Industries / Technology / Software and SaaS / ERP and CRM software
Customer service
Service software keeps the record of each customer's problem, called a case or a ticket, from the first message to the fix, along with the answers that support staff, and now AI agents, reuse.
5
Milestones achieved, from 1991 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1991
Help requests become tickets
Remedy Corporation, incorporated in November 1990, began shipping its Action Request System in December 1991. Companies used it mainly as an internal help desk to log, track and resolve support requests, and could change its forms, fields and workflow without programming; Remedy compared it to "the versatility of a spreadsheet". By the end of 1996 it had licensed the system to more than 1,850 customers at over 3,400 sites.
Credit
Remedy Corporation
1998
Sales and service, one record
In May 1998 Siebel Systems bought Scopus Technology, a maker of customer service, field service and call center software, in a share deal valued at about $415 million. Siebel Call Center then gave one agent both the sales and the service history of each customer, so that, in Siebel's words, "each service request" could "result in additional sales opportunities".
Credit
Siebel Systems; Scopus Technology
2007
A help desk bought online
Zendesk was founded in Copenhagen in 2007 by Mikkel Svane, Morten Primdahl and Alexander Aghassipour to sell a help desk as a subscription service run on the web. Most customers found it online and signed up on its website, often after a free trial, without meeting a salesperson. When it went public in 2014 it had more than 42,000 customer accounts in over 140 countries, answering their customers in more than 40 languages.
Credit
Mikkel Svane, Morten Primdahl and Alexander Aghassipour
2023
Support files hold keys
From 28 September to 17 October 2023 an attacker used a stolen service account to read files in Okta's customer support system belonging to 134 of its customers. Some were HAR files, recordings of a browser session that customers attach to a support case, and they held live session tokens, which the attacker used to take over sessions at five customers. Customers saw it first: 1Password reported suspicious activity on 29 September and BeyondTrust on 2 October, and Cloudflare says it contacted Okta before Okta notified it. Cloudflare then released a free, open-source HAR sanitizer that strips such secrets out.
LessonA support case can carry the keys to a customer's systems, so the case system needs the same protection as the product.
Credit
Okta's investigation; first reported by 1Password, BeyondTrust and Cloudflare
2024
Paying per solved case
In August 2024 Zendesk began charging for its AI agents per "automated resolution", an issue the AI settles without a person, instead of per staff seat. The trade site CX Today reported that Zendesk called this an industry first but was one of many vendors moving off seat pricing. Zendesk counts only conversations that were not passed to a person, and has a language model read each one to confirm the request was met. Intercom prices its Fin agent from $0.99 per "Fin outcome".
LessonUnder this pricing the vendor's own software decides what the customer is billed for.
Credit
Zendesk; Intercom
Watch for
Most requests handled by AI
Zendesk says customers of its Outcome AI Platform "can automate up to 84% of service requests" and will pay from a set budget for "eligible issues resolved and business work completed", with unlimited seats for staff. Zendesk itself "evaluates each outcome for quality and completeness".
Achieved whenan auditor that Zendesk does not pay samples conversations billed as outcomes at named customers and reports the share that solved the customer's problem.
Claimed by
Zendesk, November 2026
A directory for AI assistants
Zendesk published Open Service Directory, an open specification and public list meant to let people's own AI assistants find a company's verified support channel, whether or not the company uses Zendesk. Early participants include Betterment, ElevenLabs and Wireless Logic.
Achieved whenassistants from at least one other maker use the directory in public, and a test shows they reach the right channel more often than they do through web search.
Claimed by
Zendesk, November 2026
Open
01
Secrets in support attachments
A public test of the tools that are meant to strip secrets out of files customers attach to support cases. Make 500 synthetic attachments of the kinds support desks receive: HAR files, browser console logs, application log bundles, configuration exports and screenshots. Plant 1,000 fake secrets of known types in them (session cookies, bearer tokens, API keys, passwords) and 500 harmless strings that look like secrets. Run every sanitizer and scanner and count what each removes and what it wrongly flags.
The briefCredit · unclaimed02
What counts as resolved
An outside check of how help desks count the AI resolutions they bill for. Write a test knowledge base and 300 scripted customer conversations: 100 the knowledge base can answer, 100 it cannot, and 100 where the scripted customer gives up without an answer. Run them through the AI agents of help desks that offer trial accounts, then compare each vendor's count of billed resolutions or outcomes with what the script says happened. Publish the conversations with three people's labels as a test set for the language-model judges vendors use.
The briefCredit · unclaimed
Open 01 · the brief
Secrets in support attachments
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of the tools that are meant to strip secrets out of files customers attach to support cases. Make 500 synthetic attachments of the kinds support desks receive: HAR files, browser console logs, application log bundles, configuration exports and screenshots. Plant 1,000 fake secrets of known types in them (session cookies, bearer tokens, API keys, passwords) and 500 harmless strings that look like secrets. Run every sanitizer and scanner and count what each removes and what it wrongly flags.
Why it is still open
The Okta break-in ran through HAR files in support cases. Cloudflare's HAR sanitizer and another published on Google's GitHub exist, but no published catch rate was found for either, and the open secret-detection test sets found, such as hexared's on GitHub, are made of source-code and configuration files, not support attachments. A search found no test of sanitizers on screenshots or log bundles.
Built with what exists
a small test web app with fake logins, run by the tester; Chrome DevTools to record HAR files from it; the Cloudflare and Google HAR sanitizers; the open secret scanners TruffleHog and Gitleaks; Tesseract for reading text in screenshots; the redaction features of the help desks under test, on trial accounts.
First step
Stand up the test web app, record 50 HAR files with DevTools, plant 100 fake tokens, and run both HAR sanitizers on them.
Done looks like
500 files and the list of planted secrets published, and for each tool the share of secrets removed by type and file kind, with false alarms per 100 files. A second person re-runs three tools from the published files and gets the same counts.
Open 02 · the brief
What counts as resolved
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside check of how help desks count the AI resolutions they bill for. Write a test knowledge base and 300 scripted customer conversations: 100 the knowledge base can answer, 100 it cannot, and 100 where the scripted customer gives up without an answer. Run them through the AI agents of help desks that offer trial accounts, then compare each vendor's count of billed resolutions or outcomes with what the script says happened. Publish the conversations with three people's labels as a test set for the language-model judges vendors use.
Why it is still open
Zendesk has a language model confirm each resolution and Intercom bills per "Fin outcome", and neither publishes an outside check of those counts. Published comparisons of language-model judges with people, such as Purwar and colleagues' 2026 study of voice agents, score conversation quality and safety, not whether the problem was solved. A search found no public set of support conversations labeled resolved or unresolved by people.
Built with what exists
trial accounts for two or more help desks with AI agents, such as Zendesk and Intercom, used within their terms; a script or a language model playing the customer from fixed instructions; a spreadsheet of expected outcomes; three volunteer labelers.
First step
Write the knowledge base and 30 scripts, open one trial, run the 30 conversations, and read the vendor's resolution dashboard after its 72-hour window closes.
Done looks like
For each vendor, a table of scripted outcome against counted or billed resolution for all 300 conversations, and the share of unanswered or abandoned conversations that were counted. The 300 labeled conversations are published, with the labelers' agreement rate.