TL;DR: Nine RAG and agent builds in the $3–10k, four-to-six week band that sit past Zapier's ceiling and below enterprise consultancies: internal docs Q&A with source citations, customer-facing docs agent with escalation, sales-call summarizer to CRM, contract or RFP analyzer, lead enrichment and scoring agent, support ticket triage and draft, quote drafting assistant, voice-of-customer aggregator from reviews and tickets, and an internal data analyst that answers questions against your warehouse. Most agencies will not touch these because they are too technical for no-code shops and too small for enterprise. They are exactly the builds a small studio with engineering depth ships well.
There is a gap in the AI services market. Zapier handles anything simple. Enterprise consultancies charge $100k+ and quote twelve weeks. In between sits a band of genuinely useful agent and retrieval work, $3k–$10k, four-to-six weeks, that most agencies will not touch because it is too technical for no-code people and too small for enterprise shops.
These are the use cases in that band. All of them are projects we have shipped or would ship. None of them require a PhD or a team of ten.
1. Internal docs Q&A with source citations
What it does: Your team asks questions against your internal wiki, Slack history, Notion, and past support tickets. The agent answers with quotes and links back to the source doc, so nobody trusts the answer blind.
Scope: $4–7k. A simple RAG pipeline: ingest your docs, embed, store in pgvector or a vector DB, a small Next.js or Slack app on top.
Why most studios won't: Knowing what to embed and how to chunk is judgment work. Most agencies stop at "ingest everything at 512 tokens" and the retrieval is garbage. You have to know which sources to exclude (email threads, draft docs), how to rerank, when to cite, and when to say "I do not know."
Who it is for: Support teams, sales teams, onboarding for new hires.
2. Customer-facing docs agent with escalation
What it does: A chat widget on your docs or app. Answers product questions from your actual docs. If confidence is low or the user asks for a human, it escalates to support with context attached.
Scope: $5–9k. Similar to internal Q&A but with stricter output validation, safe fallbacks, and a clean escalation path.
Why most studios won't: Hallucinations in a customer-facing context are reputational damage. Building the confidence scoring, the "no answer" path, and the escalation UX is a real week of work. Most shops ship the happy path and skip it.
Who it is for: SaaS companies with product complexity, documentation-heavy tools.
3. Contract review and redline assistant
What it does: Upload a vendor contract. The agent compares it against your standard terms, flags deviations, and drafts proposed redlines. A lawyer reviews and sends. Cuts legal review time by 50%+ on standard vendor work.
Scope: $6–10k. Needs careful prompting, example-based learning, and a good review UI.
Why most studios won't: Legal-adjacent work is scary. The output has to be reliable and clearly "draft only." Building the workflow so lawyers trust it is a different skill from the RAG layer.
Who it is for: Companies with in-house counsel, law firms with standardized practice areas.
4. Multi-source sales research
What it does: Drop a company URL. Agent pulls recent news, funding, job postings, tech stack, recent social posts, and generates an ICP fit score plus a talking-points brief for the rep.
Scope: $4–8k. Requires web scraping (Firecrawl, Apify), LLM synthesis, and a Slack or Chrome extension front-end.
Why most studios won't: Scraping is fragile. Anti-bot measures change weekly. You need to build in fallbacks and a manual-override queue. Most shops underestimate the ops cost.
Who it is for: B2B sales teams with an ICP worth researching — mid-market and up.
5. Content ops agent
What it does: Takes a content brief, pulls relevant internal context (your past posts, your brand voice doc, your cases), drafts an outline, drafts the piece, flags weak sections for a human writer. Ships to a review queue, not straight to publish.
Scope: $5–8k. Needs a voice-matching layer, context retrieval, and a structured-output loop so the draft follows your content template.
Why most studios won't: Voice matching is the hard part. Generic GPT output is obvious. You need a fine-tuned prompt or a small tuned model plus a style guide retrieval layer. The ones that do it poorly ship AI slop.
Who it is for: In-house content teams shipping 8+ posts a month.
6. Email triage with summarization and draft replies
What it does: A shared inbox (support@, hello@, legal@) gets triaged by the agent. Classifies intent, summarizes the thread, drafts a reply based on your past responses. Human hits send or edits.
Scope: $3–6k.
Why most studios won't: Gmail API auth scopes, threading model, and the draft-and-review UI are all small pieces of friction that add up. Easy to start, harder to finish well.
Who it is for: Small teams where the founder or a single person is drowning in inbox.
7. Structured extraction from unstructured intake
What it does: Form responses, intake calls, or chat transcripts come in. Agent extracts structured fields — budget, timeline, key constraints, stakeholder names — and writes them to your CRM. No more typing the same info into a different shape.
Scope: $2–5k. One of the cheapest wins, one of the most useful.
Why most studios won't: They will, but they do it with Zapier AI steps, which gets expensive fast and is opaque when it fails.
Who it is for: Agencies, consultancies, service businesses with inbound inquiries.
8. Personalized onboarding agent
What it does: New user signs up. Agent personalizes the onboarding — first email, in-app checklist, sample data, first-meeting agenda — based on the user's ICP score and stated use case.
Scope: $4–7k.
Why most studios won't: The branching logic is where this goes wrong. Most shops either ship three hard-coded paths (not personalized enough) or 20 paths (impossible to maintain). Getting the branch depth right is judgment.
Who it is for: SaaS products with self-serve onboarding and enough volume to justify the build (1,000+ signups/mo).
9. Agent that books time in your calendar from Slack
What it does: Someone DMs the bot "can you find 30 minutes next Tuesday with the design team?" Agent checks calendars, finds the slot, proposes it, books when confirmed. Like Calendly for internal coordination.
Scope: $3–5k.
Why most studios won't: Calendar API edge cases are nasty. Timezone handling, recurring meetings, busy/free differences. Easy to prototype, hard to productionize.
Who it is for: Fast-moving teams of 15–100 where scheduling is a tax.
Why these projects are in a gap
Four reasons most agencies will not build these.
1. They require real engineering. Not developers, engineers. Retries, observability, structured outputs, evaluation. You cannot ship one of these in Zapier and have it work.
2. They require ops taste. Knowing where to put a human in the loop, where a model is allowed to send, how to handle failure gracefully. This is judgment that takes years of building automations.
3. They require product taste. The UI that someone actually uses to check the agent's work decides whether it gets adopted. A good agent with a bad review UI gets bypassed in a week.
4. They are not easy to sell. A $5k custom AI build has to be justified with real ROI math. Generic agency sales teams cannot have that conversation. You need someone who has actually shipped these and knows the payback window.
How we scope these
Every project above starts with a one-page scope doc: inputs, outputs, volume assumptions, failure modes, who sits in the loop, what we ship. Nothing about models, no vendor names, no stack argument. Just: what does this do, where does it run, how do we know it is working.
If you have a workflow in one of these categories and want to know what it would cost, tell us what you are doing today, how often, and what the pain is. Two-day turnaround on a scoped proposal.