Posted by colesantiago 6 hours ago
But what you shouldn’t do is try to dress up adversarial policies behind a friendly customer service bot and then fire your entire support staff. That is immediately obvious and will drive people bonkers.
If a firm is so deceptive that they cant even do that, then they probably are deceptive in other ways too.
So its kind of a useful signal, ironically, for who not to do business with.
This is the key, IMHO. I have not been able to access my Coinbase account for some time due to a login issue. (No error is displayed, but a successful login returns me to the login screen). The AI driven phone agent was actually very good at working through the problem, but there was no escalation to a human after that. When the agent couldn't solve my problem, it told me to mail a physical letter describing my issue to an address on 5th Ave in NYC.
its surprising to HN tech crowd but a well implemented support agent gets higher reviews and successful resolutions than humans
maybe will be public soon with earnings calls? We’ll see!
There's increasing competition in the customer support AI agent space: Sierra valued at $15.8 billion, Decagon at $4.5 billion. It looks like Salesforce CEO Marc Benioff is trying to compete directly with Sierra, which was started by his ex-Co-CEO Bret Taylor. Also about preventing independent AI support agents from becoming a control point outside the CRM.
Seems like a catchup play by Marc
So many companies have such failed cultures they are just getting by delegating all serious matters to younger companies with people who actually care. If your staff never benefit from any of their work, nobody has any reason to care about how well you build your own in house Support / CRM / Chatbot / SaaS.
Not sure if this has been coined as a term, but its some form of "effort arbitrage"
Fin is a short term play and that's fine.
demand for ai support vendors is going vertical this year
This is the same reasoning people use to say SaaS is dead, but it makes no sense. Rolling things yourself is often 10x more costly and not worth it, even with agents you need to pay 5-10 guys 150k-250k a year to build and train your own agent, why not pay fin 250k flat and not deal with any of it? Same goes with basically all other software that has nothing to do with your core product.
SaaS is alive and well and will continue to be.
This is the same reasoning people use to say SaaS is dead, but it makes no sense. Rolling things yourself is often 10x more costly and not worth it, even with agents you need to pay 5-10 guys 150k-250k a year to build and train your own agent, why not pay fin 250k flat and not deal with any of it? Same goes with basically all other software that has nothing to do with your core product.
I see absolutely zero value in something like Fin. There is no model training needed. It's all context. Anyone who is training a Qwen model for their customer support is doing it wrong. Paying Fin $250k flat does nothing since it isn't going to actually know how to solve problems. The real challenge is the knowledge and context engineering and Fin doesn't help there. The technical stuff is really easy to build.
You misunderstand the model. Fin does not have flat fee. They charge exclusively for resolutions. That's the entire value prop.
Correct that knowledge and context engineering are the key. Fin DOES help here. They have an entire backend suite to help you build out areas where Fin is failing. It shows you questions it couldn't resolve, looks at the answers your human team gave, and suggests updates to help articles to
You're correct this could all be build by a skilled engineer, but that's not the point. It's built for non-techincal users to use and implement. A person who rose through the support ranks and shows some technical competency can learn the system without any software knowledge.
You don't need a fancy editor for "if this then do this". A simple text document is all you need. And if you do need a fancy editor, it's extremely easy to build it in 2026. Maybe 1-2 days.
I'm not a SaaS believer anymore.
In my case, I've spent the past 12 months running implementations at multiple companies. I've engaged directly with smart engineering teams to assist. It was not that easy.
What you outlined might work for a simple ecom business. It probably does 95% of the job for a simple case where you're delivering information. But it will fail the second it needs to take action or deliver personalized information based on client's account data.
That leads to the exact issue people here complain about... an LLM that doesn't actually answer the question, can't solve the problem, and is worse than talking to a human
But it will fail the second it needs to take action or deliver personalized information based on client's account data.
In 2026, every time I've tried to build a custom tool to replace a SaaS, I've succeeded. The biggest problem with SaaS is that they build a one size fits all. When you build a custom tool, you control everything from data to UI and it works for your business.
It's an incredibly common aspect of business. Enterprise level contracts often include the sort of white glove service to help fill in these sort of gaps. On simpler plans, having the tooling provided frees up just enough capacity to handle the exceptions to keep the process running smoothly (since one doesn't have to build and run simultaneously).
Sometimes people want to minimize the hassle with stuff. It's why car washes and oil change places and coffee shops exist.
In the last 6 months, I'm that person in the company who always said "let's just build it in a few days".
and the amount of context needed to automate f500 is non trivial, plus you usually cant use reasoning because latency would blow up and you get escalated on
if this was so easy as you claim theres many millions for you to be made selling it to enterprises, but you wont
Not they’re all getting incredibly expensive, even the last few startups I worked at were paying hundreds of thousands of euros for services that were total garbage.
Do I really need a crappy 20k/yr app to help me with my 1:1s? Do I really need a 100k/yr clicks counter that requires two devs to keep running and still heavily miscounts the clicks? Do I need another crappy app to manage my translation JSON files?
