The pitch sounds familiar by now: artificial intelligence that handles customer service queries once considered too complex for automation. What's less familiar is letting support teams write actual code to control when that AI kicks in.
That's the wager Intercom made on February 26, when Chief Product Officer Paul Adams unveiled a dozen product updates to the company's Fin AI platform during a live-streamed event. The announcement, paired with a quiet same-day overhaul of Intercom's help documentation, signals the company's most aggressive push yet into territory where bots typically stumble—multi-step queries that demand context, judgment, and the kind of improvisation human agents excel at.
Whether AI can genuinely replace that human touch remains an open question. But Intercom is clearly betting that the answer lies somewhere in the space between full automation and manual intervention.
Writing SOPs in Code
The most striking addition? Python-based conditional logic embedded directly into customer service workflows. It's a notable departure from the typical chatbot playbook, which relies on the AI's ability to interpret instructions written in plain language.
Now support teams can write deterministic branching logic themselves—checking dates, wrangling timezones, processing arrays, combining multiple conditions. The advantage is precision. If you need Fin to behave exactly one way under specific circumstances, you can code that behavior rather than hoping the AI infers it correctly.
This level of control matters more as workflows grow intricate. Consider a refund request that depends on purchase date, subscription tier, regional regulations, and whether the customer has contacted support before. Previously, you'd trust the AI to parse all that from written instructions. Now you can specify it programmatically.
The company hasn't published a detailed changelog of all 12 updates, though the revised documentation fills in many blanks. Beyond Python conditions, Intercom has expanded the toolkit support teams use to construct what it calls "Procedures"—structured workflows that guide Fin through complex resolutions.
The AI Writes the First Draft
Perhaps more immediately practical for most teams: Intercom will now draft entire Procedures based on existing standard operating procedures. Paste in up to 5,000 characters of your SOP, and the system generates a workflow tailored to your industry—SaaS, ecommerce, fintech, gaming.
It's a pragmatic move. Building these multi-step workflows from scratch is tedious, and most support teams already have documentation gathering dust in Google Docs or Notion. Why not let the AI do the grunt work?
Once drafted, teams can train Procedures using examples from historical conversations, a feedback loop designed to reduce false positives. The system determines intent through "When to use this Procedure" logic, though skeptics might note that intent detection remains one of AI's most persistent weak points.
Testing Before the Real Customers Arrive
Then there's the testing problem. Complex, multi-turn support workflows are notoriously difficult to validate before deploying them live. Make a mistake, and real customers experience the frustration.
Intercom's response is Simulations—still labeled closed beta but now significantly more capable. Teams can generate AI-suggested test scenarios (including "happy path" cases) or write custom simulations manually. Tests can simulate specific users or brands, attach images, set the simulation time, define available data attributes.
When you run a simulation, Intercom provides pass/fail results with full transcripts. The goal is debugging without consequences—catching failures before they reach actual customers, building a regression library as workflows evolve.
It's the kind of feature that suggests Intercom has been listening to support leaders who've watched automation projects crater because there was no safe way to iterate.
When the Customer Changes Their Mind
Real support conversations are messy. A customer starts asking about billing, pivots to a feature request, then circles back to demand a refund. Human agents navigate this naturally. AI often doesn't.
Intercom's attempting to solve this with what it calls "Agentic Switch," allowing Fin to automatically shift between Procedures when customer intent changes mid-conversation. The system can ask clarifying questions, prioritize help center content over structured workflows when appropriate, and handle the layered, sometimes contradictory nature of how people actually communicate.
Sub-procedures add another dimension, letting teams nest reusable logic and share context across workflows. It's less flashy than the other updates, but practically speaking, it's a nod toward maintainability as Procedure libraries inevitably balloon.
The Analytics and the Bill
On the back end, Intercom has expanded analytics to include explicit outcomes: Triggered, Pending, Resolved, Handoff, Escalated. Support leaders can drill into these flows to identify where Procedures succeed and where they fracture.
There's also a billing change arriving March 12: configured handoffs will count as successful Procedure outcomes. For support teams managing automation budgets, that's the kind of detail that warrants attention.
The Fine Print and the Competition
Procedures remain under what Intercom calls "managed availability"—you need to request access through your account team. Simulations are still locked behind closed beta. And for teams currently using Fin Tasks (the predecessor to Procedures), Intercom plans to auto-migrate simple, single-block Tasks later this year. Complex Tasks will remain supported for at least 18 months.
The company introduced Procedures and Simulations as part of Fin 3 last October, positioning them as components of a "Fin Flywheel" for multi-step query resolution. These latest updates deepen that foundation, adding more authoring tools, tighter deterministic controls, broader testing coverage.
Intercom claims Fin can deliver "up to 87%" instant resolution when properly configured, though that figure comes with significant caveats around scenario complexity and content quality. Translation: your mileage will vary.
The timing isn't coincidental. Zendesk has been pushing its AI Agents with custom quality assurance and multi-turn testing. Salesforce is pouring resources into Agentforce evaluation tools. The competitive pressure to offer both flexibility and guardrails in customer service AI is mounting fast.
Intercom's bet—combining AI-drafted workflows with Python-level control and simulation-based testing—rests on a particular theory: that support teams will embrace the complexity if it means resolving queries that have historically required human judgment. At least some of the time.
Whether that theory holds depends less on the technology and more on how support teams actually use it. Because the hardest problems in customer service have never been purely technical.
