AI automation
You bought the AI add-on. Six months later you still do not know if it is working. The dashboard is green. CSAT has not moved. Your team thinks it is helping. Your customers occasionally get told the wrong refund policy.
Smart Instinct configures AI support agents across every major helpdesk: Intercom Fin, Gorgias AI Agent in both Support and Shopping modes, Zendesk AI agents and Copilot, Freddy AI for Freshdesk, Tidio Lyro, and Pylon's AI agent. We moved The San Francisco Marathon from under 30% deflection to 56% in production. The framework has almost nothing to do with which platform you bought and almost everything to do with how carefully it gets tuned.
Partner status you can verify
An AI customer support agent is a configured, language-model-backed assistant inside your helpdesk that reads incoming questions, finds answers in your connected knowledge sources, takes limited actions in connected systems such as refunds, subscription updates and order lookups, and escalates anything outside its scope to a human. The configuration is the entire difference between one that earns its subscription and one that quietly invents policy.
When most teams switch their helpdesk AI on for the first time, deflection lands somewhere under 30%. It handles the easy questions and escalates the rest. That is not failure, it is the starting line.
The work is moving from that baseline to a number that changes your support economics. On the Zendesk build we shipped for The San Francisco Marathon, the AI chat flow we configured reaches 56% deflection in production, and 67% the following year. Across the Gorgias instances we have configured, containment lands between 26 and 56%. Those are our numbers out of our own builds, which is a different thing from a vendor case study.
The range matters more than any single number, because the right target depends on your knowledge base maturity, your customer base, your ticket complexity and how aggressive you want the AI to be.
What we will not promise is a specific deflection percentage on your account. Anyone who quotes you one without seeing your data is bluffing. What we will promise is a methodology that moves the number, with you deciding how far the AI is allowed to go.
Take a team handling 10,000 conversations a month at roughly $4 of loaded agent time each. Every percentage point of deflection moves $400 a month off the cost stack, or $4,800 a year. Going from 25% to 50% is roughly $120,000 a year of cost or capacity recovered.
Your numbers will differ, but the lever is real and it scales with volume. Capacity counts too. Deflection that frees agents from order-status questions is capacity they spend on retention and the conversations that actually move revenue.
Product names below match current vendor documentation as of August 1, 2026. They change often, which is most of why this page exists.
Autonomous conversations across support, sales and ecommerce. We configure Guidance authoring, knowledge source curation, escalation rules and pre-launch testing. Note that new Fin Tasks creation was disabled on March 12, 2026, with Procedures positioned as the successor. Existing Tasks still run and can still be edited.
Support Agent handles post-purchase work such as order tracking, returns and subscription updates. Shopping Assistant handles pre-purchase questions. AI Agent 3.0 shipped in the Summer 2026 release, introducing Skills as the organising framework, alongside Gaia, their coaching agent that reads your skills and recent tickets and drafts fixes.
Two separate products. AI agents talk to customers across email, chat and messaging. Copilot helps your human agents work faster. We configure both, plus the Explore reporting to actually measure them.
Runs in three modes: Freddy Agent for autonomous resolution, Freddy Copilot facing your agents, and Freddy Insights for analytics.
Built for smaller ecommerce teams. Simpler to deploy than Fin or Gorgias AI Agent, with a cost structure that suits lower volume. We are an authorised Tidio Partner.
Works inside the shared Slack channels, Teams chats and Discord servers where B2B SaaS teams actually talk to customers, rather than only inside a help portal. Worth knowing before you build: agents executing Runbooks in Pylon cannot reach training data or knowledge base content, and will invent links if you ask them to cite docs. That constraint shapes the whole design.
The AI is grounded in help center articles written two years ago that contradict current policy. Launches and price changes never got backfilled. The AI answers from the outdated source and ships wrong information to your customers, confidently.
Most platforms support Guidance or system-prompt-style instructions covering tone, policy and what to escalate. Most teams skip them or paste a generic template, so the AI sounds nothing like the brand and falls back on whatever the vendor shipped by default.
Too cautious and it escalates everything, deflecting nothing. Too aggressive and it starts handling refund disputes, legal threats and edge cases it has no business touching.
Order tags syncing from Shopify, refund approvals routing through Slack, daily exports landing in the finance drive, returns triggering Klaviyo flows. Unglamorous work that quietly saves a support team a day a week.
We build in Make and Zapier with real error handling, retry logic and version control, and we will tell you when one fits the job better than the other rather than defaulting to whichever you already pay for. We are a Zapier Partner.
When no-code cannot do it, we write code. Rate-limited APIs, multi-step transactions, authenticated webhooks, queueing. Serverless functions, scheduled jobs and integration middleware in Node or Python, hosted in your cloud or ours, monitored and documented.
We do not quote one, and it is worth being skeptical of anyone who does before seeing your data. Across our own builds it lands between 26 and 56%. Where you land depends on knowledge base maturity, ticket complexity and how aggressive you let the AI be.
Almost always grounding. The agent is reading stale or contradictory content, or it has no Guidance telling it what it may not decide, or escalation is set so aggressively that it attempts questions it should hand to a human. All three are configuration problems, not model problems.
Yes, and you should model it before you commit. HubSpot moved Breeze Customer Agent to $0.50 per resolved conversation on April 14, 2026, billed as 50 credits, on Pro and Enterprise only. eDesk counts an AI Agent resolution when the customer does not reply within 72 hours. Definitions of a resolution differ by vendor and they materially change the bill.
Twenty platforms. The framework is the same across all of them: curate the knowledge sources, author real guidance, wire the agent to the systems that hold the answer, and design escalation that matches what your team can absorb. What changes is the vendor's own tooling, not the method.
It is scoped as part of a build, so two to four weeks on Essential through four to six on Premium. If your AI is already live and underperforming, the $500 audit tells you what is wrong before you commit to a project.
Thirty minutes, no charge. Show us where the AI got it wrong and we will tell you whether it is grounding, guidance or escalation, before you spend anything.
Smart Instinct helps provide various services to optimize your help desk or contact center.
.svg)
.svg)
