Most AI agency pitches are vague and overpriced. Here's how to evaluate them, what to ask before signing, red flags that cost you money, and what fair pricing looks like.
Hiring an AI automation agency means contracting a team to design, build, and often maintain AI-powered workflows or products for your business — as opposed to hiring an in-house engineer, buying off-the-shelf SaaS, or trying to DIY. You get speed, specialized expertise, and a system you own. When it works. The problem is that "AI agency" now covers everyone from people who genuinely build production systems to people who learned Make.com last month and built a landing page.
Here's how to tell them apart before you sign anything.
A good AI automation agency runs a scoped discovery process before quoting you anything. They ask about your current process, your inputs and outputs, your tech stack, and what done looks like. They produce a spec — a plain-English document describing exactly what will be built — before any build begins. They hand off documentation when the project ends. You own everything: the code, the workflows, the documentation.
A bad AI agency sells you a retainer for "AI strategy" with vague deliverables. They produce slide decks, not systems. They quote you $5,000 without asking a single question about your process. Their "case studies" describe outcomes in percentages with no underlying data. They structure contracts so that you need them to make any changes — no source access, no documentation, no export.
The tell is always the discovery process. An agency that won't spend 30–60 minutes understanding your specific problem before quoting you doesn't have a repeatable delivery process. They're winging it.
Ask these in your first serious conversation. The answers will tell you most of what you need to know.
1. Can I see the exact deliverable I'll own? The answer should be: "Yes, here's a spec from a similar project" or "Yes, we'll produce a spec for your project before the build contract starts." If the answer is "we'll figure that out as we go," stop the conversation there.
2. What tools and APIs will you use and what's my ongoing cost? You should get a clear answer: "We'll build this on n8n, connecting HubSpot and Apollo. Your running cost will be $180–$250/month for API credits and hosting." If they can't tell you what your ongoing costs will be, they're either inexperienced or structuring the engagement so you have to come back to them.
3. Have you built this specific workflow before? Not "AI workflows in general." This specific one. If they have, they should be able to tell you what went wrong on the first version and how they fixed it. Real experience includes failure stories. Vendors who've only done it perfectly haven't done it much.
4. What does maintenance look like after the build? A good answer: "AI models update and APIs change, so we offer a maintenance retainer for $X/month that covers monitoring and updates. If you don't want a retainer, we'll document everything so your team or another vendor can maintain it." A bad answer: "You'll need to come back to us for any changes."
5. What metrics will we track to know it worked? They should have an opinion on this before you do. "We'll measure time-to-response on inbound leads, lead-to-meeting conversion rate, and hours saved per week compared to the baseline we document at kickoff." If they can't define success in advance, they can't be accountable to it afterward.
Agencies that can't give you a plain-English spec. If they can't describe what they're building in plain language, they either haven't scoped it properly or they're trying to avoid accountability for what gets delivered. A spec doesn't need to be technical. It should read like a business description of the workflow.
Agencies that bill by the hour for "AI strategy." Hourly billing for strategy work with no defined output is a structurally bad deal. The clock runs. You don't know what you're getting. Good agencies charge for outcomes — per project, per workflow, per milestone — not for their time.
Agencies whose case studies have no numbers. "We helped a financial services firm improve efficiency" means nothing. "We automated a 12-person ops team's lead qualification process, reducing their research time from 20 hours/week to 3 hours/week, with a 4-month payback on the $18k build cost" is a case study. If the numbers aren't in there, ask for them. If they can't provide them, move on.
No IP transfer clause. Any vendor who won't transfer full IP ownership to you upon project completion is building a dependency, not a solution. You should own every workflow, every script, every API connection, and every document when the project ends.
Discovery: $500–$2,000 depending on complexity. This is the scoping and spec phase. Some agencies roll it into the build cost. Others charge separately. It's worth paying for separately because it forces accountability — if the spec doesn't match what you asked for, you haven't started a $20,000 build yet.
Build: $3,000–$60,000 depending on complexity. Simple single-workflow builds (inbound lead routing, report automation) run $3k–$8k. Full department automation systems run $25k–$60k. The ranges I've seen in the market are in the AI automation agency pricing breakdown.
Maintenance: $300–$2,000/month. Justified if the system runs daily on live data and a failure affects revenue. Not justified for a weekly reporting workflow that can sit broken for a few days without costing you anything.
Milestone-based, not hourly. Tie each payment to a specific deliverable: discovery complete + spec delivered, build complete + testing passed, handoff with documentation. Never pay a final milestone before you've accepted the delivered system.
IP ownership clause. "All intellectual property, code, workflows, and documentation created under this agreement transfers to [your company] upon receipt of final payment." If they negotiate this clause, walk away.
Source code or workflow export. For n8n workflows, you want the workflow JSON file. For custom code, you want the repository access. For Make.com or Zapier, you want the scenario exported. Whatever they built, you should be able to move it to another platform or hand it to another developer without going through them.
SLA for the build phase. When does it deliver? What counts as acceptance? What happens if it misses the deadline? These should be in the contract, not discussed later.
---
If you want a second opinion on a proposal you've received, or want to understand what a fair quote looks like for your specific project, book a 30-minute call at cognival.co/book.
30-min strategy call. No pitch, real look at your stack.
Book a strategy call →