1.5y in B2B AI - here is what I see working
and here is what I think is VC selling "the future"
I’ve been building agentic systems for over a year now. And selling them. I hear quite a bit about what others are buying in the AI space and what they are churning from.
tl;dr
There is a few use cases you probably heard about. They are working and people are buying the tools to get them done. Then there is a long tail of use cases which have 20 startups chasing it but very few companies are committing to those tools.
Everyone in B2B is buying
coding (Anthropic, Cursor, Openai) for their employees
general purpose AI (claude, chatgpt) for other functions like marketing. If employer doesn’t pay, employees use it privately and copy paste
Why everyone? This take no setup, is the highest ROI and kind of the starting point on AI for an org. Let’s people play around with MCPs etc.
Many in B2B are buying
integrated customer support tools that automate
triage of support tickets
suggest best customer success actions
automate responses to tickets
marketing tools that automate
personal outreach
enrichment of leads
tracking where brand shows up (e.g. octolens.com )
Why many? They hear from other companies that this saves $$ or headcount. Especially in EU cost cutting seems to be a motivational driver. In the US growth still plays a bit more of a role.
Some in B2B are buying
general purpose agent platforms
every employee can build their own slack “agents” that have access to certain data sources (e.g. dust.tt)
broad AI intelligence tools
at scale indexing of e.g. all customer interactions (e.g. nextapp.co)
specific niche tools like
specialized code review tools (e.g. coderabbit.ai)
specialized marketing asset generation tools
Why some? Seen as experimental. The cut-off points of the products are unclear. Some go broad, some go narrow. Some go concierge some go “build you own agents”. Unclear what works for who and what is here to stay.
There is a lot of opportunity to apply the capabilities provided by models like codex 5.5 and opus 4.7 to existing workflows.
We definitely do not need better models. The work now is to figure out product UX and scope that makes these AI applications usable without having 4 forward deployed engineers and 10k a month in token budgets.