Red Flags When Hiring an AI Development Company: A Buyer's Diligence Guide

Most failed AI development engagements were predictable from the first sales call. The patterns are visible if you know what to look for. The most common buyer mistake is reading early signals charitably — assuming the vendor will rise to the occasion during execution rather than treating the sales call as their best foot forward.

This article is the twelve red flags that predict bad AI development engagements, organized by where they appear: in sales conversations, in proposals, in portfolio reviews, and in team composition.

Red flags in sales conversations

Red flag 1: Buzzword density without specifics. The vendor talks at length about "innovation," "AI-powered solutions," "transformation," "intelligent automation" without explaining how they handle data, what their model performance looks like, or how their systems work in real-world deployments. The Reddit-famous test: count the abstract terms vs concrete terms in 10 minutes. Above 70% abstract is a red flag.

Red flag 2: No willingness to describe a real deployment in detail. "Tell me about a recent engagement" should produce specifics: what the customer did, what data flowed through the system, what broke, how it was fixed. Vendors who deflect to "we have case studies on the website" or describe deployments in vague success terms usually don't have depth.

Red flag 3: Inability to describe failure. "Tell me about an engagement that went badly" filters out vendors. The strong ones describe a specific failure, what they learned, and what changed in their methodology. The weak ones claim they don't really have failed engagements or describe a "challenging client" without specifics.

Red flag 4: Sales pressure to close before evaluation. Healthy vendors give you space to evaluate. Vendors who push for closes before reference calls happen, before security review, before technical evaluation are protecting something. Discount-with-deadline tactics ("price is only good if you sign this week") are the same pattern.

Red flags in the proposal

Red flag 5: Vague scope with optimistic timeline. "Build an AI-powered solution to address your priority use case" in 8 weeks for $80K, with the actual scope undefined. The vendor is committing to a timeline before they understand the work. Either they're padding or under-scoping; both end badly.

Red flag 6: Missing MLOps and production infrastructure. Per multiple practitioner accounts, "if a vendor's proposal does not include MLOps infrastructure from day one, treat that as a red flag." Production AI systems need monitoring, observability, deployment automation, model versioning, eval pipelines. Proposals that focus on building the model without addressing these are demo proposals positioned as production proposals.

Red flag 7: Fixed-fee on undefined scope. When a firm commits to a price before scope is concrete, they protect their margin by cutting work — usually evaluation, governance, integration depth. The cuts produce a delivered project that fails to ship.

Red flag 8: Generic compliance language. "Compliance review" or "governance considerations" without specifying which regulations apply or how the work will produce auditable artifacts. Per the data practices and privacy guide, specificity is the signal.

Red flags in portfolio review

Red flag 9: Pilots and demos positioned as production deployments. The vendor's portfolio shows many "AI deployments" but on examination they're 4-week pilots, demo videos, or one-time consulting reports. None are running systems with 12+ months of production usage. Per the buyer's guide, conversion rate from pilot to production is the metric that matters.

Red flag 10: Heavy reliance on screenshots and marketing language. Strong portfolios contain technical detail: architecture diagrams, eval methodology, specific challenges and solutions. Weak portfolios contain marketing language and screenshots without backing detail. The Reddit-credible test: ask the vendor to explain a portfolio item at architecture level for 15 minutes and see if it holds up.

Red flags in team composition

Red flag 11: No named team committed to the engagement. The proposal lists firm credentials and senior partner bios but doesn't commit to who will actually deliver. The classic bait-and-switch lives here. Strong vendors name the team and commit to continuity in the contract.

Red flag 12: Junior-dominated team with one senior name on top. Inspect the proposed team carefully. Five-person team with one principal and four juniors means juniors will deliver most of the work. Sometimes that's appropriate; usually the senior name was sold and the junior delivery wasn't expected.

Patterns that compound

Red flags rarely appear alone. The proposals that fail tend to have multiple from the same family:

The under-scoped family: vague scope + fixed fee + no MLOps + no specific systems. The firm is planning to deliver less than the proposal implies, expecting the buyer not to catch the gap until late.

The bait-and-switch family: no named team + buzzword sales + missing failure stories. The pitch is sales fiction; delivery will be by a different team than the sales conversation suggested.

The trap family: fixed-fee + sales pressure + opaque portfolio. The firm is structuring to extract value before delivery is measurable.

When you see the family pattern, walk regardless of price.

A quick diligence test

A 30-minute diligence test that surfaces most red flags:

Minute 1-5: ask "what's a recent production deployment, and walk me through what broke after launch and how you fixed it." Listen for specifics.

Minute 5-15: ask to see a portfolio item at architecture detail. "Walk me through the data flow, the eval methodology, the integration points, and what you'd do differently."

Minute 15-25: ask "tell me about an engagement that went badly and what you learned." Listen for honesty.

Minute 25-30: ask "who specifically would deliver our engagement, by name, with their bios and committed allocation."

Vendors that handle this 30-minute test well are usually capable. Vendors that struggle with it are usually not.

What to do when you find red flags mid-engagement

If you've already signed and red flags emerge during execution:

Document specifically what's failing. Vague concerns are easy to dismiss; specific failures are not.

Escalate to the firm's senior leadership. Often the team doing the work is fine; the problem is misalignment between sales commitments and delivery resources. Senior leadership can sometimes fix this; sometimes can't.

Be ready to change firms. The cost of changing AI development firms mid-engagement is real but lower than the cost of completing a failing engagement. Pull the plug at month 3 if it's clearly not working; don't wait for month 9.

Recover what you can. Eval sets, documentation, and architectural decisions from the failed engagement are sometimes salvageable. Code is often not.

The honest takeaway

Twelve red flags. Buzzword sales, no real deployment description, no failure stories, sales pressure, vague scope, no MLOps, fixed-fee on undefined scope, generic compliance, pilots positioned as production, marketing-heavy portfolio, no named team, junior-dominated team.

Most bad AI development engagements show 4-6 of these. The diligence required to spot them is 30 minutes. Buyers who do the diligence pick measurably better firms. Buyers who skip it sign engagements that were predictable from the sales call.

Read the early signals. Walk away when the patterns appear. The 30 minutes saves months.