The sales process for AI work has a specific problem: the person presenting is almost never the person building. We have seen what that gap looks like on both sides, when it goes well and when it goes badly. These are the signals that show up before you sign.
Why the Agency Sales Process Is Structurally Misleading for AI Work
Most agency sales processes were designed to sell web design or marketing retainers. AI integration work is fundamentally different. It requires upfront feasibility analysis, architecture decisions, model selection, and failure-mode planning before a single line of code is written. The standard pitch deck format hides all of that.
The Salesperson–Engineer Split No One Admits
At most agencies, the person who wins the contract is not the person who builds the thing. Senior salespeople and account directors exist specifically to communicate competence. They are polished, technically fluent in conversation, and very good at making ambiguous capability sound like certainty. The engineers are usually in the background, on a call if you’re lucky, absent entirely if you’re not.
For a brochure site, that split is manageable. For an AI integration that touches your CRM, your customer data, or your fulfillment workflow, it’s a structural problem. The salesperson cannot make reliable commitments about what the engineering team can actually deliver.
Ask early: “Will the person I’m talking to right now be on the delivery team?” If the answer is “we’ll introduce you to the team after we scope the project”, that’s your first flag.
Why AI Projects Fail at a Higher Rate Than Web Projects
McKinsey’s 2025 State of AI report identifies inadequate upfront feasibility alignment as a leading failure mode for AI projects. A separate survey found that 38% of businesses lack the technical expertise to evaluate AI deployment proposals they receive. That’s not an indictment of buyers, it reflects how new the category is. Agencies know this, and the less scrupulous ones structure their pitches to exploit it.
A web project fails in visible ways: pages look wrong, links break, the checkout doesn’t work. An AI integration can appear to function while doing the wrong thing, returning hallucinated summaries, mis-routing tickets, silently skipping edge cases. The failure modes are harder to detect without technical oversight, which makes the initial vendor selection more consequential.
Red Flags in the Proposal and Discovery Phase
The Proposal Arrived Within 24 Hours
If you sent your brief on Monday morning and a 30-page proposal landed in your inbox by Tuesday, one of two things happened: they sent a templated document with your name swapped in, or they didn’t think seriously about your problem. Either way, you don’t want that team building anything that runs in your production environment.
Scoping an AI integration takes time. The agency needs to understand your existing systems, data structure, edge cases, and failure tolerance before they can make honest commitments. A rapid-fire proposal is a proposal that skips all of that.
Discovery Is a $15K–$30K Line Item Before Any Code
Some agencies charge a substantial discovery phase before committing to any architecture or deliverable. That’s occasionally warranted, complex enterprise integrations may genuinely need three weeks of investigation. But for most SMB AI projects, a paid discovery phase that produces a document rather than running code is often a way to get money in the door before the real scoping conversation happens.
Ask what the deliverable of discovery is, and whether it obligates them to a fixed-scope build at the end. If discovery is open-ended and the output is a “recommendations report,” be skeptical.
The Proposal Is Long but Has No Architecture Decisions
A 30-page proposal that doesn’t specify which AI model will be used, how your data will move between systems, where the integration will live in your tech stack, or what happens when the model is wrong, is not a technical proposal. It’s a sales document formatted to look like one.
Credible AI integration proposals include at least preliminary architecture thinking: the model or API being used and why, the data flow from input to output, the fallback behavior when outputs are incorrect, and where human review sits in the workflow. If none of that is in the proposal, they haven’t thought about it yet.
Red Flags in the Technical Conversation
They Can’t Name the Model or Explain the Trade-Off
“We use the latest AI technology” is not a technical answer. The specific model matters, different models have different context windows, pricing structures, latency profiles, and accuracy characteristics for specific tasks. An agency that can’t tell you whether they’d use GPT-4o, Claude 3.5, Gemini 1.5 Pro, or a fine-tuned open-source model, and why, for your specific use case, doesn’t have a technical opinion. They have a sales pitch.
Push directly: “Which model would you use for this and why not the alternatives?” A competent team will have a considered answer. An AI-washed team will deflect or use vague capability language.
The Person Pitching Won’t Be on the Delivery Team
This was mentioned above, but it deserves its own flag. Request to speak with the developer or technical lead who would actually build the integration before you sign. If that meeting keeps getting deferred to “after we’ve agreed terms” or “once we’ve kicked off,” treat it as a serious warning. You are entitled to evaluate the builder, not just the seller.
The “Prototype” Is a Figma File or a Slide Deck
Mockups and wireframes are not prototypes. A credible technical demonstration for an AI integration shows the system actually running, an input going in, the model processing it, an output coming out, in a real environment. If the “demo” is a series of screenshots or a clickable Figma prototype showing what the interface will look like, the engineering work hasn’t started.
The one question that resolves this immediately: “Can you show me this running in production for a client similar to me, right now, on this call?” If they can’t, the conversation is effectively over. They may have the capability to build it, but you have no evidence of that yet.
They Agreed With Every Spec Item Without Pushback
This sounds counterintuitive. Pushback from an agency during scoping is a positive signal. Competent engineers know where the hard parts are. They ask uncomfortable questions about data quality, existing system limitations, what happens when the AI returns a low-confidence output, who owns the model’s errors. If an agency walks through your entire spec and says “yes, we can do all of that” with no friction, they either haven’t thought about it, or they’re agreeing now and planning to re-scope later.
