Help Desk and Ticketing: from Remedy to AI agents

Support in this era lived in three places: a phone queue, a shared email inbox, and, in larger companies, an on-prem ticketing system. Remedy was the name most IT and support leaders knew. Its Action Request System shipped in 1991 and became the default way big enterprises logged, assigned, and tracked requests.

The software ran on your own servers and needed admins to customise forms and workflows. Changing a field or a routing rule was a project, not a setting. Smaller teams skipped all of it and worked from a shared inbox, where the real routing system was whoever replied first.

Phones carried the urgent work. Customers waited in ACD queues, agents typed notes into the ticket afterwards, and knowledge lived in binders, wikis, and the heads of senior agents. Remedy itself changed hands twice in two years, bought by Peregrine in 2001 and sold to BMC in 2002 after Peregrine's bankruptcy.

What broke was visibility. Tickets were invisible to customers, duplicates piled up across email and phone, and nobody could say how long anything really took.

Zendesk launched in 2007 from Copenhagen and showed that a help desk could be a browser tab you signed up for in an afternoon. Freshdesk followed in 2010 from Chennai with the same idea at a lower price. Desk.com, Help Scout, and Kayako filled in the rest of the market.

The unit of value became the seat. You paid per agent per month, and every new hire meant another licence. That model worked because support capacity really was a function of headcount: more tickets needed more people.

The SaaS desks added what on-prem never had. Customer-facing help centres, web forms that created tickets, live chat widgets, macros for common replies, triggers and SLA timers, and dashboards anyone could read. Zendesk went public in 2014 on the strength of that model.

What broke was scale. Macros and trigger rules multiplied until nobody understood the routing. Help centre articles went stale. Volume still grew linearly with customers, so the only lever was hiring more agents or offshoring to a BPO.

Machine learning arrived as a suggestion engine. Zendesk introduced Answer Bot in 2017, which read an incoming ticket and offered the customer help centre articles before an agent picked it up. Intercom shipped its own Answer Bot in 2018 and the successor Resolution Bot in 2020.

These bots matched questions to known answers. They could not take an action, so a refund, an address change, or a password reset still went to a human. Deflection was the metric, and deflection often meant a customer gave up rather than got helped.

Inside the console, ML did triage: intent and sentiment tags, language detection, suggested macros, and skills-based routing. Agent assist tools proposed replies and pulled up related articles. The human still read every ticket that mattered and clicked every button that changed anything.

The market consolidated as growth slowed. Zendesk was taken private in 2022 by a Hellman and Friedman and Permira led group in a deal worth about $10.2 billion, just before ChatGPT reset expectations for what a support bot could do.

The turn came in March 2023, when Intercom launched Fin on GPT-4, answering only from a company's own content and handing off when unsure. By 2026 Fin reports a 76% average resolution rate across more than 12,000 customers (vendor-reported), and it now updates accounts and processes refunds and payments, not just answers. Klarna showed the volume case in February 2024: its assistant handled two-thirds of service chats in its first month, which the company equated to 700 full-time agents. By May 2025 its CEO was telling Bloomberg the company would hire humans again, because cheaper AI had meant lower quality.

Incumbents bought their way in. Zendesk agreed to buy Ultimate in March 2024 and completed its purchase of Forethought in March 2026. In June 2026 Salesforce signed a definitive agreement to acquire Fin, which had renamed itself from Intercom, for about $3.6 billion, and it closed the deal on 10 September 2026. Fin now runs inside Salesforce AI Labs beside Agentforce, which Salesforce reported at $1.2 billion in annual recurring revenue in its first fiscal quarter of 2027. Agent-native companies are well funded: Decagon raised $250 million at a $4.5 billion valuation in January 2026, and Sierra announced a $950 million raise at a valuation above $15 billion in May 2026. They sell agents that rebook, refund, and change plans for brands such as Gap, SiriusXM, Duolingo, and Chime.

Pricing is shifting too. Zendesk moved to outcome-based pricing for AI agents in August 2024, and Fin charges $0.99 per outcome. Salesforce started at $2 per conversation, moved to $0.10 per action in 2025, and its Agentforce Service page now sells on a pay-per-resolution basis. When software does the work, the seat is a weaker unit, which is the argument I made in AI didn't kill SaaS, it killed the seat.

Vendor resolution rates are self-reported, and Gartner found in late 2025 that only 20% of service leaders had actually cut headcount because of AI. A later Gartner survey of more than 320 leaders found 85% expanding what human agents handle, with 31% implementing or planning AI-driven layoffs. Customers are not asking to be rid of people either: 87% told Gartner that a route to a human is essential. The phone line is now a target too: as I wrote in your phone line became the front door, automating it hands attackers a patient, polite agent that never gets suspicious.

My bet is that by 2031 the help desk stops being a place where tickets wait. Most routine requests will be resolved by an agent before a ticket would have existed, across chat, email, and voice, and the queue that remains will be exceptions, disputes, and anything that changes who controls an account.

Gartner predicts agentic AI will resolve 80% of common customer service issues without a human by 2029. I think that holds for common issues at companies with clean knowledge and good backend APIs, and fails everywhere else. The constraint is not the model. It is whether the agent can safely call the refund system, the order system, and the identity system with the right scope.

Two things change the shape of the work. First, the customer side gets agents too: Gartner expects machine customers to start a large share of service requests, and its 2026 customer survey already found 58% of customers who use GenAI have used it to complete a task for them (74% in B2B). I expect support agents to negotiate with other agents. That is the world Future Tech describes as agent-answered calls. Second, pricing will likely finish its move to outcomes, as I argued in the pricing model that proves you're actually AI-native, and I expect vendors to compete on audited resolution more than on seats.

The human team gets smaller and more senior. It owns procedures, policies, escalations, and the identity checks no agent should run alone.