Skip to content
MakerGain
← Back to project

When AI Can't Answer, This Conversation Summary Can't Be Turned Off

Edited by MakerGain

My AskAI makes the conversation summary generated before an AI customer-service handoff impossible to turn off, so the human agent taking over gets the context directly; co-founder Mike Heap recalls that the company moved from general-purpose Q&A to customer service because customer service accounted for 70% of revenue at the time. Its Zendesk migration emphasizes being ready within 24 hours, backtesting with historical tickets, and running parallel tests with internal notes before charging by equivalent chat ticket; the TravelJoy case says the resolution rate rose from 24% to about 80%.

In chats or tickets where AI is enabled to reply directly, when a handoff to a human occurs, the person taking over does not see a blank ticket but an automatically generated conversation summary from the AI. My AskAI makes this summary impossible to turn off: when the AI cannot answer, the customer does not have to repeat what they just said. [1]

This is a company that provides AI customer service. Its product plugs into Zendesk, Intercom, Freshdesk, Gorgias, and HubSpot, so businesses do not have to move off their existing helpdesk. [2]

Image: source My AskAI actual image source[3], retrieved 2026-11-10; the original interface is retained.

That summary that cannot be turned off is the design in the product most worth pausing on. It does not explain how smart the AI is; it simply fixes the context needed for the handoff. The triggers for human escalation listed in the support documentation include the AI being unable to answer, the user asking for a human, user frustration, and preset rules; the summary's job is to make the preceding conversation clear so this baton pass is complete. [1]

First General-Purpose Q&A, Then Revenue Numbers Pulled Them into Customer Service

Co-founder Mike Heap recalled in Pinecone's RAG Brag interview that the company did not start with customer service. It was a general-purpose 'input your material, ask questions' product, used by students and law firms. The problem lay in two places: feature priorities were hard to sort out, and marketing was also difficult. Later they found that customer service accounted for 70% of revenue at the time, so they switched to it and rebuilt the product. [4]

This pivot sounds logical in hindsight, but 70% was an outcome, not a plan. A general-purpose tool was being used by different users, and in the end the revenue structure made the choice for them. I am more inclined to believe this kind of story of being pushed along by numbers; it is closer to what startups really look like than 'we identified a customer-service pain point.'

Technically, it also makes sense. Alex Rainey said in the same interview that they retrieved relevant snippets from websites and documents and fed them to the model. This explains why customer service is a fitting use case: the answers were already in the help center; if retrieval is accurate, the answers have sources. RAG Brag interview[4] They use Pinecone for vector storage and retrieval, and PortKey as the LLM gateway, handling fallback and response caching. This stack was as of the time of the interview; the interview did not give a publication date.

The words 'fallback' and 'caching' are worth a look: in the interview, Alex explained PortKey's fallback as a backup for when the primary model service fails, and response caching as a way to save time and cost. He also compared calling an LLM to dealing with a beast—powerful but unpredictable; change the prompt a little, and the output may change. This helps explain why they test in such detail. Mike also mentioned something more concrete: updates to OpenAI models significantly affect answer quality. Part of the product's quality is not in their own hands.

Mike was quite candid about customer acquisition: early on they got customers through AI newsletters and influencers, but the channels had short lives; finding repeatable distribution channels was hard; and there was a lot of noise in the AI space. RAG Brag interview[4] The About page on the official site also says that the two founders still personally do demos, onboarding, and human support; on plans that include human support, customers work directly with the people building the product. [2]

Wrong Answers Need Someone to Fill the Gap

The trouble with customer-service AI is not only at the moment of answering. If that item is not in the knowledge base and the AI cannot answer, after the ticket is handed off to a human, how does the knowledge base get filled?

What Self-Learning does is compare the AI's replies and the human's replies on escalated tickets, automatically draft knowledge articles, and fill the gap. About Us[2] Whether these drafts go through human review is not mentioned on the About page; it turns 'maintaining the knowledge base' from waiting for someone to write it to having the system automatically draft it first.

Tools goes further: it connects to the customer's API to look up orders, process refunds, cancel subscriptions, and update accounts. But there are prerequisites—the corresponding systems must be connected, and the rules set by the business must be followed. About Us[2] The FAQ on the pricing page adds: the technical team may need to create or update API endpoints, and Tasks and Tools also require the corresponding system connections and access rules. Pricing page[5] In other words, what the AI can do is bounded by what the customer has connected and allowed.

The triggers for human escalation are listed in detail: the user asks for a human, the AI cannot answer, the user is frustrated, preset guidance rules, and tagged topics. Intercom, Zendesk Messaging+Tickets, HubSpot, Freshdesk, and Gorgias support native handover; Freshchat can also hand off to a human, but email handoff is not supported. Support documentation[1] Being detailed enough to say that a particular channel is not supported—this does not read like marketing material; it reads like something written down after running into pitfalls.