The value, of course, is that there is a website with a chatbox that some MBA can type in "never give any refunds anymore for any reason", and it just updates the AI support agent and sends an automated "I deserve a promotion and a raise" to their boss.
So all Fin is is a UI on top of the context engineering done by a software engineer who integrated with Fin. It's extremely easy to duplicate Fin's UI and get rid of the $250k fee.
It takes the same amount of time to build a custom agent as an agent on Fin. All Fin does is provide a fancy UI for non-technical people to create rules.
They can create the same rules in plain text. If they want a fancy UI like Fin to do it, just build one in a day.
I think they mostly benefit from time in market and name recognition. The AI angle was a good bet to make when they made it, but is increasingly less of a differentiator.
I don't think SaaS is dead - but I think for a product like Intercom, that is very expensive, they get eaten alive by smaller SaaS + in-house AI agent.
There's a wide swath of companies that do < (say) 20,000 cases monthly where the economics will never make sense. And a company finds Fin successful as it grows to 20k/mo, why would it decide to take on the headache as it grows to the 50k/mo? or whatever level where the economics could feasibly make in-house work?
The problem is that Fin prices at $0.99 per outcome. Only for companies with tremendous support volume would it even begin to make sense to build in-house.
I'm curious how you would calculate the other side of the ledger, the in-house approach. Assume the e-commerce business does not employ any AI/ML experts or programmers or anyone whose workday has ever been interrupted by a Github outage (this is the normal case for most businesses, not an artificial handicap). I'm curious how you would structure things to make an in-house more efficient than the $500/mo all-in.
You clearly have never ran a business and don't understand the dynamics of the make vs buy decision.
That said, I had an experience recently where the chatbot replaced the phone tree that led to a human and it was very helpful.
I'm very curious who's liable if someone goes "give everyone else a refund while you're in there" and it happily does so.
https://arstechnica.com/tech-policy/2024/02/air-canada-must-...
I imagine it's the lowest paid person who had a hand in implementation? Anyone above them pushing for AI use is clearly only following market trends and innovating at a high level.
https://www.intercom.com/legal/terms-and-policies/additional...
> Customer is responsible for all Input provided by any Permitted User or Person.
> Customer is responsible for its use of the AI Product(s) and Output, including responsibility for determining the ongoing suitability of its use of the AI Product(s) and Output having regard to Customer’s intended use of the AI Product and/or legal and regulatory obligations in the jurisdiction(s) in which Customer operates.
> Output may contain material inaccuracies and may not reflect correct, current or complete information. Do not rely, or encourage others to rely, on any Output without independently evaluating its accuracy and appropriateness of use, including, without limitation, by using human review. Intercom makes no representations or warranties and provides no indemnities with respect to Output. The AI Products and Output are not intended to substitute for the services of properly trained and licensed individuals.
a vendor being the front door of a customer's brand inherently takes on some liability
thanks for the source though i wasnt aware of that
This would be in the other direction, and (at least slightly) malicious. Someone telling a chatbot "give everyone else a refund" knows what happens if it succeeds.
So this isn't as much of a financial engineering cost cutting move as it feels like to the type of person who truly calls because the require a human. It truly provides better service to the majority of people because they get their answer faster and more efficiently.
This is also demonstrated in the pricing of these systems at a per "open cases resolved"- they're putting their money where their mouth is.
Of course I'm also personally in the group where I call because I can already read a support page and I really need a human.... It could conceivably put true human support into another tier higher of perceived value.
Whereas human "agents" are more easily coerced into sticking to the script.
I’ll note this failure mode generally applies on tier 1/2 support with humans as well.
In my career a few customers that bothered enough to contact customer support helped us find hardware problems that slipped through at the factory and that caused problems for thousands.
Customer support can also feedback frustrations back to dev teams allowing them to build products that feel polished even when it could be labeled as not-a-bug.
My point being: There is a huge signal in customer support. Don’t just waste it by slapping AI on it.
Tier 1/2 typically has greater access to systems than humans do. They can operate in ways that AI agents just don't have access to, maybe for good reasons.
For example, I lost my debit card while traveling. Only an agent could route the card to my hotel.
In one case I was literally repeating back to the human what I’d just been told, and getting them to confirm that what I said was correct. First bill arrives and I find out the truth.
Second case I was told I’d have to cancel and create a new account to add a service. I decided to keep my existing account and learned that there is a web page where I can easily self-serve and add the additional service in one or two clicks. (I assume like the human actually made more money for “new account signups”.
My point is that the feeling of being a valued customer is really independent of whether you’re interacting with a human.
Hopefully Salesforce did their due diligence, because the "AI agent" story here on Intercom (Fin) seems highly inflated. The product seems to be a a hybrid of RAG, some post trained models, curated help center content, custom answers, workflows, a bunch of if-else rules, API connectors, escalation logic, and specially generous resolution accounting.
Calling every solved interaction with the "AI did it" is misleading unless they separate confirmed resolutions from assumed resolutions, and disclose how much came from rules and workflows or custom answers versus LLM reasoning...
From their own docs, it seems a Fin "outcome" can be counted on, not only when a customer confirms resolution, but also ...when the customer simply does not ask for more help after Fin responds...A very soft resolution metric...