Red Flags in the Contract and Commercial Terms
Pure Hourly Billing With No Outcome Gates
Time-and-materials billing is not inherently a red flag. But hourly billing with no milestones, no defined deliverables, and no point at which you can review what’s been built before continuing to pay, is. AI projects can expand in scope quietly. Hourly billing without gates makes it easy for an agency to run up hours on a problem that turns out to be harder than promised, with no mechanism for you to pause and reassess.
Look for contracts that define specific deliverables at specific points, with a decision gate before each new phase of work is authorized. If the contract is simply “X hours per week billed monthly,” ask what the exit conditions are.
”Done” Is Not Defined Anywhere in the Contract
What does completion mean for your integration? Is it code deployed to staging? Code in production? Code in production with no critical errors for 30 days? Accuracy above a defined threshold on real data? If the contract doesn’t answer that question, the agency can declare the project complete at the point that’s most convenient for them.
This is standard due diligence for any software contract. It’s especially important for AI work because “working” is subjective, a model that’s 80% accurate might be commercially acceptable for one use case and completely unacceptable for another. Your definition of done needs to be in writing.
Long Lock-In With No Performance Benchmarks
A 12-month retainer to maintain an AI integration is reasonable. A 12-month retainer with no clause allowing you to exit if the integration underperforms is not. Any agency confident in their work should be willing to include performance benchmarks and an off-ramp if those benchmarks aren’t met. Resistance to that kind of clause is a signal about how much confidence they actually have in what they’re building.
The AI-Washing Problem: Web and Marketing Agencies Rebranding Overnight
Since 2023, hundreds of web design and marketing agencies have added “AI integration” to their services page. Some of them have genuinely built the capability. Most have not, they’ve learned to talk about AI fluently, experimented with a few off-the-shelf tools, and are now offering custom AI integrations they’ve never built before.
This isn’t unique to AI. Web agencies rebranded as “digital transformation” agencies in 2015 and “growth agencies” in 2018. The pattern is predictable: a high-demand category appears, the pitch language is easy to adopt, the actual skill takes years to build. SMBs pay for the gap.
What to Ask Any Web or Marketing Agency Claiming AI Capability
Ask them to describe the last AI integration they built for an SMB. You want: the client’s industry, the specific workflow that was automated, which AI model or API was used, and what the measurable outcome was. A team with genuine track record will answer that in 90 seconds. A team that’s been doing AI marketing longer than AI engineering will give you a vague answer about “a client in retail” or pivot to showing you a case study PDF.
Agencies that do custom WordPress development or other technical work alongside AI services are sometimes better positioned than pure-AI shops, they’ve already built the discipline of delivering complex technical work to real clients. The question is still whether the AI-specific capability is real.
The One Question That Ends the Conversation Quickly
“Can you show me an AI integration running in production for a client similar to me, right now, on this call?” Not a case study. Not a demo environment. Production. A paying client. Running today.
If the answer is no, the agency is asking you to fund work they haven’t yet proven they can do. That may be acceptable at a much lower price point, with appropriate risk protections in the contract. At $30K+, it’s a significant bet on unproven capability.
Frequently Asked Questions
What questions should I ask an AI agency before signing?
Ask to speak with the developer who will build your integration before you sign. Ask them to name the specific model or API they’d use and explain why. Ask what happens when the model produces incorrect output. Ask to see a live production example for a similar client. Ask how “done” is defined in the contract. Those five questions surface more than any proposal document.
How do I verify an agency’s AI track record?
Request a direct introduction to a past client in a similar industry and ask that client whether the integration is still running, what the original scope was, and whether it delivered the promised outcome. Agencies with genuine track records welcome this. Agencies without them will offer to connect you “once we’ve agreed in principle”, which is never.
What does a credible AI integration proposal look like?
It specifies which model or API will be used and why. It shows the data flow from input to output. It addresses what happens when the model fails or returns low-confidence output. It defines “done” with measurable acceptance criteria. It includes at least preliminary thinking about data security and access controls. If any of those elements are missing, the agency is proposing before they’ve scoped.
Is a discovery phase before development always a red flag?
Not always, but a discovery phase should produce a fixed-scope build commitment, not just a report. If you pay $15K for discovery and the output is a document that says “here are our recommendations,” and the agency then asks you to approve another $50K to start building, you’ve paid for their scoping work without gaining any rights to what comes next. Discovery that’s worth paying for results in a defined deliverable, a timeline, and a price.
How do I know if an agency is AI-washing vs. genuinely capable?
Ask for a live production demonstration of AI work they’ve done for a real client, on the call, right now. Ask them to explain the specific technical decisions they made, model choice, fallback logic, data pipeline design. Ask what the hardest problem was and how they solved it. Salespeople can learn to talk about AI. They can’t fake specific technical decisions under direct questioning.
If you want to talk through what this looks like for your operation, start a conversation. We can tell you honestly whether what you’re scoping is buildable and what the real constraints are. See how we scope and build this at designodin.com/ai.