A pragmatic, opinionated read on what's actually shifting — and what's hype.
There's more noise in this category than in any other technology shift we've worked through. Part of that is genuine — the underlying capability is changing fast. Most of it is marketing dressed up as inevitability. This is our attempt to separate the two, specifically for the kind of business that hires us: service firms in the 20-80 person range, professional services or B2B.
We'll list three trends we think genuinely matter, then three we think most service business founders should ignore for now.
What matters: retrieval-augmented generation that actually works at scale
The technology has been around for two years. What's new is that it's finally cheap and reliable enough to build serious systems on.
The use case that matters for service businesses isn't a customer-facing chatbot. It's internal: every consultancy, law firm, accounting practice, and recruitment agency we work with has years of accumulated written work — proposals, deliverables, case files, memos, contracts. Most of that work is searchable in theory and unfindable in practice.
A well-built retrieval system over a firm's own work turns this from dead weight into a force multiplier. A junior lawyer can query "what's our standard position on this specific clause" and get back the actual language the firm has used, in context, with links to the original documents. A consultancy can ask "how did we frame the strategic options for this kind of client last year" and get its own prior thinking back, not a generic LLM response.
The reason this matters now and didn't 18 months ago: embedding costs have collapsed, retrieval quality has improved, and the engineering pattern is well-understood. A custom retrieval system for a 40-person firm is now a three-week build, not a six-month one.
What matters: voice-first document workflows
Every service business has people whose work happens in conversation — sales calls, client interviews, intake meetings, advisory sessions. The output of those conversations historically lives in notes, transcripts, or worse, nobody's head.
The combination of cheap transcription, structured extraction, and CRM integration changes this. A 45-minute client call now produces: a clean transcript, a structured summary, extracted action items, updated CRM fields, and a draft follow-up email — without the operator doing anything except being present in the call.
The reason this matters: the bottleneck in most service businesses isn't the work, it's the documentation of the work. Senior people don't write notes because they're busy doing the next thing. The next thing has historically been the cost of doing the first thing well. That equation just changed.
We've built versions of this for three clients in the last six months. The pattern is now mature enough to deploy confidently.
What matters: AI quality control loops
The hype version of AI says "automate this workflow with AI." The reality version says "automate this workflow with AI plus the systematic catch-and-correct loop that handles the cases AI gets wrong."
This is unsexy, but it's the actual unlock. Generic AI tools fail in the 5-10% of cases that don't match the average. Most service businesses can't tolerate that error rate on high-stakes work — legal documents, financial calculations, regulated communications.
What works in practice: a structured architecture where AI handles the bulk of the work, a clear confidence score on each output, and a human review queue for low-confidence cases. The human review feedback gets captured and used to improve the system over time. Done well, you get AI throughput on the easy 90% and human judgment on the hard 10%, with the system actually learning where its limits are.
This is what we mean when we say "AI implementation done properly." It's not a model. It's a workflow that includes the model.
What doesn't matter (yet): autonomous AI agents
The most-hyped category of the moment. The version that's being sold — AI agents that independently take actions on your behalf, navigate the web, complete multi-step tasks, coordinate with other agents — is mostly demos. The reliability isn't there. The error modes are catastrophic. The auditability is non-existent.
For service businesses where actions have consequences — sending client communications, executing transactions, making commitments — autonomous agents are currently a risk that exceeds the upside. The same outcomes are achievable today with deterministic workflows that include AI for the parts AI is good at. The "agent" framing adds nothing except risk.
Will this change? Probably, in two or three years. Right now, anyone selling you autonomous agents for production use is selling you a beta test you didn't sign up for.
What doesn't matter: most generative AI for client-facing content
Marketing copy. Sales emails. Proposal drafts. The use cases are real, but the lift is marginal once you've got a good template library, and the brand risk of letting AI write directly to clients without heavy human editing is significant.
The exception is internal first drafts that humans then edit. Useful. Not transformative. Most of the productivity claims in this category come from operators who weren't writing well to begin with, were never going to write better, and have now stopped trying to write well because they have AI for that. The work product gets worse, not better, because the AI baseline becomes the ceiling rather than the floor.
Service businesses compete on the quality of their thinking, expressed in their writing. We've never seen a serious firm gain real ground by automating its written communications.
What doesn't matter: "AI-first" platform rewrites
Every major SaaS vendor is in the middle of an "AI-first" rewrite of their product. HubSpot AI. Salesforce Einstein. Zendesk AI. The features ship; the value proposition for the buyer is mostly noise.
The pattern: the AI feature does about 70% of what the vendor demo suggests, on the use cases the vendor cherry-picked. For your specific use case, it does less. Meanwhile the product gets more expensive, the UI gets more cluttered, and the actual jobs you hired the tool to do get marginally better at best.
The right move for most service businesses is to ignore the AI features in their existing SaaS stack and not pay more for them. Use the tools for what they were originally good at. Build the AI capability you actually need elsewhere, on your own data, where you control quality.
The frame to use
When you read an AI announcement, ask three questions. First: does this work on your specific data or only on demo data? Second: what happens in the 10% of cases where it's wrong, and who pays for that? Third: are you the customer or the beta tester?
The trends that matter are the ones that survive all three questions. The ones that don't survive are the ones you'll read most about, because the marketing budget is correlated to the gap between the demo and the reality.
We'll revisit this list in three months. The half-life of "what matters in AI" is shorter than founders are being led to believe.