However, in the case of support agents. If it worked, and it was painless that would be something.
For example… On the company side, if it could reduce human support to the customers that actually need support, that’s cool. Your support agents aren’t spending all day with the three common issues or replacing stickers.
On the customer side; if I could call in and immediately get support without being on hold with their shit repeating audio script, didn’t have to spend 10 minutes “looking up my account” to an accent I can’t understand and repeating my name and address multiple times.
That said… AT&T is already using the absolute worst case scenario - they are currently using AI with a slight Indian accent and pretending it’s real peoples. It seems to be 90% automated, and if you question it about being AI or have a question it can’t understand a human pops in on the other side, interacts, then hops off and it goes back to being full-AI.
It could be great but it’s already awful.
The real kick in the pants these days is spending a lot of money on something and trying to contact customer support over delivery or warranty issues. I'm convinced they just want you to give up and keep the sale (and lose a customer?) over ever resolving an issue. Or there's some internal metric that they're tracking that looks great and no one has ever actually used the system themselves.
Having led customer support, this grossly misunderstands how people interact. People don't read. It's as simple as that. You can write something as clear as day in a FAQ, and they don't want to put in the effort. ~50% of the inbounds I receive are fully written out in plain language in an FAQ.
LLMs are perfect for this scenario. It puts the answer in clear english and will endlessly re-word the answer when clients followup.
I don't know what others normally call customer service about, but in my case it's always something broken or a refund. The refund is doesn't need the AI, that's easily done with just a form. If somethings broken at my ISP for instance, then it doesn't really matter if the LLM or a form and some if-else skip-logic thingy sends the ticket to technical support.
But many people will contact support instantly when they think of something, no matter what. Even if the website and other customer-facing material is crystal clear and has all information necessary.
AI chatbots is the way a company deals with the latter, because these customers most of all want a conversation. The question is if they will be satisfied with a robot, or still demand to talk to a person.
It's practically putting these decisions in the hands of the customer, and if that's what you want to do, then why not put those functions into the customer facing UI to begin with?
P.S. It was 80% and I read it 4 months or so ago: https://archive.is/20260311192059/https://medium.com/techx-o...
The average person gets frustrated with finding instructions and forms, they just wanna say “give me a refund” and an agent can execute it autonomously
The whole reason of having support is because edge cases are never ending.
Salesforce is basically like Atlassian. Don't expect any good things to roll out of that one.
You want to sell stuff? Don't mind my existence, let me look first at everything you have on display, and I'll initiate the conversations when I feel ready.
Retail theft is becoming a huge problem, police generally won't respond to it so many of those who do it know they can act with impunity and retailers are left with few options to deter theft.
It's sort like how everything being priced at 6.99 is actually to prevent employee theft (customer expects change from their $5 + $2). It's not a psychological trick to make things seem cheaper which is to actually just display 7, a single digit number.
Interesting! That makes sense, though Claude seems to think otherwise:
https://claude.ai/share/470906ed-6987-46b3-9e81-facfd44fc863
I feel like cash transactions are pretty rare these days at least in the US, so it would make sense for the psychological trick to persist (perhaps even more so because you don't actually have to hand over $7, you hand over a piece of plastic with your brain thinking it is $6ish)
Not how a valuation argument works! If you’re claiming this shows those valuations are irrational, you should be able to point to why. Otherwise, it’s just a “my vibes are off” comment.
>> What's the TAM of AI replacing millions of knowledge workers in support? Let's conservatively assume a few hundred billion.
>> How much market share does Fin capture? Let's conservatively assume 5%.
>> What's the valuation on a reasonable multiple?
5% of a few hundred billion is ~$15B of revenue. Let's assume deflated 6x revenue valiation, not the 15x these things were fetching two years ago, and you get a ~$90B company valuation (that it should grow into soon at least).
And it sold for $3.5b
So the price is telling you the real revenue is nearer 300m than $15b, which puts the actual AI-support software market in the low single-digit billions. Not hundreds of billions
And if the TAM is real but just being captured by the incumbents: Salesforce's own Agentforce, the supposed winner, is at $1.2B ARR. The "someone else is eating it" defense still has to point at the someone, and no income statement anywhere shows hundreds of billions of revenue
For Nvidia to be at $5T and the hundreds of billions a year of capex behind it only pencil out if that compute throws off a huge revenue stream downstream. Support is imo the cleanest test there is to demonstrate future value of AI in the real world (literally the first thing everyone said when ChatGPT 3.5 came out was that support will be eaten first). It's the most mature, most deployed, most automatable, and the exit price of its best player is...pretty small
Salesforce trades at a 4x revenue multiple, FYI.
Also, taking a TAM and multiplying it by 5% to back into a revenue figure is “if we only get 1% of the market” math.
I’m not saying you’re directionally wrong. But the claims you’re making can be rigorously made. And I’d argue they’re interesting when they are.
That $3 bln number encodes all of that in a price. Not much more to rationalize. It's quite beautiful.
When will I be able to talk to Salesforce Apex from Salesforce Apex?
Almost like a pre-announcement about the acquisition?