Migration Is the Dirty Work

Understanding My AskAI only as 'an AI customer-service tool plugged into an existing helpdesk' misses what it is really selling. The migration page says: after migrating from Zendesk AI, once the required access permissions and setup details are obtained, testing can be ready within 24 hours; Zendesk tickets, macros, views, and help center are retained; first backtest with historical tickets, then run parallel tests in Zendesk Tickets using internal notes; Zendesk Messaging does not support note mode, so it has to be tested and launched separately. [6]

Not a single sentence in this process talks about how smart the model is. It talks about how to let something new come in without interrupting the existing customer-service operation. For a customer-service team that is already running, this is more persuasive than any resolution-rate number—backtesting, parallel running, launch; each step leaves room for the customer to decide.

The Dividing Lines Between the Three Tiers

Pricing has three tiers: Pro is $199 per month and includes 1,000 equivalent chat tickets per month; Scale is $499 per month and includes 2,000, with unlimited team seats; Enterprise starts at $999 per month and is aimed at 20,000+ tickets per month. Pricing is based on 'equivalent chat tickets': 1 conversation credit equals 2 AI replies, and a ticket with only one reply counts as 0.5 credit. [5]

Among the add-ons, API, Slack, and Teams access is a $49/month add-on on Pro and is included only on Scale; Tasks, Tools, and image replies cost $0.02 per reply; AI tagging costs $0.05 per tag and is unavailable on Zendesk Messaging and HubSpot. Pricing page[5] These conditions change the actual bill: teams that need automatic categorization pay by the number of tags; to use My AskAI's API, Slack, or Teams access, Pro requires an additional $49 per month.

Numbers and Boundaries

The company describes itself as bootstrapped, profitable, and growing, and says it does not need outside financing. The About page on its official site also says it serves 200+ businesses, processes 200,000 tickets per month, has uptime above 99.99%, has SOC 2 Type II, and Enterprise supports HIPAA; the pricing page says 250+ businesses and 1.9 million+ tickets resolved cumulatively. [2] · Pricing page[5] These are all the company's own statements, with no statistical period or audit details provided. Profit remains unclear—the costs of model calls, vector storage, scraping, and labor are all undisclosed, and how profit is calculated is not explained either; 'profitable' remains the company's own claim.

In the customer case, TravelJoy's path is fairly typical: it first tried Zendesk Advanced AI for several months, did not meet expectations, and switched over. It retained Zendesk Ticketing and Messaging; knowledge comes from website syncing and internal Notion; Messaging had direct replies enabled from day one; email tickets use Copilot drafts and only reply to the first message. At launch, the AI resolution rate was about 76%, now about 80%, AI CSAT 86%, covering 2,500–2,700 tickets per month, with more than 24,500 resolved cumulatively. [7]

The same page also gives two time-saving figures: the opening and overview say it saves 193 hours per month, while later text, calculating at 5 minutes of handling time per ticket, gives 205 hours. Both numbers come from the vendor's case, and the method, sample, and statistical period are not fully disclosed.

This case was published by My AskAI itself, so the numbers have to be viewed with that scope in mind. But at least in this case, the customer did not start using AI from zero; it moved over from another AI.

The hardest part of this business is getting customers to dare to hand over the front line. My AskAI breaks migration into steps where customers can first backtest and then decide whether to open up direct replies, but the risk of switching systems is still borne by the customer.

Continue reading

AI supportAI searchAI commerce
Customer Support Bots Start Looking Up Orders for You, and Chatwith Bills “Talking” and “Acting” Separately

Chatwith bills customer-support answers and external actions separately: messages use credits, while actions successfully executed through MCP are billed separately; plans start at 9 US dollars, and an extra 1,000 MCP actions cost 6 US dollars. The founder says he put about 21,000 US dollars of AppSumo revenue into Google Ads in 2024 and at one point a Google Ads bill went unpaid; TrustMRR shows MRR of 3,622 US dollars and 42 active subscriptions in November 2026.

AI supportAI productivity
Card first, then entry: ChatDash has turned payment collection for AI voice agents into an annual-fee business

ChatDash turns a white-label AI agent dashboard for agencies into a business selling customer slots annually: customers must first add a card before entering the dashboard, and Stripe automatically handles subscriptions and usage-based billing; SaaS Mode is limited to a Growth add-on and supports only Retell voice and OpenAI text. As of 2026-11-14, the TrustMRR page reports cumulative revenue of $603,545 and MRR of $17,167.

AI supportAI searchAI commerce
When a Visitor Clicks Handoff to a Human, Chat Data Sends a Colleague an Email with a Login-Free Link

Chat Data trains company documents, webpages, and files into an AI customer-service agent that can be deployed to websites and messaging channels, and uses a login-free email link to connect visitors who ask for a human. It charges by message credit and supports custom models and MCP hosted by Composio; the TrustMRR page lists MRR of 17,645 US dollars and 304 active subscriptions, with a team of 2–5 people.

AI supportAI search
Nine Assistants, One Subscription: What CustomGPT.ai Sells Is “Answers That Point Back to the Page”

In CustomGPT.ai's anonymous pharmacist association case, nine assistants divided by department content share one entry-level subscription; in their busiest month they handled 1,006 questions in total, with zero support tickets over three years. The product retrieves only from published materials and attaches a source link to each answer, turning “adding an assistant” from procurement into configuration. Standard costs 99 United States dollars per month and Premium costs 499 United States dollars per month, tiered by number of agents and query quota, similar to DocsBot's tiering logic.