a reply, human agent reviews and sends. The safe default.
• Auto-respond with review window: Model sends, but a senior agent has 10 minutes to recall and edit. Useful for high-volume Tier-1 questions.
• Full auto: Reserved for narrow, low-risk categories — order status, hours of operation, password reset. Everything outside that falls back to Suggest.

You can move ticket categories between tiers as confidence builds. We don’t lock you in.

Where the model runs

Same three options as our other integrations, picked by data sensitivity:

• Vendor API (default): Claude, GPT, or Gemini through no-retention endpoints. Fastest path to live.
• Your cloud: Integration runs inside your AWS, Azure, or GCP account; vendor models reached via your own keys.
• On-prem / open-source: Llama or Mistral hosted in your infrastructure. For regulated or air-gapped environments.

How we control AI cost

Support volume is unpredictable, and an under-budgeted bot is how you wake up to a five-figure bill from a Reddit post. We ship with:

• Per-ticket token caps and per-conversation context windows
• Daily and monthly spend ceilings with automatic shutoff and Slack alerting
• Aggressive KB-answer caching — same question, same answer, no second model call
• Tiered routing — small model classifies, premium model drafts only when classification confidence is low
• A real-time cost dashboard your ops team can watch without us in the loop

The 30-day proof of value

We pick one ticket category that’s draining your agents — password resets, order status, billing FAQs, integration troubleshooting, whatever is loudest. We ship draft-answer integration for that category in 30 days, measured against your current handle-time and CSAT. If those don’t move, you don’t scale to category two. We put the cost ceiling in writing on day one.

Frequently asked

Will customers know they’re talking to an AI?+

In the default Suggest tier they aren’t — your human agents send every reply. In auto-respond tiers we recommend transparent disclosure, and most jurisdictions are moving toward requiring it anyway. We’ll help you write the disclosure copy.

What if our knowledge base is out of date?+

Common. Week one of every engagement we audit what the KB actually covers versus what tickets actually ask, and give you a gap report. We can also seed the KB from past resolved tickets so the agent has something useful to work with even before docs are cleaned up.

How do you handle hallucinations?+

Every drafted answer is grounded in retrieved documents, and the model is instructed to refuse rather than guess when retrieval is weak. Refusals route to a human. We log every answer and the documents that produced it, so you can audit any reply your customers receive.

What does this cost ongoing?+

Two components: token spend (your own API bill, usually $300 – $3,000 / month at SMB-mid-market volume) and an optional monthly retainer for monitoring, KB tuning, and prompt updates as your product changes. We size both before you sign.